Thirty People in Almaty, Number One in America

Music artists started using Higgsfield AI.

音乐人开始用 Higgsfield AI 了。

Will Smith used it, and Snoop Dogg shot something with it too.

Will Smith 用过，Snoop Dogg 也拿它拍了片子。

Elon Musk, one of your users.

Elon Musk 是你们的用户之一。

When did you find that out?

你们什么时候发现的？

We went viral on TikTok.

我们在 TikTok 上爆了。

Which trend went viral?

是哪个玩法爆的？

The biggest one is Eyes In.

最大的是 Eyes In。

It's the transition through the eye.

就是从眼睛里穿过去的那个转场。

We went down for 13 hours.

我们宕机了 13 个小时。

We had no idea what was going on.

完全搞不清发生了什么。

I mean, we see that everything is growing, revenue is growing, the number of subscriptions is growing, the number of generations is just growing.

我们看到的是一切都在涨，revenue 在涨，订阅数在涨，生成量一路往上涨。

Yeah.

对。

Everybody's sad in the office, everybody's sad.

办公室里所有人都很丧，所有人都很丧。

We call Alex, we say: "Damn, we're down."

给 Alex 打电话，说：“操，我们挂了。”

And he's all happy, he goes and posts on Twitter: "We've got this traffic, we went viral," he's super happy.

他倒是一脸开心，跑去 Twitter 上发：“我们流量起来了，爆了”，他超级高兴。

And he says: "Not once in my entire career have I seen traffic grow 10X and the prod not go down a single time."

他说：“我整个职业生涯里，从没见过流量涨 10X 而 线上服务 一次都没挂的。”

I say, let me check right now how many times our traffic grew.

我说，我现在就去看看我们流量涨了多少倍。

I look, and we're at more than 20x.

一看，我们超过 20x 了。

Personally I was really sad, because basically, well, I'd been dreaming about this since 2013.

我个人特别难受，因为说白了，我从 2013 年就在做这个梦。

Yeah, to make a product that goes viral all over the world.

对，做一个能在全世界爆火的产品。

Uh-huh.

嗯。

And that's the moment, and we just went down.

就是这个节骨眼上，我们直接挂了。

By then it's already, I don't know, like 12 at night, night time.

那会儿已经，大概半夜 12 点，反正是深夜。

Everyone's telling us: "DevOps."

所有人都跟我们说：“找 DevOps。”

Everyone we called said, you're being DDoSed.

我们打电话找的每个人都说，你们被 DDoS 了。

This is unreal.

这不现实。

This is unreal traffic.

这种流量不现实。

And we're like: "Damn, what if competitors really did start DDoSing us, like, I don't know, the Chinese, full on."

我们就想：“操，万一真是竞争对手开始 DDoS 我们了，比如中国人，往死里打。”

Well, a really tense atmosphere, like really tense.

气氛特别紧张，是真的很紧张。

Before, for example, we had to at least not highlight anywhere that, well, the whole team is Kazakhstani,

比如以前，我们至少得做到哪儿都不主动点破：整个团队是哈萨克斯坦的，

You've got 30 people, right?

你们 30 个人，对吧？

We have 30 people and only my cofounder Alex, he's in the Valley.

我们 30 个人，只有我的联合创始人 Alex 在硅谷。

Everyone else is in Kazakhstan — usually that's like a flag, when investors found out, when some partners found out, they'd react, well, with a lot of confusion.

其余人全在哈萨克斯坦——通常这就是个减分信号，投资人知道了、有些合作方知道了，反应都是不太理解。

But now there are no questions.

但现在没人有疑问了。

Everyone calls Higgsfield, like, a team of superstars.

所有人都说 Higgsfield 是一支全明星团队。

And now we can say that everyone is from Kazakhstan, because the product already speaks for itself.

现在可以直说所有人都来自哈萨克斯坦，因为产品已经自己会说话了。

Because, probably, the viewers should understand, we're number one in America.

因为观众大概得知道，我们在美国是第一。

A video gen AI startup.

视频生成 AI 创业公司。

Our competitors raised rounds five to 7 times bigger than ours, and we beat them purely on speed, well, and on product quality.

我们的竞争对手融的轮次比我们大五到七倍，我们纯靠速度赢了他们，还有产品质量。

Hello, everyone.

大家好。

Another episode of the nFactorial Podcast.

又一期 nFactorial Podcast。

The nFactorial Podcast is conversational master classes, inspiring ones, with masters from different industries.

nFactorial Podcast 是对话形式的大师课，嘉宾是各行各业里的高手，很给人启发。

Today our guest is Yerzat Dulat, founder of Higgsfield AI.

今天的嘉宾是 Yerzat Dulat，Higgsfield AI 的创始人。

An explosive episode is coming — Higgsfield, 11 million ARR, 20,000 paying users, 2 million uhh, I mean 2 million monthly actives, 600,000 peak, 86% M1 retention, 200 million total reach on social media, incredible, incredible products,

这期要炸了——Higgsfield，1100 万 ARR，20,000 付费用户，200 万，我是说 200 万月活，600,000 峰值，86% 的 M1 retention，社交媒体上 2 亿总触达，产品好到不可思议，

by the way, 11 million in 8 weeks from launch — a phenomenal product.

顺便说，上线 8 周做到 1100 万，现象级的产品。

Season 3 we traditionally kick off with Yerzat, right?

第三季照例从 Yerzat 开始，对吧？

That's how the beginning of every season starts, with an episode with Yerzat.

每一季的开头都是 Yerzat 这一集。

This is already his third appearance on our podcast.

这已经是他第三次上我们播客了。

We're very inspired.

特别受启发。

Before we start, let's take a look at what Higgsfield can do.

开始之前，我们先看看 Higgsfield 能做什么。

Uh, Maria, [music] unclear.

Maria，[音乐] 听不清。

And here's the popular feature that blew up on TikTok.

这就是在 TikTok 上爆了的那个功能。

Wel[music]come.

欢[音乐]迎。

Thanks for finding the time.

谢谢你抽出时间。

Yeah, really glad to be on the podcast for the third time already.

对，很高兴已经第三次来这个播客了。

You're our record holder here.

你是我们这儿的纪录保持者。

We hope we'll keep this tradition going.

希望这个传统能一直延续下去。

And we're recording this in the evening, on a Saturday evening.

我们是晚上录的，周六晚上。

And one of the interesting things is that you usually wake u— wake up at 12, and today what, you woke up at 15.

有意思的一点是，你平时 12 点才起，今天好像是 15 点起的。

Yeah, yeah, because right now there's a lot of work and it's remote, on American time, right, so my schedule shifted a bit once again.

对对，因为现在活儿多，还是远程按美国时间来的，所以作息又往后挪了一次。

Let's talk about your current product, yeah, let's start the screen share.

我们聊聊你们现在的产品，来，开始共享屏幕。

So let's explain what Higgsfield is, yeah?

先讲讲 Higgsfield 到底是什么，好吧？

So Higgsfield AI is, well, the main product is image-to-video generation, I mean, the original product, but, well, it's a broad topic.

Higgsfield AI 主打的产品是 image-to-video 生成，也就是我们最早那个产品，不过这话题挺大的。

image-to-video generation.

image-to-video 生成。

And, first of all, Higgsfield is aimed at several audiences.

首先，Higgsfield 面向的是好几类人群。

And our first version was aimed at an audience of professional cinematographers.

我们第一版针对的是专业摄影师这批人。

I mean, uh, we could genera—, well, we can generate, uh, certain professional, uh, camera techniques.

就是说我们能生成一些专业的运镜手法。

So, for example, there's the dolly zoom-in.

比如有 dolly zoom in。

It's a camera technique where, uh, the person kind of comes closer and the background pulls back.

这是一种运镜手法，人往前推，背景往后拉。

Mostly it's used to show some emotional moment in films.

电影里主要用它来表现某个情绪时刻。

I mean, I don't quite remember, well, I know, I think it was definitely Hitchcock who popularized it, well, if I'm not mistaken.

我记不太清了，但我知道，应该确实是 Hitchcock 把它带火的，如果没记错的话。

So, I mean, it's a popular film technique, right, specifically a camera technique.

这是个很流行的电影手法，确切说是运镜手法。

Then, for example, there's Car Grip — that's also a camera technique, where they literally, uh, mount the camera onto the car.

再比如 Car Grip，也是一种运镜手法，直接把摄影机挂在车上。

Mm-hm.

嗯。

The characters, the actors are driving and the camera is strapped to the car.

角色、演员在开车，摄影机绑在车上。

Also, for example, there's a very expensive camera technique called Boltcam.

还有一种特别贵的运镜手法，叫 Boltcam。

And here, with the robot.

这个就是 机械臂拍的。

Mm-hm.

嗯。

Yeah.

对。

I mean, uh, this Boltcam technique, it's very expensive.

Boltcam 这个手法特别贵。

I mean, uh, the Boltcam itself costs, it can cost hundreds of thousands of dollars.

Boltcam 本身可能就要几十万美元。

It's this robotic arm, uh, whose movement a separate programmer programs.

就是一条机械臂，得有专门的程序员给它的运动编程。

And it's a very popular technique in music videos — like Travis Scott, like Kendrick Lamar, they love using the Boltcam in their videos.

这在音乐录影带里特别流行，Travis Scott、Kendrick Lamar 都爱在自己的 MV 里用 Boltcam。

And, well, it's this democratization for cinematographers, that now, well, they don't have to buy or rent super expensive equipment.

这对电影摄影师是一次平权，现在不用买、也不用租超贵的设备了。

As far as I know, in Almaty there are only two Boltcams in total, for the whole million people, right, for the whole city.

据我所知，Almaty 一共就两台 Boltcam，一百万人口，整座城市就这两台。

And renting one, well, costs several million tenge.

租一次要好几百万坚戈。

And here you can, well, use it for a 9 dollar Higgsfield subscription.

而在这儿，9 美元的 Higgsfield 订阅就能用。

Also, for example, there's Bullet time.

再比如还有 bullet time。

This, uh, the creator, well, one of the cinematographers on The Matrix came up with this technique, specifically the guy who did the special effects there.

这个手法是《黑客帝国》的一位摄影师想出来的，就是做特效的那个人。

And it's when, well, there's this slowing of time and this epic camera movement, uh, time slowing down.

就是时间慢下来，配上很史诗的运镜，时间放慢。

That's bullet time.

这就是 bullet time。

And the dude who came up with it, well, bullet time itself, he's one of the first users of Higgsfield — Jason Zada — he even wrote an article about us in Forbes, that, well, this really is a democratization of these tricks, camera movements, slo-mo.

想出 bullet time 的那哥们儿，是 Higgsfield 最早一批用户，Jason Zada，他还在 Forbes 上写文章讲我们，说这真的是把那些花活、运镜、慢镜头都平权了。

I re— re—, by the way, I remember, after The Matrix this, this effect was used really well in the film Wanted by Timur Bekmambetov, right, when the hero, John— James McAvoy's character, I think, shoots at his opponent and the bullet flies off at an angle like this, and yeah, very, very striking, yeah, slo-mo

我记得，《黑客帝国》之后，这个效果在 Timur Bekmambetov 的《通缉令》里用得特别棒，就是 James McAvoy 演的主角朝对手开枪，子弹这样斜着飞出去，非常非常带感，慢镜头

there's also, uh, things like, for example, Mouth In, it's very popular in music videos to use that effect, transitions through the mouth, when a rapper is spitting bars.

还有像 Mouth In，这个效果在 MV 里特别流行，从嘴里穿过去的转场，说唱歌手在那儿念词的时候。

Then the camera flies in.

然后镜头飞进去。

It's a very trendy transition.

是个非常时髦的转场。

So.

就是这样。

And after camera movements we went into VFX, I mean generating special effects, like, for example, houses blowing up.

做完运镜之后我们进了 VFX，就是生成特效，比如房子爆炸。

And here's an example.

这就是个例子。

I mean, uh, effects like that are also very expensive to make.

这类特效做起来也非常贵。

And when something's exploding there, you can also blow up cars, and even blow up people.

画面里有东西爆炸，也可以炸车，甚至炸人。

With us you can — back in the eighties Jackie Chan blew things up for real.

在我们这儿可以——80 年代 Jackie Chan 是真炸。

Yeah, yeah, now, yeah.

对对，现在，对。

So.

就是这样。

And all sorts of superhero-style effects, turning into metal, all kinds of transformations.

还有各种超级英雄风格的特效，变成金属，各种变身。

So.

就是这样。

And, here, for example, something similar, if anyone watched the film Ghost Rider back in the 2000s.

比如这个也差不多，看过 2000 年代那部《恶灵骑士》的话。

Uh, a similar effect too.

也是类似的效果。

that transformation into a man of fire, or like the Fantastic Four.

变成火人那种，或者《神奇四侠》。

This is a teaser for this year's Burning Man.

这是今年 Burning Man 的预告片。

Yeah.

对。

Yeah.

对。

And so the product comes down to being able to, uh, generate, well, both for professional cinematographers who actually shoot, and for the people doing post-production, special effects for them too.

所以产品就是能生成——既给真正在拍片的专业摄影师，也给做后期的人，特效也是给他们的。

Uhh, so those are the two main products.

这就是两个主要产品。

And then the third product, which literally came out a couple of days ago.

然后第三个产品，就这两天刚上。

It's called Higgsfield Speak.

叫 Higgsfield Speak。

Two main ones.

两个主要的。

What's the first one called?

第一个叫什么？

What's the second one called?

第二个叫什么？

The first one is, well, it was just Higgsfield I2V, right, Motion Controls, I mean basically it's camera control.

第一个，其实就是 Higgsfield I2V，Motion Controls，主要就是控制镜头。

Mm-hm.

嗯。

And the second product is VFX.

第二个产品是 VFX。

And, uh, the third of the big products is Speak.

第三个大产品是 Speak。

And now, on top of being able to animate an image with some kind of control, now you can also make the character talk.

现在除了能靠某种控制让图片动起来，还能让角色开口说话。

Uh, I mean, for example, podcasters, they can deliver lines.

比如播客主，可以让他们念台词。

I mean, I upload my photo, I upload the text, and it will animate me and and it will speak the text that I wrote.

我上传自己的照片，上传文字，它就会把我动起来，然后念出我写的那段话。

Yeah, yeah, yeah.

对对对。

And this is, you could say, state of the art, because avatars were originally invented by HeyGen, a pretty old company, well, and a big one.

可以说这是 state of the art，因为数字人最早是 HeyGen 做出来的，一家挺老、也挺大的公司。

They're at something like 50 million ARR.

他们大概有 5000 万 ARR。

And their avatars are static, I mean only the face gets animated, but the hands, the body — they can't walk, they can't gesture.

他们的数字人是静态的，只有脸会动，手、身体都不动，不能走路，也不能比手势。

And our product is cool because our avatars are more alive, I mean they can walk up, they can gesture, they can, for example, cry, get angry, uh, show different, I mean, emotions.

我们产品厉害在数字人更像活的，能走过来，能比手势，还能哭、能发火，能表现各种情绪。

And, well, right now this is state of the art.

目前这就是 state of the art。

State of the art, meaning, well, the best solution on the market in the world right now, right, that exists among avatars at the moment.

state of the art，意思是当下全世界市场上最好的方案，数字人这块目前最好的。

And this has enormous potential, because a huge number of AI influencers are showing up now, AI podcasters, there's a big trend right now for AI podcasts, AI influencers who talk about stuff, they can run ads, they can sell things, they can do storytelling.

这个潜力巨大，因为现在冒出来一大批 AI 网红、AI 播客主，AI 播客和 AI 网红正是大趋势，他们在那儿讲东西，可以接广告，可以带货，可以做故事内容。

And all of that is done with avatars.

这些全靠数字人来做。

And right now we have the best solution on the avatar market.

而数字人市场上现在最好的方案就是我们的。

And since our expertise was originally camera movement and VFX, on top of our avatars just being able to emote and gesture, they can also work together with VFX, for example — and here, our avatar is about to load.

因为我们最早的强项就是运镜和 VFX，所以我们的数字人除了会表情、会手势，还能跟 VFX 配合，比如——我们的数字人马上就加载出来了。

Uh, for example, here's the avatar and here's the VFX starting to appear.

比如这是数字人，然后 VFX 就开始出来了。

I mean her eyes started, uh, glowing, lightning appearing in the frame.

她的眼睛开始发光，画面里出现闪电。

Here, for example, a grandma is saying something, and tentacles are coming out of her eyes.

再比如这儿有个老太太在说话，眼睛里伸出触手。

I mean, the same as what I showed before, they can burn, they can explode, all kinds of special effects can, uh, happen.

跟我前面展示的一样，他们可以着火，可以爆炸，各种特效都能发生。

There, the hair burns up, right, and you can customize the avatars, you can pick them, generate your own.

你看头发烧起来了，数字人是可以自定义的，可以挑现成的，也可以自己生成。

They can be cartoonish, it doesn't have to be just people, it can also be animals, some kind of animated characters and so on.

可以是卡通的，不只是人，也可以是动物、动画角色等等。

So.

就是这样。

And we have a lot of these little drops of various products.

我们还有很多这种小的产品 drop。

For example, we have a product—

比如我们有个产品——

Uh, can I ask a quick question?

能插个问题吗？

I noticed, when we were talking about the product, there's this reporting use case, and in the background behind her it's like there's some kind of combat going on.

我注意到刚才讲产品的时候，有个 reporting 的 use case，她背后就像正在打仗。

Uh, I have to upload a photo of that combat, right, is that how it works?

那我得自己上传那个战场画面的照片，是这样吧？

Uh, well, you can generate it, I mean inside Higgsfield you can generate the image, besides image-to-video.

也可以直接生成，就是在 Higgsfield 里面除了 image-to-video，还能生成图片。

There's also text-to-image, you can generate an image from a prompt, I mean, uh, so, yeah, you can design the setting, you can design the avatar character itself, I mean full control, you can design the voice, pick a voice or prompt your own voice.

还有 text-to-image，可以按 prompt 生成图片，所以场景可以自己设计，数字人角色本身也能设计，完全可控，声音也能设计，可以选一个现成的，也可以用 prompt 生成自己的声音。

What's the duration of this video going to be?

这个视频时长是多少？

Uh, right now the maximum duration is 13 seconds, but that's just so we could release fast, we set 13 seconds, but actually it's not limited.

目前最长 13 秒，但这只是为了快点 release，我们就设了 13 秒，实际上并没有限制。

I mean in the next releases the time will be unlimited.

接下来的 release 里时长就不限了。

So.

就是这样。

And we also have, uh, we had a drop with ads, meaning you can upload a photo of your product and it generates, uh, an ad video.

我们还有——之前做过一个广告方向的 drop，就是你上传自己产品的照片，它给你生成一条广告视频。

So.

就是这样。

Uh, you can look at this Coke, yeah, for example, or like what's interesting here, yeah, brands are already using it, yeah, it—

可以看看这个可乐，比如说，或者看这里有意思的地方，品牌方已经在用了，

for advertising, right?

用来打广告，对吧？

Yes.

对。

I mean, basically, uh, big brands are already using us, for example, uh big ones, so literally just the other day, uh I just saw on social media a big eyewear brand, Meller, uh and they generated a big post there, a lot of video slides, on our Higgsfield.

大品牌已经在用我们了。比如前几天，我在社交媒体上刷到一个做眼镜的大品牌 Meller，他们在我们 Higgsfield 上生成了一整篇帖子，好多条视频。

So.

就这样。

And, well, I mean we don't have any kind of B2B track, like where we'd go out to that brand and try to sell to them.

而且我们根本没有 B2B 那条线——不是我们去找这个品牌、想把东西卖给他们。

It's just that, well, the product right now is so popular among creators that, most likely, well, the creator at Meller, the biggest one there, probably it's one of the biggest eyewear brands, they generated it with us.

纯粹是产品现在在创作者里太火了，很可能是 Meller 自己的创作者做的——那大概是最大的眼镜品牌之一——他们就在我们这儿生成了。

So, maybe some other well-known brands are working with you there, I don't know, or from among the celebrities, who's using you, right?

那还有别的知名品牌在跟你们合作吗，或者名人里，谁在用你们？

I mean, uh, since, uh, globally the product is positioned for, uh, originally for music videos, because everything we have is very stylish, cool, kind of youthful, there are a lot of cool transitions there that get used in music videos.

产品在全球的定位，最早就是给音乐 MV 用的，因为我们的东西都特别有型、酷、偏年轻，里面有很多 MV 里会用的炫酷转场。

That's why there was immediately fast adoption among, well, music videos.

所以音乐视频这块马上就用起来了。

And also we don't have, well, some B2B department there that reaches out to music productions.

而且我们也没有什么 B2B 部门去对接音乐制作公司。

And the music artists themselves started using Higgsfield.

是音乐人自己开始用 Higgsfield 的。

And so one of the latest ones, Will Smith used it in his, uh, music, in his music video that he posted on Instagram, and that video has millions of views.

最近的一个是 Will Smith，他用在自己的音乐视频里，发在 Instagram 上，那条视频有几百万播放。

And he used a specific effect, Soul Jump.

他用的是一个具体的效果，叫 Soul Jump。

There it is.

就是这个。

I mean it's this move where, uh, the soul kind of flies out of the person.

就是那种手法，灵魂从人身上飘出来。

Such a fun, very stylish, uh, example.

特别好玩，非常有型的一个例子。

And Snoop Dogg also shot one.

Snoop Dogg 也拍了一条。

Snoop Dogg, he actually, well, went for, it seems, the maximum subscription.

Snoop Dogg 好像直接上了最高档的订阅。

Uh, well, whoever did it for Snoop Dogg, probably his assistant.

给 Snoop Dogg 做这个的，大概是他的助理。

And he, it seems, tried to use up all the credits, because there, well, on the order of, probably, ten, possibly more than ten effects in total were used.

他好像是想把 credit 全用光，因为那条里大概用了十个、可能十几个效果。

I mean Snoop Dogg just used absolutely everything across the board and made a music video out of it.

就是说 Snoop Dogg 把效果一股脑全用了一遍，然后拼成一条 MV。

[music]

[音乐]

Madonna, I think, right, recently, the one that came out.

还有 Madonna 吧，最近出来的那条。

Yeah, yeah.

对，对。

So we literally just released that like 2 days ago.

那个我们就是 2 天前刚发的。

Today's Saturday, we released it on Friday, on Thursday.

今天周六，我们是周五发的，周四。

in the evening, and 5 hours later Madonna posted in her Instagram story with, well, there she is an AI generated Madonna advertising uh some new vinyl album of hers, and she used exactly VFX, I mean she first

晚上发的，5 小时后 Madonna 就在自己的 Instagram story 里发了，里面是个 AI generated 的 Madonna 在推她的一张新黑胶专辑，她用的正是 VFX，就是说她先

I mean she used specifically avatars with special effects, and that hadn't existed before.

她用的正是带特效的虚拟形象，这种以前没有过。

This, well, we did it just for the fun of it, like, why not, let's combine avatars and special effects.

这个我们纯粹是图好玩做的，想着为什么不呢，把虚拟形象和特效合到一起。

And 5 hours after the release Madonna was already using it.

上线 5 小时后 Madonna 就用上了。

That was, yeah, cool.

这个确实挺爽的。

And among famous people, well, a lot of them use it, it somehow gets back to us.

名人里用的人挺多的，多多少少会传到我们这边。

But the nicest thing, probably, for the engineers, our ML guys, was that Elon Musk became a user, because, probably, I'll talk a bit later about our marketing.

但最让工程师、让我们做 ML 的那帮人开心的，大概是 Elon Musk 成了用户，因为——我待会儿再讲我们的营销。

And we do a ton of marketing on Twitter.

我们在 Twitter 上做了特别多营销。

And it so happened that he constantly saw our clips and also joined in there, became a Higgsfield user, liked our posts, left comments and, well, that's very, in general, yeah,

结果就是他一直刷到我们的视频，然后也加进来了，成了 Higgsfield 的用户，给我们的帖子点赞、留言，这个真的特别，总之，对，

your, your, the guy using you right now, right, I mean a legend of entrepreneurship, right, I mean who has a 200 million audience on x.com.

现在在用你们的这个人，是创业圈的传奇，在 x.com 上有 2 亿受众。

Yeah.

对。

Yeah.

对。

And, well, on on X we have crazy numbers there on views.

我们在 X 上的播放数据特别夸张。

Uh, well, our posts rack up tens of millions of views, because we originally targeted X, to try to do marketing there, and only then started moving over to other social networks.

我们的帖子能跑出几千万播放，因为我们一开始就瞄准 X，先在那儿试着做营销，之后才往别的社交平台扩。

Since our team is small, and so we started with X.

因为团队小，所以就从 X 开始。

Right now we're doing marketing on TikTok, on Instagram.

现在我们在 TikTok、Instagram 上做营销。

In the future, hopefully, we'll be on Twitch.

以后希望能做到 Twitch，

on Reddit and so on.

还有 Reddit 之类的。

Very interesting.

太有意思了。

So you did 12 updates over over the last 8 weeks.

你们过去 8 周做了 12 次更新。

Uh, tell me, how did you pull that off?

讲讲，你们是怎么做到的？

Yeah, well, probably because it worked out that we have a super powerful team.

大概是因为我们碰巧凑出了一支超强的团队。

And in itself, well, it's just, probably, some kind of luck with how circumstances went, that we managed to put together some incredibly powerful team.

这件事本身大概就是机缘巧合的运气，才凑齐了这么一支强得离谱的团队。

of engineers, designers, a creative team which, well, well, in my view, we've just got world-class superstars working there, at the beginning, well, only, for example, well, internally we understood how cool all the guys are.

工程师、设计师、创意团队——在我看来我们这儿就是一群世界级的超级明星在干活。一开始只有内部知道这帮人有多强。

And, uh, well, before, for example, we kind of had to at least not highlight it anywhere — like, the startup is American, you know, Silicon Valley, but actually all the engineers are Kazakhstani, well, the whole team is Kazakhstani, 30

而且以前我们至少得做到哪儿都不提这事——对外是美国创业公司、硅谷，但其实工程师全是哈萨克斯坦人，整个团队都是哈萨克斯坦的，30

…people do you have, right,

……你们有多少人,对吧,

we have 30 people, and only the cofounder, my cofounder Alex, he's in the Valley.

我们有 30 个人,只有联合创始人——我的联合创始人 Alex——在硅谷。

Everyone else is in Kazakhstan.

其他人全在哈萨克斯坦。

And we didn't, we never highlighted this anywhere.

我们从来没在任何地方拿这事出来说。

Usually it's like a red flag, when investors found out, when some partners found out.

这通常算个危险信号——投资人知道了,某些合作方知道了。

It was always a red flag, or at the very least, well, people just don't get it.

这一直是个危险信号,再不济也是人家理解不了。

Well, well, it's a normal reaction, when people run into something unknown they're very, well, they treat it with incomprehension, and, especially Americans.

这也是正常反应,人碰上陌生的东西就是不理解,更别说美国人了。

But now, uh, there are no questions.

但现在没人有疑问了。

Everybody knows that, well, in the comments, in private conversations, in DMs, uh, well, everyone calls Higgsfield a team of superstars.

所有人都知道——评论区里、私下聊天里、私信里,大家都说 Higgsfield 是一支超级明星团队。

And now, well, obviously, you can say that everyone is from Kazakhstan, because the product already speaks for itself.

现在当然可以直说所有人都来自哈萨克斯坦,因为产品已经自己说明一切了。

And by the way, Ole Mir gave me a really good piece of advice.

顺便说,Ole Mir 给过我一个特别好的建议。

We get on calls with each other often.

我们经常通电话。

also a guest of the nFactorial podcast, an MIT graduate, a really cool engineer, a founder.

他也上过 nFactorial 播客,MIT 毕业,非常厉害的工程师,创始人。

And when I was telling him about this, he gave me the advice, back when we hadn't launched the product yet, uh, say that the team is, well, not just from Kazakhstan, say that it's simply a world-class team, but the one thing that unites everyone is that they're all from Kazakhstan, but in reality they're all world-class.

我跟他讲这事的时候,产品还没上线,他给我的建议是:别只说团队来自哈萨克斯坦,就说这是一支世界级团队,只不过所有人有一个共同点——都来自哈萨克斯坦,但实际上人人都是世界级的。

I mean there are a lot of winners of international olympiads in physics, in math.

队里有一堆国际物理、数学奥赛的获奖者。

And on the creative team we've got directors who made music videos with tens of millions of views.

创意团队里有拍过几千万播放量 MV 的导演。

So there you go, I mean, somehow that's just how it turned out.

反正,不知怎么就成了这样。

Uh, yeah.

对。

And, uh, tell me about your — you've got a wonderful team, right, you've got, and, and, yeah, and product guys, and ML engineers, a creative team, prompt engineers, you've got directors, right, actual film directors work at your company, marketing, right.

跟我讲讲你们的——你们团队特别棒,有做产品的人,有 ML 工程师,有创意团队,有 prompt 工程师,还有导演,真的有电影导演在你们公司上班,还有市场,对吧。

Uh yeah, come on, let's talk about your, well, your workday.

来,咱们聊聊你们一天怎么干活的。

I've been at your office several times.

我去过你们办公室好几次。

It was a day off, a workday, 9:00 pm, the whole team in the office.

有休息日,有工作日,晚上 9:00,整个团队都在办公室。

Tell me, what's the magic here.

讲讲这里头的魔力在哪。

7 days you work, how many days do you work there, 100 hours a week, by the looks of it.

你们一周干 7 天,到底一周干几天,看样子一周 100 小时。

That's probably why Elon Musk likes your stuff, because it's his style.

所以 Elon Musk 才给你们点赞吧,这正是他的路子。

Yeah, yeah.

对,对。

We, well, we try to work a six-day week, but it doesn't work out for us, often we have to, because it's the other way around, I have to hold the guys back, and tell them: "Come on, let everybody rest, we'll release on Thursday."

我们试着搞六天工作制,但常常做不到,因为反过来是我得拦着这帮人,跟他们说:「大家都歇一歇,周四再 release。」

But the guys are super fired up, like no, we're going to release on Monday, and on Sunday they don't go home to rest, they prep the release.

但这帮人特别上头,「不行,我们周一就要 release」,周日也不回家休息,在那儿准备 release。

I mean, well, the team is super ambitious, so, and everyone's a superstar.

团队野心特别大,而且个个都是超级明星。

If you take the product, we've got a top-tier site, and the product itself, if you compare it with competitors, well, if you look at the other competitors, with a lot of them, for example, the mobile site just doesn't work, or the design is, well, really outdated,

说到产品,我们的网站是顶级的,产品本身跟竞品比——你去看别的竞品,很多家的移动端网站根本打不开,或者设计非常过时,

for example, well, design right now has also started changing really fast in the AI era, and that totally flat minimalist design, it's already on its way out, uh, receding,

比如说,AI 时代里设计本身也开始飞快变化,那种彻底扁平的极简设计已经在退场了,

for instance, Brian Chesky, the founder of Airbnb, recently — I mean, he's a designer himself, he went to design school, and Airbnb was always kind of the, uh, gold standard of design, a pioneer.

比如 Airbnb 创始人 Brian Chesky 最近——他本人就是设计师,念的设计学院,Airbnb 一直被当成设计的标杆、开路者。

And he recently posted a tweet saying that this minimalism, flat design, all of that is yesterday's news, it's not cool.

他最近发了条推,说极简、扁平设计这些都是昨天的东西了,不酷了。

What's cool is exactly that kind of maximalism.

酷的恰恰是那种极繁。

And even before that tweet, before all this buzz, we were of the same opinion.

在那条推之前、在这波风潮之前,我们就是这么想的。

We have an insanely cool designer, Madi.

我们有个牛到不行的设计师 Madi。

We looked for him for a really long time, uh, for a cool product designer.

为了找到这么个厉害的产品设计师,我们找了非常久。

Originally he was a graphic designer, I mean we hired him for a graphic design position and, yeah, graphic design, but with us he became a product desi— designer.

他起初是平面设计师,我们招他就是做平面设计的,但到了我们这儿他变成了产品设计师。

And then I realized this, when I was reflecting on it, that among product designers there are a lot of people who are more about metrics, about that kind of thing.

后来我复盘的时候想明白了:产品设计师里有很多人更看重指标之类的东西。

But Madi, he was originally a graphic designer, and he really gets design, I mean fonts, the visuals.

而 Madi 本来就是平面设计出身,他是真懂设计——懂字体,懂视觉。

Well, basically, he gets it, like, he doesn't A/B test the way it looks.

基本上他门儿清,外观这块他不拿 A/B 测。

He says, my opinion is this, my taste says that this here would be more right, yeah, but he A/B tests it anyway.

他会说,我的看法是这样,我的品味告诉我这样更对,但他还是照样跑 A/B 测试。

I mean we're still, yeah, a fully data-driven company, but at the same time he's very, well, high aesthetic, I mean, well, I think if a person is a designer, he should, well, get design globally, like, not ignore his own taste, in general, yeah, he should get architecture, get fashion, uh, well, typography and so on.

我们终归是一家彻底 data-driven 的公司,但他同时审美非常高。我觉得一个人既然是设计师,就该整体地懂设计,别把自己的品味丢掉——得懂建筑,懂时尚,懂排版这些。

So.

就这样。

And the design came out really cool.

所以设计做得非常出彩。

And that's why, well, everyone, well, if you take the industry, ours, Higgsfield has one of the coolest designs, that's already a fact.

所以在我们这行里,Higgsfield 的设计是最好的之一,这已经算是公认的事实。

So.

就这样。

And the product — also an nFactorial graduate, Almaz Theolden, he basically owns the entire product, and he's, well, he's an engineer himself.

产品这块,也是 nFactorial 出来的,Almaz Theolden,整个产品基本上归他扛,而他本人就是工程师。

Uh, and that's probably important too, that the Product Lead was originally an engineer.

这一点可能也很关键——Product Lead 最好是工程师出身。

Almaz is an iOS developer, right, originally,

Almaz 最早是 iOS 开发,对吧,

yeah, he was originally an iOS developer.

对,他一开始是 iOS 开发。

And while he was doing iOS development, he leveled up insanely in product analytics.

做 iOS 开发那阵子,他在产品分析上练得特别猛。

So he's the one running the A/B tests, everything's covered with events, everything's covered with Amplitude.

A/B 测试就是他在跑,所有环节都埋了事件,全接上了 Amplitude。

I mean these are the kinds of things so that you can see and understand every single move the user makes, and what the funnel of product usage looks like, so you can optimize it further.

做这些就是为了让你能看见、看懂用户的每一个动作,看清产品的使用漏斗长什么样,再接着往下优化。

And in the end, well, obviously, we're a business, revenue matters to us, so it's very important to run a whole lot of A/B tests, to raise the conversion into paying users, to raise … retention.

说到底我们是做生意的,revenue 才是关键,所以必须跑大量 A/B 测试,把付费用户转化率拉上去,把 … retention 拉上去。

And Aidar, oops, Almaz, he's a superstar at this.

Aidar——啊说错了,Almaz,他在这上面是超级明星。

He knows how, I mean, to own the product and squeeze the maximum out of it.

他知道怎么把一个产品完全扛起来,再从里面榨出最大值。

So.

就这样。

And your creative team?

那你们的创意团队呢?

Yeah,

对,

the creative team.

创意团队。

It — well, that's exactly why to this day I say it was luck.

这支团队——所以我到今天还是说,这是运气。

I don't get it.

我不明白。

I mean, sure, I'm a machine learning engineer, and, well, putting together a top machine learning team — the one I talked about in a lot of detail back on the previous podcast — assembling guys like that was, well, a clear-cut task for me, because I'm an ML guy, an engineer originally, but I had never really done any products, let alone, uh, well, at the level we ended up with.

当然,我是机器学习工程师,组一支顶级的机器学习团队——上一期播客里我讲得很细的那支——招这样的人对我来说路径很清楚,因为我本来就是搞机器学习的,工程师出身,但我从来没做过产品,更别说做到我们后来这个水平。

And, uh, I was even further removed from the creative industry.

而且我离创意行业更远。

And it worked out that, uh, here my wife helped me, she's closer to it, uh, Togzhan, she's closer to the creative industry.

结果是,这块儿是我太太帮了我,她离这行更近——Togzhan,她更贴近创意行业。

And she recommended some guys to me who are in our creative scene, really cool young directors who had shot music videos with tens of millions of views.

她给我推荐了几个人,都在我们的创意圈里,是很厉害的年轻导演,拍过几千万播放量的MV。

Uh, that's Ziya, Karim, and the two of them, uh, originally started out, I think, at Ozen.

就是 Ziya、Karim,他们俩最早好像是在 Ozen 起步的。

I mean that's also Aizotula, our super cool production house.

也就是 Aizotula,我们这儿超牛的制作公司。

Uh, they started there, they made videos that — well, I just didn't get any of this before, uh, about this industry, I knew little about it.

他们从那儿起步,拍MV,那些片子——我以前根本不懂这行,对这个行业知道得很少。

But it turns out, I mean, in Kazakhstan the music video industry is top-tier on a world level.

结果发现,哈萨克斯坦的MV产业是世界顶级水平。

And they were exactly the ones doing them at Ozen, shooting all these cool videos there.

他们在 Ozen 干的就是这个,那些牛片子都是在那儿拍的。

And, well, and later, after Ozen, they also did a whole lot of cool projects, both commercial and advertising ones, and music ones, creative ones.

后来离开 Ozen,他们也做了很多厉害的项目,商业的、广告的、音乐的、创作类的都有。

And, uh, and they also started pulling in more people after that.

然后他们又开始往里拉人。

We've got, uh, Danil there, a top cinematographer actually, who shot, well, all kinds of technically complicated stuff, there, I think, he shot Morgenshtern's video Volmaty, as the cinematographer on it.

我们这儿有 Danil,顶级摄影指导,拍过各种技术上很复杂的东西,好像 Morgenshtern 那支 Volmaty 的MV就是他掌镜的。

And so all this expertise came to us — how to properly shoot music videos, how to properly do the visuals, how to properly do the presentation.

整套这些经验就这么进来了:MV该怎么拍、视觉该怎么做、呈现该怎么做。

So that we — uh, and after that a really strong reputation specifically as a creative team appeared.

后来我们就有了非常硬的口碑,而且正是作为一支创意团队的口碑。

Probably the indicator of that is that after one of our cool releases — and our releases really do look like music videos, because Ziya and Karim direct them — and uh, after one release OpenAI wrote to us.

最能说明这点的是:有一次我们做完一个很棒的 release——我们的 release 看起来真的就像MV,因为是 Ziya 和 Karim 执导的——那次之后,OpenAI 主动写信来找我们了。

They were, well, just blown away by that level.

他们完全被这个水准惊到了。

Because, the way I understand it, Silicon Valley and LA — well, LA is the creative scene, right, of America.

因为据我理解,硅谷和 LA——LA 是美国的创意圈,对吧。

They apparently don't overlap at all, becau— well, because the competitors who do video, they don't have anything even close to the visuals we have.

这两边好像根本不交叉,因为做视频的那些竞品,视觉水准跟我们完全没法比。

With them everything is really sad, depressing, if you look at it.

他们那边的东西很惨,看着让人难受。

A human being just can't watch that.

正常人根本看不下去。

And when OpenAI saw that everything is, well, that cool on our side, they decided, well, they wanted to work with us, and we're their early partners now, and we have early access to all of their models, uh, I mean, well, they want to collaborate with us — I mean OpenAI themselves — it's not like we came to them.

OpenAI 看到我们这边做得这么好,就决定要跟我们合作,我们现在是他们的早期合作伙伴,他们所有模型我们都能早期用上——是他们想跟我们合作,是 OpenAI 自己找上门,不是我们去求人家。

So there.

就这样。

And it worked out, yeah, we just managed to put together some magnificent team.

总之就是凑齐了一支特别棒的团队。

Also, let's move on for a moment, before you get to the other members of your team.

我们先往下走一点,在你讲团队其他成员之前。

Let's quickly show it.

咱们快速放一下。

So you mentioned that Ziya joined you literally a week ago, right?

你刚提到 Ziya 是一周前才加入你们的,对吧?

Am I saying that right?

我说得对吗?

No, he's already been working with us for several months.

不,他在我们这儿已经干了好几个月了。

Uh, so who joined you a week ago and then right away, in 3 days, shipped Speak?

那一周前加入、三天就把 Speak 上线的是谁?

Yeah, that's a super cool, uh, director.

对,那是个超厉害的导演。

And everybody knows him.

所有人都认识他。

Well, I think, in the CIS, in the creative scene, he shot music videos for the biggest Russian rap artists.

我觉得在独联体的创意圈里,他给俄罗斯最大牌的说唱歌手拍过MV。

I mean, I just don't know this stuff, but it's like, well, the top actors — oops, rappers of Russia.

我自己不懂这些,但那些就是俄罗斯最顶级的演员——口误,说唱歌手。

That's Ilya Karchin, he's himself, uh, from Almaty, he graduated from School 134, the physics-math one.

他叫 Ilya Karchin,他本人就是 Almaty 的,毕业于 134 物理数学中学。

I mean he — he's actually a mega-powerful techie on top of that.

而且他其实还是个超级硬核的技术人。

And I mean he gets everything.

什么都懂。

VFX, uh, motion design, uh, how to properly shoot live-action footage, how to do, uh, well, he's a director himself, I mean and he, yeah, he literally joined a week ago and got up to speed with the process fast.

VFX、motion design、实拍该怎么拍、怎么做——他本人就是导演,而且他真的是一周前才加入,很快就融进流程了。

And so, yeah, the Speak release, which Madonna then posted in her stories, that was Ilya Karchin's work.

对,Speak 那次 release,后来 Madonna 发到自己 story 里的那个,就是 Ilya Karchin 做的。

Let's take a look right now, yeah, at the — at the result.

我们正好来看看成果。

Good morning and the top about [music] Now [laughter] [music] ser [music]

Good morning and the top about [音乐] Now [笑声] [音乐] ser [音乐]

Uh-huh.

嗯哼。

So you have a creative team, uh, there's a product team, there's an infra team, right?

所以你们有创意团队,有产品团队,有 infra 团队,对吧?

I mean ML engineers, MLOps, and uh a marketing team.

就是 ML 工程师、MLOps,还有市场团队。

What does the team — the marketing team do?

市场团队具体做什么?

Uh, the marketing team, it works very closely with the creative team, because all of our marketing is content based, and so, uh, we have a flat-out genius ad guy, well, just a genius of a person, that's Sultan.

市场团队跟创意团队配合得非常紧,因为我们整个市场打法都是 content based 的,所以我们有个天才级的广告人,真的就是天才,他叫 Sultan。

Sultan Unasbekov.

Sultan Unasbekov。

He used to do advertising, he came to us for a prompt engineer position, learned to generate video, and now he, well, leads our content based marketing.

他以前做广告,来我们这儿是 prompt engineer 的岗位,学会了生成视频,现在负责我们的 content based 市场。

Uh, I mean, uh, our — well, like I said before, our videos are mega-viral on Twitter, and then they went viral after that on TikTok, on Instagram.

就像我前面说的,我们的片子在 Twitter 上爆得一塌糊涂,后来在 TikTok、Instagram 上也跟着爆了。

And he, being a professional ad guy, he, uh, comes up with the ideas, executes them himself, and then even younger prompt engineers join us who learn from him.

他是专业广告人,自己想创意、自己动手做出来,后来又有更年轻的 prompt engineer 加进来,跟他学。

And we, well, I mean he does the ad videos end-to-end, from the idea all the way to the generation.

也就是说,从想法到生成,广告片他一个人 end-to-end 全做。

And we all riff on it together with him too.

我们也都会跟他一起把点子聊开。

Well, in advertising the narrative and the message matter a lot.

广告里,叙事和要传达的那个点非常重要。

It has to be very simple, clear, and it has to have some kind of twist to it so that it goes viral.

必须非常简单、好懂,而且里面得有点花活儿,这样才能爆。

I mean usually it has to be something like, as it's trendy to say, contrarian, well, yeah, controversial.

通常得是那种,用时髦话说叫 contrarian 的东西,对,就是有争议的。

And often it can piss off big audiences, that's why it goes viral, because they, well, people start getting mad, posting it, criticizing it.

而且经常会惹毛一大批人,所以才会病毒式传播,因为大家开始生气、转发、骂。

And, well, this is already kind of a standard playbook in genAI.

这在 genAI 里已经算标准打法了。

Actually, for example, Suno built all of its own PR marketing on this too, on the fact that, uh, Billboard I think wrote an article about them saying that, uh, well, they're going to replace music artists

比如 Suno 整套 PR 打法就是这么建起来的:好像是 Billboard 写了篇文章,说他们要取代音乐人

and there was immediately huge pushback against it, well, the industry reacted very sharply

马上就有巨大的反对声,行业反应非常激烈

and everybody found out about Suno on the — on the back of that criticism

大家就是顺着这些批评知道了 Suno

roughly that's what we do with Sultan, he comes up with an idea, for example how to make an ad that cost $100,000 for $9, and he goes Hollywood, right, or like, yeah, RIP Hollywood and stuff like that.

我们跟 Sultan 大概也这么干,他出主意,比如怎么用 $9 做出一支价值 $100,000 的广告,然后打上 Hollywood,或者 RIP Hollywood 之类的。

And uh he generated so many ads on Higgsfield, and $9 was written everywhere in them, that it became a meme.

他在 Higgsfield 上生成了太多广告,里面到处都写着 $9,结果这成了梗。

$9 — that's basically now become — meme accounts post stuff in the comments, they write things like, we're sick of you and your $9 ads.

$9 现在基本上成了——那些梗号在评论区发东西,写什么「你们那 $9 的广告烦死了」。

I mean $9 just became a global meme.

$9 就这么成了全球性的梗。

And he — we had this general idea, uh, to generate, uh, corporate wars like that, where it's as if, I don't know, someone disses Jeff Bezos.

我们还有个共同的想法,就是生成那种企业互撕的内容,比如搞得像有人在 diss Jeff Bezos。

just a totally unconnected thing, to hook somebody.

完全不搭界的东西,就为了勾住人。

And he generated a lot of videos like that.

他生成了很多这种片子。

And that also became a global meme, this thing about corporate wars on Twitter.

这也成了全球性的梗,就是 Twitter 上的企业互撕。

And others started picking it up

然后别人开始跟着做

creators.

创作者。

And one creator, actually he captioned it there, inspired by Higgsfield, I also generated a clip where it's just World War III, where McDonald's is at war against Starbucks and there are soldiers with, who've got Starbucks backpacks on their backs there with with some kind of chemical weapon against McDonald's tanks there.

有个创作者受 Higgsfield 启发,配文说他也生成了一条片子,内容就是第三次世界大战,McDonald's 跟 Starbucks 开战,士兵背着 Starbucks 的背包,拿着某种化学武器去打 McDonald's 的坦克。

That clip also pulled a ton of views.

那条片子播放量也高得离谱。

So, things like that.

就是这类东西。

Uh, well, really, if you formulate a clear message, uh, one that's controversial, viral, and properly, uh, generate it there, uh, a video, to generate it, there are a lot of rules there too.

说真的,你要是能提炼出一个清晰的信息点,有争议、能传播的,再把它漂亮地生成成视频——生成这一步本身也有一大堆规矩。

It has to hook hard at the start there, it has to hold the person to the end.

开头必须狠狠钩住人,还得把人一直留到最后。

Very, well, all these metrics matter for for especially for advertising and especially, like, in the era of social media.

这些指标都非常重要,尤其是广告,尤其是在社交媒体这个时代。

Uh, because they just optimize for the user's watch time.

因为平台优化的就是用户的观看时长。

And so, taking all of these things into account, we, well, we learned how to make clips on Twitter that pull tens of millions of views there.

把这些都算进去之后,我们摸出了在 Twitter 上做几千万播放量片子的办法。

And basically, probably, well, when Sultan and I are riffing on ideas, uh, we have this, I don't know, whether it's a dream or not, but uh, when we come come up with some crazy ideas, we're like, well, we always say: "Damn, 100% they've got to call us in as PR-marketing consultants at the next American elections."

我跟 Sultan 一起瞎想点子的时候——我们有个念头,不知道算不算梦想——每次想出什么疯狂主意,我们都会说:“靠,下届美国大选百分之百得请我们去当 PR 营销顾问。”

And what's interesting, we were once riffing on something about Trump, we were just about to post it, and in the end Trump, and he posted an AI-generated clip too, roughly the thing we wanted to riff on there.

有意思的是,我们有次策划了一个关于 Trump 的东西,都准备发了,结果 Trump 自己也发了一条 AI 生成的片子,内容跟我们想搞的差不多。

Well, let me say right away, it's not not the Palestine clip, the one he made that was of super awful quality, well, and in terms of the message too, uh, this was much later.

先说清楚,不是他做的那条巴勒斯坦片子——那条质量差得离谱,立意也差——这是晚得多的事。

So this—

就是这个——

And where did he post it?

他发在哪儿了?

On Twitter.

Twitter 上。

He posted it on Twitter, yeah.

发在 Twitter 上,对。

And, well, we know exactly how the playbook works.

我们非常清楚这套 playbook 怎么运转。

Also we can either later or right now start, well, talking about TikTok.

TikTok 我们可以晚点讲,或者现在就讲。

On TikTok we also went super hard viral there.

在 TikTok 上我们也火得一塌糊涂。

Uh, we were the main trend of the week overall even, well, several weeks, uh, well, at the peak, probably a week.

我们直接成了那一周——甚至好几周——的头号趋势,峰值大概持续了一周。

It was just our our trend was at the peak, I mean all over the world, in every country.

就是我们那个趋势冲到了顶,全世界、每个国家都是。

Uh, I mean we understand how to go viral on Twitter, how to go viral on TikTok.

也就是说,我们搞明白了怎么在 Twitter 上爆,怎么在 TikTok 上爆。

I mean we understand these playbooks very clearly.

这些 playbook 我们摸得非常清楚。

So let's lay out this playbook.

那就把这套 playbook 拆开讲讲。

So it turns out this is what you call situational marketing, right, probably, or or not necessarily?

这大概就是你们说的情境营销吧,还是说不一定?

Yeah, yeah, ours is situational marketing.

对对,我们做的就是情境营销。

Uh, I mean, well, basically, the way startups work, like ours, we don't try to plan very far ahead.

像我们这种创业公司的运作方式,我们不会去规划太远。

I mean we have a strategic vision, we, as experts, understand where the market is going to move.

我们有战略层面的 vision,作为这行的专家,我们清楚市场会往哪儿走。

I roughly, well, have this clear picture of what the market will look like in 3, 6 and 12 months.

我大致能看清 3、6、12 个月后市场长什么样。

But at the same time, locally we don't build any plans in the company.

但具体到眼下,公司内部不做任何计划。

I mean we don't have a plan like in 2 weeks we're going to release this.

我们没有那种“两周后 release 这个”的计划。

We don't have that kind of plan, because everything changes very fast, the market develops fast.

没有这种计划,因为一切变得太快,市场跑得太快。

But in marketing everything is even more situational for us.

而营销这块,我们更是完全看情境。

Uh, and I actually even ask the guys, please don't plan content 3-4 days ahead, because, well, that will really start making you kind of, well, inflexible, because in those 3-4 days a ton of stuff can happen in the world.

我甚至反过来跟大家说:拜托别把内容提前排到 3-4 天以后,那会把你们弄得特别僵,因为这 3-4 天世界上能出一堆事。

And our idea, like, won't be a trend anymore at all.

到那时候我们这个点子根本就不是热点了。

You have to feel in real time what's happening right now, what the news hooks are, what people are discussing.

你得实时感知当下在发生什么、有哪些话题由头、大家在聊什么。

When, for example, Katy Perry, like, when Jeff Bezos sent Katy Perry into space.

比如 Katy Perry 那次,Jeff Bezos 把 Katy Perry 送上太空。

Got it.

明白。

That very same day we started making a clip about it and it's super, yeah, our marketing is situational, literally real-time, and uh yeah, well, in marketing it's very, well, the work is fun, I'd say, I mean it's, well, it's of course also very stressful, I mean, uh, well, the atmosphere in our marketing department there probably resembles those movies about trading in the nineties, where in real time everyone's calling like And you gotta buy the stock, you gotta react in real time, everyone's yelling at each other.

当天我们就开始做这条片子了,对,我们的营销就是看情境,完全实时;营销这活儿挺好玩的,当然也非常有压力——我们营销部的气氛大概像九十年代那些讲交易员的电影,所有人实时打电话,喊着“快买那支股票”,必须实时反应,所有人互相吼。

That's roughly, like, like what our marketing department looks like.

我们营销部大概就长这样。

We have awesome ops guys, they're uh also top guys from Nazarbayev University, mathematicians, engineers, who, well, whatever the creative team comes up with, they then have to execute it properly, set up the distribution of that content properly.

我们有一批很强的运营,也都是 Nazarbayev University 出来的顶尖人,数学的、工程的,创意团队想出来的东西,他们负责准确落地,把内容的分发铺对。

And so the ops people there, who don't run on super emotions, they just do their job cleanly.

这些做运营的不靠情绪上头,就是干干净净把活儿干完。

So we've got Ali, Nargis, uh fresh grads, uh Kamazhai, uh Madiyar - these are our youngest, newest employees.

我们有 Ali、Nargis,都是应届生,还有 Kamazhai、Madiyar,是我们最年轻、最新的员工。

They literally just graduated university yesterday, and they're such professional ops people.

他们真是昨天才从大学毕业,可已经是非常职业的运营了。

They build the process.

流程是他们搭起来的。

And so whatever the creative team, the crazy ideas they come up with, they then distribute it.

创意团队想出的那些疯狂点子,后面由他们分发出去。

So.

就这样。

And it works out very, well, very cleanly.

配合得非常利落。

And data driven, I mean we then, we have a top data analyst who is basically responsible for most of the company's business processes, so that we, well, we try to move fully data driven.

而且是 data driven 的,我们有个顶尖的数据分析师,公司大部分业务流程基本都归他管,我们想做到完全 data driven。

And he then analyzes what our CPM costs, what our worldwide reach was, what the, uh, information food chain in general looks like, uh how to deliver content properly, what time you need to post them and so on.

然后他分析我们的 CPM 是多少、全球触达有多少、整条信息食物链长什么样、内容怎么投递才对、该在什么时间发,等等。

Mhm.

嗯。

Do your marketing teams have some kind of KPI?

你们营销团队有 KPI 吗?

How many videos do they have to ship, to publish a week?

他们一周得推出、发布多少条视频?

For for us, I wouldn't say we really run on KPIs, like I said.

我不会说我们真是按 KPI 在跑,前面也说过。

With us the guys, how much how much they publish in practice, roughly, yeah, they just, well, on the contrary you have to, basically, my bigger problem is actually asking an employee to go get some rest, to stop generating, stop riffing, because, well, everyone is super charged up.

他们实际发多少……反过来说,我更大的麻烦是得劝员工去休息,别再生成了、别再想点子了,因为所有人都太上头了。

So, for example, well, Sultan, he really does one clip a day.

比如 Sultan,他真的是一天一条片子。

I mean one clip that pulls millions of views, he makes it in a day.

一条能拿到几百万播放的片子,他一天就做出来。

And here's what's interesting, there's Dor Brothers.

有意思的是,有个 Dor Brothers。

That's the world's top AI production.

全世界最顶的 AI production 就是他们。

Joe Rogan talked about them on his podcast for quite a while, praised them a lot, like these guys are insane.

Joe Rogan 在播客里讲了他们好一阵,把他们夸得不行,说这帮人是疯子。

And Dor Brothers wrote to us themselves, invited us onto a call to talk.

结果 Dor Brothers 主动来找我们,约我们开个电话会聊聊。

And when we were on the call, they, uh, and this was really nice, they, uh, told us that we're number one in vibe marketing on Twitter.

通话时他们说——这话我们特别受用——说我们是 Twitter 上 vibe marketing 的第一名。

I mean we got, well, respect from the most top-tier guys, and they called us top-1 on Twitter in vibe marketing.

也就是说,我们拿到了最顶尖那批人的认可,他们说我们是 Twitter 上 vibe marketing 的 top-1。

So.

就这样。

Then, by the way, they tried to poach Sultan.

顺带一提,后来他们想把 Sultan 挖走。

Well, that's a global thing overall, I mean it's also a problem, since everyone became superstars.

这其实是个全球性的问题,大家都成了超级明星之后就有这麻烦。

A lot of American companies try to poach our guys from us.

很多美国公司想从我们这儿挖人。

For example, take Kairym , uh, the one from Oze , our creative guy, he's a director there, Pika wrote to him, like they tried to get in touch with him somehow and so on, because he also made a very viral clip, and, well, and obviously they go after the engineers and so on.

比如 Kairym ,从 Oze 过来的,我们的创意,是个导演,Pika 就给他写过信,想办法联系他之类的,因为他也做过一条特别火的片子;工程师那边当然也一样。

Yeah, that's another sharp problem we have, that, well, the guys are very good and competitors write to them and so on.

对,这也是我们一个很扎手的问题——人太强了,竞对天天给他们写信。

Let's double-click on this.

咱们就这件事往下 double-click。

So let's imagine that day when they flew, when she flew into space, uh, Bezos's wife, his fiancée, Lauren Sanchez, and her girlfriends.

想象一下那天,Bezos 的妻子、当时的未婚妻 Lauren Sanchez 和她的闺蜜们飞上太空。

And so you guys there, it becomes a meme there, how Vika Katy Perry now landed back on earth, kissed the ground and all the rest.

然后这事就成了梗,Vika Katy Perry 落地之后亲吻大地什么的。

So tell me, like, you write the script,

你讲讲,你们怎么写脚本的,

So tell me about how the script gets written afterwards, and, I don't know, like what steps you go through next, how this script turns into a storyboard, right, some kind of storyboard, probably, and then you describe it, and for them it's a picture plus a prompt, and then you glue it all together.

讲讲剧本后面是怎么写的，还有，我也不清楚，你们接下来走哪几步，这个剧本怎么变成分镜，对吧，大概是某种分镜，然后你们把它描述出来，给到那边的就是一张图加一个 prompt，最后再把这些拼起来。

What does that look like?

具体是什么样的？

Yeah, yeah.

对，对。

Well, basically, yeah, the generation of consistent content, we do it, uh, with the help of a storyboard, as for what to say now about this specific case garbled ASR.

嗯，基本上是这样，一致性内容的生成，我们靠分镜来做，至于这个具体案例现在要怎么讲 ASR 含混。

What was that one of yours, what was the idea there?

你们那条当时是个什么东西，想法是什么？

Do you remember that clip?

你还记得那条片子吗？

I don't remember anymore, it was a long time ago, but yeah, at the start you need to, well, get things rolling, and well, obviously, we have a lot of creative guys in the company, and we just start riffing, throwing out all sorts of crazy ideas.

我已经记不清了，太久之前了，不过对，一开始得先把想法带起来，我们公司里搞创意的人特别多，大家就开始互相激，往外抛各种疯狂的点子。

Most of them you can't generate, because, well, it's just inappropriate.

大部分是没法生成的，因为那些东西根本不合适。

Well yeah, yeah, they'd probably just ban us on Twitter, yeah.

对对，很可能我们直接就被 Twitter 封了。

And yeah, then the problem is to filter all of it, to find some more or less, well, content that is, well, obviously, maybe on the edge, but so that it could still exist on Twitter.

然后问题就是把这一堆全筛一遍，挑出多少还过得去的内容，那种可能踩在边缘上、但至少能在 Twitter 上活下来的。

And then it's very important to formulate exactly why this is going to be viral.

接下来特别重要的一步，是把它为什么会爆讲清楚。

Like, there has to be a really clear message.

必须有一条非常明确的 message。

We have to understand, uh, why this went viral.

我们得搞明白，它到底为什么会爆。

And when we understand for sure that this one, well, basically has to hit, then, yeah, we start the generation.

等我们确定这条肯定能打中，那就开始生成。

And, well, often it works the first time, but often you need to finish it off with some post-production.

很多时候第一遍就成了，但也经常要再补一轮后期。

For example, we generated a video there.

比如说，我们生成了一条视频。

So this was Danil's idea, he's a professional cinematographer.

这个点子是 Danil 出的，他是职业电影摄影师。

And he decided, well, to imagine the idea of what if Tarkovsky used modern AI tools.

他决定设想一下：要是 Tarkovsky 用上现在的 AI 工具会怎么样。

And obviously, well, Tarkovsky is that for the industry, but for Danil himself, for Danil himself.

当然，Tarkovsky 对整个行业来说就是那种存在，但对 Danil 本人来说，对他本人来说——

He applied to VGIK, I think.

他好像考过 VGIK。

Well, I mean for him he's an icon, a great man.

对他而言那就是偶像，是伟人。

An icon.

偶像。

And why not, yeah, imagine if he used AI tools, modern ones.

那为什么不设想一下呢，他要是用现在这些 AI 工具会怎样。

And so he made this clip himself, we posted it, but it didn't get that many views.

他自己把这条片子做了出来，我们发了，但播放量没多少。

And then, well, we had to think about how to, well, make it get a lot of views.

然后就得琢磨，怎么让它跑出大量播放。

And there was an idea, well, we decided to just, uh, stick the real footage right next to the AI generation, so that people would actually see the contrast, the difference.

当时有个想法，我们决定直接把实拍镜头贴在 AI 生成的旁边，让人一眼就看到反差、看到区别。

And when we did that, it just started going viral, well, obviously, first on Twitter it immediately got 11 million views in a couple of days, and then it went all around the world.

这么一改，它就开始疯传，先是在 Twitter 上，几天就冲到 1100 万播放，接着传遍了全世界。

And the media started posting it and, well, everywhere.

媒体也开始转，到处都是。

And why did this video become popular, in your view?

在你看来，这条视频为什么会火？

Well, how do you explain it?

这该怎么解释？

Uh, right.

嗯，是这样。

I mean, it's no secret that generative AI gets a ton of hate from people, and a lot of, well, obviously, objective hate there.

这不是什么秘密：生成式 AI 挨了非常多的骂，其中很多骂，明摆着是有客观道理的。

It, well, changes the industry a lot, changes, well, the modern world a lot.

它确实把这个行业搅动得很厉害，也把当今世界搅动得很厉害。

And, uh, well, we hit that sore spot of people's, in that for some of them it was of course the opposite, it was like wow, but for some, well, even if the generation comes out well, even if it comes out cool, interesting, fresh, well, people, if they find out that it's AI, then they, well, immediately start hating.

我们正好戳中了大家的这个痛点：对一部分人来说恰恰相反，那是「哇」的感觉；但对另一部分人，哪怕生成得很好，哪怕做得酷、有意思、新鲜，只要知道这是 AI，立刻就开始骂。

So.

就是这样。

And, well, if people, for example, don't find out that it's AI, they might, like, they might like it, it's cool, it's made well, high quality.

反过来，如果大家不知道这是 AI，可能就会喜欢：挺酷，做得讲究，质量高。

So our camera movements, they work really well, like, well, because Danil, he shot all of it himself, and he, well, clearly understands when it works and when it doesn't.

我们的运镜就做得非常扎实，因为 Danil 这些东西全是自己拍出来的，他很清楚什么时候管用、什么时候不管用。

So.

就是这样。

And so are there already some patterns, like one way to go viral is to make people not indifferent, some kind of, well, rage, right, for example, envy, jealousy.

那已经总结出一些规律了吗，比如说，爆的一条路是让人没法无动于衷：愤怒，对吧，或者羡慕、嫉妒。

What else?

还有呢？

What other feelings trigger virality?

还有哪些情绪能带来病毒式传播？

Well, of course, these feelings are the, like, fairly simple way, probably.

当然，这几种情绪大概是最省事的一条路。

And also when something is made really well, that can also go viral nicely.

另外，东西做得特别精良，也一样能爆。

Uh, because it surprises the audience.

因为它让观众感到意外。

Yeah, yeah, it surprises you when it's made with quality.

对对，做得够精良就会让人意外。

And, for example, uh, one creator, he actually shot a clip where he, uh, shot a real car chase, I mean he mixed real footage with our special effects, well, obviously it's expensive to shoot a car professionally with an FPV drone, with a car grip there, and to do the vehicle rigs, and also the VFX explosions, collisions.

比如有个创作者，他真拍了一条片子，拍的是真实的汽车追逐，就是把实拍和我们的特效混在一起——用 FPV 无人机专业地拍车、上车载摄影支架、搞那些车辆装置，再加上 VFX 的爆炸和碰撞，这些都很贵。

And he, I mean, he shot the frames that are fairly simple to shoot.

他只拍了那些相对好拍的镜头。

Well, he shot it himself, I mean he's, well, a professional director who makes films.

都是他自己拍的，他本来就是拍电影的职业导演。

And so he made a really high-quality clip and it went super viral on Instagram.

他做出来的片子质量非常高，在 Instagram 上一下就爆了。

So there are cases like that too.

这样的案例也有。

Our announcement videos, they also go viral well on social media, because they're made really well by our, uh, creative guy Ziya, he directs them, the way he used to direct music videos, and he did music videos, well, the most popular music videos — like Ivan Urgant, for example, shot a parody of a video that Ziya made, for instance.

我们的发布视频在社交平台上也传得很好，因为做得非常讲究——我们的创意负责人 Ziya 来执导，跟他以前拍 MV 一个路子，他拍过那种最火的 MV，比如 Ivan Urgant 就照着 Ziya 拍的一支 MV 做过恶搞。

So.

就是这样。

And our announcement videos are at that level, they go super viral.

我们的发布视频就是这个水准，传播得非常猛。

You publish the same thing.

你们发的是同一个东西。

If you make one video, right, for example, do you publish it to three platforms at once, or do you adapt it to each platform separately?

比如你们做了一条视频，是直接同时发到三个平台，还是针对每个平台单独适配？

I mean TikTok, Instagram,

就是 TikTok、Instagram，

we, yeah, we of course adapt it, because the formats still have to be adjusted a bit.

我们当然会适配，格式总归得改一点。

And talking with Ziya, it turns out this is very similar to when he was at Oz в Озе, they had this, what's it called, label producing, production, something like that, I mean what he did exactly was, they recorded a huge number of artists there, and then you had to distribute their tracks across social media.

跟 Ziya 聊下来发现，这跟他在 Oz в Озе 的时候特别像，他们那边叫什么 label producing、production 之类的，他干的活就是：那边签下大量歌手录歌，然后要把这些歌分发到各个社交平台。

And that's a pretty huge, high-volume job, because there are a lot of tracks.

这活儿的体量非常大，因为歌太多了。

Each one has its own cover art, each one has its own messages.

每首都有自己的封面，自己的主打信息。

You have to drop it on all the social networks on the same day.

这些必须在同一天全平台一起放出来。

For YouTube it's one format, for TikTok another, for Instagram another.

YouTube 一个格式，TikTok 另一个，Instagram 又是另一个。

You need to prepare the covers.

封面都得提前做好。

For example, Ziya is also really good at graphic design because of that.

所以 Ziya 平面设计也玩得很溜。

I mean cool fonts, cool visuals, so that all of it, well, really looks cool.

字体要酷，视觉要酷，整套东西看上去得真够劲。

So.

就是这样。

Well, our competitors don't have anything even close to that kind of visual at all, yeah.

我们的竞品在视觉上根本连边都摸不着。

So you managed to put together this superstar team that seems to bring together, let's say, professionals who in ordinary life, in many companies, don't cross paths, right?

所以你们攒出了一支全明星团队，把那些平时在很多公司里根本碰不到一块的专业人士聚到了一起，对吧？

I mean creative, yeah, Hollywood, Silicon Valley, product, marketing, everything all in one.

创意，对，好莱坞、硅谷、产品、市场，全都塞在一起。

Obviously Pika is this Stanford kind of thing, engineers, uh, who arguably might have problems with taste.

很明显，Pika 就是那种斯坦福路数，一帮工程师，审美上可能多少有点问题。

Yeah.

对。

Yeah.

对。

I mean, well, globally all the competitors have big problems with taste.

整体上看，所有竞品在审美上都有很大问题。

very big ones.

非常大的问题。

I mean even, well, OpenAI, go into their Sora Instagram.

就连 OpenAI——你去他们 Sora 的 Instagram 上看看。

Well, everything there is going to look very sad, just sad.

那上面的东西看着都特别惨，就是惨。

And yeah, and the funny thing is that everyone sits in one office, in one room, in one open space.

还有个有意思的点：所有人坐在同一个办公室、同一个房间、同一片 open space 里。

And somehow, well, in general I always understood this, well, then I just confirmed it in practice.

这一点我其实一直都懂，后来只是在实践里验证了一遍。

I always understood that, uh, people who are super creative, creative, who are, uh, well, pros at their craft, I mean it doesn't matter, whether it's a great machine learning engineer, a great backend programmer, a great mathematician, a great fashion designer, a great director.

我一直明白一件事：那些超级有创造力的人，那些在自己这行做到专业的人——不管是厉害的机器学习工程师、厉害的后端程序员、厉害的数学家、厉害的时装设计师，还是厉害的导演。

All the people, well, who adore their field, who are like fixated, obsessed with their field, they are all, on a meta level, they're all very similar to each other, that's what unites them.

所有热爱自己这行、近乎走火入魔、对自己这行着迷的人，在元层面上其实彼此非常像，这就是把他们连起来的东西。

And this, if you've seen the meme, one of my very favorite memes, it's people who make X and people who use X.

如果你看过那个梗图——我最喜欢的梗图之一——就是「做 X 的人」和「用 X 的人」。

Well X could be fashion design, it could be AI, like, well, the kind of people who make X.

X 可以是时装设计，也可以是 AI，反正就是那种做 X 的人。

Well, you can drop this meme in later.

你们后期可以把这个梗图插进来。

That one, that's that Gaussian line thing.

那个，是不是那张高斯直线的图。

No, it's not Gauss.

不，不是高斯那个。

It's, uh, the Harry Potter actor on top, I think, and, well, or some other actor.

是那种，上面好像是 Harry Potter 的演员，或者别的演员。

And at the bottom they're like, I mean at the top there's this wrecked guy lying there all torn up.

下面是那种——不对，上面是一个被折腾得不成样子的人，衣衫褴褛地瘫在那儿。

Well, you can see that he works a whole lot, is really obsessed with his craft.

一看就知道他干活干得极多，对自己这行极度着迷。

If it's some designer, he looks really bad, because, well, he doesn't, well, he himself doesn't consume what he creates.

如果他是个设计师，那他自己会显得很邋遢，因为他并不消费自己做出来的东西。

Right.

对。

And the people who use it, they're these, well, coolly dressed, stylish people.

而使用这些东西的人，一个个穿得又酷又时髦。

I mean, I don't know, the ones who write machine learning libraries are just these, well, maximally nerdy-looking programmers.

比方说，写机器学习库的那批人，看上去就是极度技术宅的程序员。

And the people who use machine learning libraries are these stylish ones who work in the corporate world.

而用机器学习库的那批人，都是在大公司上班的时髦人。

There's a huge difference, what you're saying, between the viewer and the participant, the creator and the user.

你说的这个差别非常大：观众和参与者之间，创造者和使用者之间。

Yeah.

对。

Yeah.

对。

Well, this is probably connected to work-life balance, in general, to things like that in the end, that people who are obsessed with their industry are actually very similar to each other.

这大概跟 work-life balance 这类事情有关：那些痴迷于自己行业的人，其实彼此非常像。

And the real chemistry happens when our ML engineers are sitting there, the ones who want to build a really cool model.

而真正的化学反应发生在这里：我们的机器学习工程师坐在那儿，想做出一个特别牛的模型。

And they also understand that, well, a cool model, if it's video, has to generate aesthetically cool stuff.

他们也明白，牛的模型如果是视频模型，生成出来的东西在美学上得够酷。

And they sit there together with the creative team, uh, together.

他们就跟创意团队坐在一起，坐在一块儿。

And it just happens for them, I mean they explain to them how it works in the industry, how it's supposed to look.

然后事情就自然发生了：搞创意的人给他们讲行业里是怎么运作的，东西该长成什么样。

That, well, it's not just some random video, there are rules there, some even academic ones, some cultural ones.

这不是随手拍的一条视频，里面是有规则的，有些甚至是学院派的，有些是文化层面的。

I mean, if we're talking about music videos, that's the culture of the last 30 years, the nineties, the two-thousands and the present day.

拿 MV 来说，那是过去 30 年的文化，九十年代、两千年代，一直到当下。

There's a completely huge cultural layer there, how the industry developed from music videos on TV all the way to cloud rappers who made videos on the cheap.

这里面有一整层巨大的文化积累，行业怎么从电视上的 MV 一路走到 cloud rap 歌手用最土的办法糊出片子。

And then all of it, well, this whole cultural layer moves over into AI, and, well, AI is made by machine learning engineers.

接下来这整层文化都迁移进了 AI，而 AI 是机器学习工程师做出来的。

I mean there's no way it passes them by.

所以这些东西不可能绕过他们。

Because they train all of it, they make it.

因为都是他们在训练、在做。

And so literally, yeah, chemistry happens.

所以真就是字面意义上的化学反应。

I, well, I see it every day.

我每天都能看见。

It's just, well, mind-blowing.

太震撼了。

So here's one of your great qualities, right, I mean, uh, this one, the one that developed in you, right?

你身上有个特别棒的特质，是后来慢慢长出来的，对吧？

I mean, originally you were a terrific engineer, right, the author of really cool, well-known libraries.

你最早是个非常出色的工程师，写过几个很牛的知名库。

I remember how back in 2018 Schulman was already interviewing you, right, from OpenAI.

我记得 2018 年 Schulman 就面试过你了，OpenAI 的那位。

So.

就这样。

And it's the ability, uh, to ask the right questions during an interview, to figure out whether the person across from you is a manager or a practitioner who can actually do it with his own hands.

这个特质就是：面试时问对问题，判断坐在你对面的是个管理者，还是真能自己动手干的实干派。

Tell me, please, about this approach.

讲讲你这套打法吧。

How do you tell talent from, uh, I mean from non-talent?

你怎么把有才的人跟没才的人分出来？

Right.

对。

Yeah.

对。

I mean, yeah, this is super important.

这事儿超级重要。

I mean, uh, being able to, well, interview people properly — a lot, well, goes on there.

会不会正确地面试人——这里头门道特别多。

Uh, I mean it's like, well, an interview, it, well, when I interview, it consists of two parts.

我面人的时候，一场面试分成两部分。

The first part is, yeah, I want to figure out whether the person is a practitioner or more of a consumer.

第一部分是，我想弄清楚这人是实干派，还是更像个消费者。

Uh, and usually that starts becoming clear when you just go into really deep long-tails and and well you can tell from the person how he talks about them.

通常只要往特别深的长尾细节里扎，看他怎么讲这些东西，就能看出来。

Uh, usually you just kind of feel it.

一般就是能感觉出来。

Uh and, well, for most of the interview, probably, I'm actually trying to talk people out of working with us.

而面试的大部分时间，我其实是在劝人别来我们这儿干。

That's something I've just started noticing lately, when I interview people and there, obviously, right now in general, well, when we came out of stealth, the product started blowing up, a huge number and a huge inbound, everybody wants, well, to join us, and I, well, I really just talk people out of it.

这是我最近才注意到的：我面人的时候——现在情况摆在那儿，我们从 stealth 出来之后产品直接炸了，量特别大，inbound 也特别大，所有人都想加入我们——而我是真的在劝退人。

I mean, I tell them all the downsides of our job.

我会把我们这份工作的所有坏处都摆给他看。

Tell me about those downsides, yeah, because, well, you just, well, you also don't want to ruin somebody's life if the person doesn't, well, fit.

讲讲这些坏处吧，毕竟要是这人不 fit，你也不想去毁人家的人生。

I mean, well, so, like I said, it's a seven-day week, and in some cases a six-day week, and work with no schedule at all, completely absent, because, well, anything at all can happen, and competitors might release something, or something might situationally happen in the world that you have to react to.

就像我说的，一周七天，有些情况是六天，而且完全没有作息表，因为什么事都可能发生——竞争对手可能突然发个版，世界上也可能突然出点什么事，你必须反应。

And plus uh the things we want to deliver, they, well, if competitors take like 2-3 months to deliver them, we deliver them in two, like, a week and a half.

再加上我们想交付的东西，竞争对手要花 2-3 个月才交付的，我们两周、一周半就交付了。

I mean, and working in that mode, uh, well, it's a very stressful job.

在这种节奏下工作，这是压力非常大的活儿。

And to people, well, I really try to explain that this is a very stressful job.

我会明确跟人讲清楚：这是一份压力非常大的工作。

Plus, uh, we try as hard as we can to get away from any kind of bureaucratic, corporate stuff.

另外，我们尽最大努力避开任何官僚的、大公司式的东西。

So if a person is used to that, if he worked somewhere in a corporation, then most likely it'll be really hard for him.

所以如果一个人习惯了那套，如果他在大公司待过，那他多半会非常难受。

I mean we try to minimize any kind of, well, bureaucratic written communication, so that, well, I mean corporations do more of that, uh, I mean with us, obviously, everything gets tracked, we keep, uh, all sorts of Notions, trackers and so on, but that's people arranging it among themselves.

我们尽量把官僚式的书面沟通压到最低——这类东西大公司做得更多；我们这儿当然也都有追踪，各种 Notion、tracker 都在用，但那是大家自己之间约好的。

It's not like we have some process that has to run, everybody is required to mark things in the calendar somehow, to track tasks.

不存在什么必须走的流程、所有人必须在日历上标记、必须 track 任务。

We don't have that at all, basically.

这种东西我们压根就没有。

Everybody has to navigate on his own, uh, as an individual contributor.

每个人得自己找方向，作为个人贡献者。

If, in order to close your task, you need to, well, coordinate people, and you own the task, then we don't have that corporate hierarchy where there's, we have no managers at all in the company, basically.

如果为了把你的任务关掉，你需要协调别人，而任务归你 own，那我们这儿没有那套大公司层级——我们公司根本就没有管理者。

And the person has to coordinate it himself, go to the other person.

得他自己去协调，自己去找那个人。

And, for example, a corporation works differently.

而大公司是另一套玩法。

If you tell a person: "Do this task". and he goes to the other guys, and the other guys, well, most likely, him, well, there'll be some little bottlenecks, they'll say: "I can't, I'm busy with other stuff".

你跟一个人说：“把这个任务做了”，他去找别的同事，别的同事多半会卡他一下，说：“我做不了，我手上有别的事”。

And usually, well, in corporations the person who was given the task, he'll come back and say: "Damn, I can't do this task, because this other person is blocking me, he's the one at fault, let's sort it out, let's prioritize differently, let's, I don't know, put together a committee, right, let's all get on a call". something like that.

通常在大公司里，被派任务的那个人会回来说：“靠，这任务我做不了，因为有人把我卡住了，是他的问题，咱们捋一捋吧，重新排一下优先级吧，要不组个委员会，对吧，大家开个会”，差不多这种。

With us, well, we don't have that.

我们这儿没有这一套。

We have, and the person, well, he owns the task, it's kind of on his responsibility to push all the people so that it gets done.

在我们这儿，谁 own 这个任务，就由谁负责去推动所有人，把事做成。

He, well, we don't have this thing where you can go and complain.

我们这儿没有那种你可以跑去告状的事。

I didn't get it done because somebody blocked me.

“我没做完，因为有人把我卡住了。”

So, well, our guys, well, that's why a lot of our guys, they can cover a lot of breadth, because, well, as an individual contributor, starting with, like I was telling you, the creatives, who can do the motion design themselves, edit it themselves, generate everything themselves.

所以我们很多同事能覆盖很宽的面，因为作为个人贡献者——就像我讲过的，创意团队那帮人自己能做 motion design、自己剪辑、自己把素材全生成出来。

Same with the engineers, who can spin up the backend and everything fast.

工程师也一样，能自己把 backend 搭起来，什么都快。

Or else you have to push the person, explain to him why it's really important.

要不然你就得去推那个人，跟他解释为什么这事儿非常重要。

You have to agree on priorities yourselves.

优先级得你们自己商量出来。

Uh, I mean people have to agree among themselves in real time and deliver fast and, well, and understand the global context of the company, what matters and what doesn't.

人得 real time 自己商量、快速交付，还要理解公司的全局背景——什么重要，什么不重要。

I mean it's not like somebody has to come up to me and say: "So, Yerzat, what do you think, what's more important, do we ship this feature or not?"

不存在谁得跑来问我：“Yerzat，你觉得哪个更重要，这个 feature 我们发不发？”

I mean our guys kind of figure out for themselves what matters right now and what doesn't.

我们的同事自己就明白眼下什么重要、什么不重要。

And so, well, for some people that can be very stressful, because from the outside it just looks like chaos, well, chaotic processes.

对某些人来说这压力会很大，因为从外面看就是一团混乱，流程都是乱的。

Plus, obviously, there can be, uh, some conflicts and you need to, well, resolve all of it without conflicts.

另外显然也会有冲突，而这些都得不起冲突地解决掉。

And so, well, the job is, yeah, really, I mean, stressful.

所以这活儿确实是有压力的。

You reminded me, you know, there was this ad from 100 years ago.

你让我想起来，知道吗，一百年前有过一则广告。

Men wanted for hazardous journey.

招人，参加危险旅程。

Low pay, bitter cold, long hours of complete darkness.

报酬低，酷寒刺骨，长时间处在完全的黑暗里。

Return home doubtful, honour and recognition in case of success.

能否回家未知；若成功，可得荣誉与认可。

Yeah, yeah, literally.

对，对，字面意思就是这样。

Well, the low pay is the only thing that isn't true for us.

不过报酬低这一条，在我们这儿不成立。

I mean with you, with you it's all true, except, well, the pay is high.

也就是说在你们那儿全都是真的，只有报酬是高的。

Yeah, yeah, yeah.

对对对。

Tell me how you raise, let's say, these red flags during an interview, that the person won't fit you speed-wise, I mean he's way too, too used to a slower pace.

讲讲你在面试里怎么把这些红旗立起来——判断这人在速度上跟不上你们，就是他太习惯那种慢节奏了。

Do you have any special questions specifically,

你有没有什么专门的问题，

well, I don't have a specific list.

固定的清单我是没有的。

How do you tell?

那你怎么判断？

I just, well, I usually get it from free-flowing conversation.

我一般在随便聊天的过程里就看出来了。

I mean for me it happens informally, I haven't formalized it.

对我来说这是非正式的，我没把它形式化过。

I've just, probably, done a lot of these interviews and plus, well, then I had a sample, well, and at other companies I also did interviews, and I just have a sample I trained myself on, so I can tell.

大概是这类面试我做得太多了，再加上后来手里攒了个样本——在别的公司我也做面试——我就有了一个用来训练自己的样本，靠它来判断。

And it's kind of, well, and it somehow got learned in a way that it's cross-domain.

而且这东西不知怎么就学成了跨领域的。

But, for example, I saw how Almaz interviews, and it's, well, I thought that I'm the one who interviews harshly.

不过比如说，我见过 Almaz 怎么面人，我原本以为我面得已经够狠了。

But the way he interviews, he turns out to be way harsher than me.

结果他面起来比我狠多了。

I was just in shock.

我当场就震住了。

He just, well, he's got questions like this.

他的问题是这样的。

So the interviewer comes in, well, the interviewee, well, the one being interviewed, uh, and Almaz just asks him these really sharp questions.

面试者进来——我是说被面的那个人——Almaz 直接甩给他非常锋利的问题。

For example, imagine you've been working 7-8 days straight.

比如：想象你连着干了 7-8 天。

Mm-hm.

嗯。

At least 12 hours each.

每天至少 12 小时。

Mm-hm.

嗯。

Then you were promised that after this we ship the release, we hit the goal, and you can rest for one day, and you've already arranged it with your loved ones, with your family, that you'll come over, you'll, I don't know, go somewhere, I don't know, you'll set a table, like, well, like with us Kazakhs there's always got to be some spread, you'll all sit together with your loved ones, you've got it planned, or you, I don't know, you dreamed about some concert, bought a ticket.

然后有人答应你：这波之后我们就发版、达成目标，你可以休一天；你已经跟家人、跟亲戚约好了，说你会过去，你们要一起去个什么地方，或者摆一桌——我们哈萨克人嘛，必须得摆一桌，一家人坐在一起——你都安排好了；或者你一直想去某个演唱会，票都买了。

So imagine that.

你想象一下。

And on top of that you delivered everything.

而且你把活儿全干完了。

I mean you were promised that you'd take that one day.

人家答应过你，你可以休那一天。

This is Almaz asking.

这是 Almaz 在问。

But on that day, and so you finished the work, you hadn't slept, you'd slept only like 3 hours, 4 hours.

但就在这一天——你刚把活干完，你没睡，总共就睡了 3 小时、4 小时。

And on that day our competitors ship a release — your move.

就在这一天，我们的竞争对手发版了——你怎么做。

That's the kind of question it is.

就是这种问题。

Awesome, right.

牛，确实。

And well it's really interesting to watch the reaction, because I later sat in on several interviews where Almaz interviews, and there really, well, in the first 5 seconds you can basically just tell from the person's reaction whether he fits or not, because, well, with people, with a lot of them there's just shock on their face and you immediately see there's no point going further.

看反应特别有意思，因为我后来旁听过好几场 Almaz 主面的面试，真的，头 5 秒钟从人的反应基本就能判断他合不合适，因为很多人脸上就是懵的，一眼就看得出后面没必要谈了。

But why ruin a person's life if, well, he doesn't fit?

既然不合适，何必去毁人家的人生？

And there are, of course, well, people who without any hesitation at all immediately say: "Well, I'm a team player, I understand that, well, we want to win. And I mean, if everybody stays, I stay too".

当然也有人完全不带犹豫、当场就说：“我是团队型的，我明白我们是要赢的。所有人都留下，我也留下”。

So.

就这样。

And and he, and it's not the only one like that, he just asks a large number of such questions back to back, really harsh ones that shock you.

而且这不是唯一一道这种题，他会一连串抛出大量这种非常狠、非常让人震住的问题。

And and it's not that he essentially just made them up, these are real situations.

而且这些基本上不是他编出来的，都是真实发生过的情况。

He literally just describes real situations.

他就是在原样复述真实发生的事。

And this specific situation — that's us delivering a product, we did absolutely everything that we

具体这个情况，就是我们当时在交付产品，把我们

we wanted to, we celebrated.

想庆祝，就庆祝了。

So how do you guys celebrate, by the way, tell me.

对了，你们怎么庆祝的，讲讲。

So we have a tradition, well, like all Kazakhs probably do,

我们有个传统，估计所有哈萨克人都这样，

like, a sheep, you slaughter a sheep there?

宰羊？你们那儿宰羊？

Salzhyk name garbled does it himself, right there.

就是 Salzhyk人名存疑 亲手弄的。

He— does it, like, happen right there in the office?

他——这是直接在办公室里干的？

No, not in the office, obviously.

不不，当然不在办公室。

But we've got Salzhyk name garbled, our ML lead.

但我们有 Salzhyk人名存疑，我们的 ML lead。

He, he does it, he catches the sheep and slaughters it himself for you.

他，他来弄，他抓羊，亲自给你宰。

Well, not himself, obviously, he organizes it.

当然也不是他亲自动手，是他张罗。

It was honestly hilarious, when on the first day of the release, so we shipped the release,

太逗了，release 第一天，我们刚 release 出去，

Uh-huh.

嗯。

everybody's like applauding.

所有人都在鼓掌。

We were, like, prepping the product there, just working.

我们当时在打磨产品，就是干活。

So, the very first release, March 31.

最早那次 release，3 月 31 号。

Everybody's like, "We stand up, we applaud."

所有人都是：“我们起立，鼓掌。”

Obviously, the guys are creative, they also love filming themselves there, for the archive.

毕竟这帮人是搞创意的，还爱拍自己，留个档。

They bring in a sheep, and it— there's just, like, a crowd of people at our place there.

把羊抬进来，我们那儿乌泱泱一堆人。

No, not a live one, it's already all butchered, everything, basically.

不是活的，已经全分好了。

We sit down.

坐下。

The dastarkhan is laid out in the office.

办公室里摆开了一桌 dastarkhan。

Yeah, in the office.

对，在办公室。

Uh-huh.

嗯。

And then I'm, like, looking over, our Kuka, uh, DBK garbled, the engineer who's, well, in charge of our whole product infrastructure, he's sitting there typing something.

然后我一看，我们的 Kuka、DBK听不清，就是管我们整个产品基础设施的那个工程师，坐在那儿敲着什么。

I run up to him, I say, "What happened?"

我跑过去问：“出什么事了？”

He says, "Damn, found a bug, I'm fixing it, and we're already celebrating everywhere, basically."

他说：“操，发现个 bug，我在修，结果我们已经到处庆祝上了。”

And so, well, I'll never forget that picture, him sitting there.

那个画面我永远忘不了，他就那么坐着。

It later became a meme, that we're, like, shipping something, everyone's already started celebrating.

后来这成了个梗：我们一发东西，所有人就开始庆祝。

And Kuka's sitting there like, well, at his computer, like, alone in the corner.

而 Kuka 一个人坐在角落的电脑前。

And it turns out, basically, he's already fixing a bug there.

结果他早就在那儿修 bug 了。

And yeah, why I was telling this, how we, yeah, so, I mean we're celebrating, we'd already celebrated and out comes the GPT API, a big thing, well, a really powerful feature, I mean the one that, yeah, Image, and we have video generation, our own model, and we have an image generation model, but at the moment we don't have a model, we'll soon have our own, which can, from a prompt, take an already finished image that we generate

对，我讲这个是想说，我们在庆祝，庆祝完了，GPT API 就出来了，大事，一个真正强的 feature，就是那个，对，Image，我们做的是视频生成，自己的模型，我们也有图像生成模型，但目前还没有那种模型，很快会有我们自己的，能按 prompt 把我们已经生成好的图

You can edit it further, I mean it adds some— our Higgsfield model, it can generate a really beautiful shot.

可以再继续编辑，就是加一些——我们的 Higgsfield 模型能生成非常漂亮的镜头。

For example, OpenAI can't generate a shot like that, but on the other hand their model can edit it with a prompt, if you need to add something, some details and so on.

比如 OpenAI 生成不出这种镜头，但反过来，他们的模型能用 prompt 去改它，需要加点什么、加些细节之类的。

That's a pretty, well, major feature.

这算是个挺大的 feature。

Well, it was, it was viral, the Ghiblification of everything, right.

当时特别火，全世界都在 Ghibli 化。

And so it comes out, and they, Open— we'd already celebrated everything there, we'd shipped, and we all just get together like, what do we do.

它一出来，他们，Open——我们这边已经庆祝完了，也 release 完了，大家就凑到一起：怎么办。

We understand that now all the competitors are going to integrate it.

我们清楚，接下来所有竞品都会去集成它。

And obviously, well, all the guys say, "Damn, we have to be the first in the world to integrate OpenAI."

然后很明显，兄弟们都说：“操，我们必须做全世界第一个集成 OpenAI 的。”

And, well, everyone just, well, we'd celebrated there, everyone had worked long hours there, everyone sits back down and starts coding and there and, well, and obviously everyone's a super-optimist like, "Yeah, we'll integrate it in 5 hours, we'll all head home at 5:00 in the morning."

大家刚庆祝完，之前已经连轴干了很多小时，所有人又坐回去开始写代码，而且个个都是超级乐观主义者：“行，5 小时就集成完，凌晨 5:00 都能散了。”

In the end it's already lunchtime, and we're still integrating.

结果到了午饭点，还在集成。

In the end everyone's literally already going to sleep, well, there in the office in shifts, we've got sleeping bags in the office.

结果所有人真的开始睡了，就在办公室轮着睡，我们办公室里备着睡袋。

Everyone sleeps there, uh, in in shifts.

大家就在那儿睡，轮着来。

Well, and in the end, yeah, we were the first in the world to integrate OpenAI there and we pulled it off.

最后，对，我们是全世界第一个集成 OpenAI 的，做成了。

I mean it's, well, everybody's work.

这是所有人一起干出来的。

The creative team already started making a video on how to present it right.

创意团队马上开始做片子，琢磨怎么把它讲对。

The product engineers already started building the product, the ML guys started testing out the capabilities there, the prompt engineers started writing prompts there so it would work stably.

产品工程师开始做产品，ML 那帮人开始试各种能力，prompt 工程师开始写 prompt，让它跑得稳。

So I mean everyone in the company is mixed up in this.

就是说，公司里每个人都卷进来了。

And, well, so, yeah, we published first, our video went viral, one of, I think, the most-viewed announcement videos we've ever had overall, that's this OpenAI integration, even though the feature, well, the one that, it really was built there in less than 24 hours, but in the end it brought us huge reach, which, well, was then strongly reflected in revenue временю garbled.

对，我们第一个发出去，片子爆了，大概是我们所有 announcement 视频里播放量最高的之一，就是这个 OpenAI 集成，虽然这个 feature 真是不到一天做出来的，但它最后给我们带来了非常大的曝光，后来也明显反映到了 revenue 上временю 含糊。

[music]

[音乐]

So basically, I upload uh my own picture, right, I upload a prompt for how I'd like it edited, like in Ghibli style, in Pixar style, Simpsons, Rick and Morty, Lego, Minecraft, uh, and I get a video after that, right, as the output.

也就是说，我上传自己的图片，对，再上传一个 prompt，说我想怎么改它，比如 Ghibli 风格、Pixar 风格、Simpsons、Rick and Morty、Lego、Minecraft，然后输出拿到一个视频，对吧。

Yes, yes.

对，对。

Great.

太好了。

Oh, a really cool story.

哦，这故事真酷。

Do you have any other, like, comp— traditions at the company, besides Besides the fact that you eat a sheep, right, during a release?

你们公司还有别的传统吗，除了——除了 release 的时候吃羊？

Well, I'd have to think back a bit.

这得想一想。

With us, well yeah, we don't really have that, well, because everyone's roughly of the same mindset, right, we have a really friendly atmosphere.

我们这儿倒没什么特别的，因为大家 mindset 差不多，所以气氛特别融洽。

For us, probably, well, the unusual thing is that once my wife came by and sat in our office and was very surprised.

要说不寻常的，大概是有次我老婆过来，在我们办公室坐了会儿，特别吃惊。

We have a lot of, well, obviously, everyone loves sports, especially our engineers, well, and the creatives, and they try there, well, obviously nobody has time for sports, so, like, here, for example, somebody brought a ball and the guys constantly love kicking a football around right in the office, and the ball might just fly into a comp—, into a monitor, well, it definitely looks chaotic.

我们这儿很多人——大家都爱运动，尤其是工程师，还有搞创意的那帮人，他们也想练，但谁都没时间运动，所以就有人拿了个球来，兄弟们老爱在办公室里踢球，球可能直接飞进电——飞进显示器，看着确实挺乱的。

And also we have, well, a lot of the engineers for some reason, well, I don't know why, love all kinds of martial arts, we constantly have all kinds of sparring there, someone's constantly boxing with someone else there and wrestling.

还有，我们不少工程师不知道为什么特别爱搏击，天天在那儿对练，老有人互相打拳、摔跤。

Is that, I mean are there actually gloves, huh?

这个——就是说真戴拳套？

Well, it's more like slap-tag.

更像是玩拍人游戏。

I mean it's cardio like that, it's, well, probably unusual, when you're in our office, someone, someone's constantly playing slap-tag there, because, well, yeah, nobody has time to do any sports outside the office.

就是那种有氧运动，在我们办公室待着大概会觉得挺怪，老有人在那儿玩拍人游戏，因为确实谁都没时间在办公室外面运动。

So, well,

就这样，

like at PayPal, there was wrestling there too, arm wrestling, right, arm wrestling matches between Max Levchin and Elon Musk,

就像 PayPal 那样，他们那儿也摔跤，掰手腕，对，Max Levchin 和 Elon Musk 的掰手腕比赛，

yeah, I didn't know that, cool, yeah.

是吗，我还真不知道，挺有意思。

So, well, something like that, well, some kind of physical activity, basically, the guys need it, because everyone's stuck in the office.

就是这类，大家总得有点体力活动，因为所有人都困在办公室里。

And, well, it probably looks just a little weird from the outside, because my wife once texts me, I was, well, somewhere not in the office, she was sitting in the office, and she writes, "There's a fight going on here."

这从外面看可能是有点怪，因为我老婆有次给我发消息，我当时不在办公室，她在办公室坐着，她说：“这儿有人在打架。”

And I'm, like, in shock.

我当时也懵了。

Well, I figure that, well, the guys do this all the time.

但我知道，这是他们天天在干的事。

So you have a lot of openings, right, you always have, you have openings for an ML engineer, a DevOps engineer, there are openings in the marketing department.

你们有很多在招的岗位，对吧，一直都有，ML 工程师、DevOps，市场部也有岗位。

Tell me what these openings are and what kind of people are a fit for you, what kind of people are not a fit for you.

讲讲这些岗位是什么，什么样的人适合你们，什么样的人不适合。

Yeah, well of course we, yeah, we need people who have a mindset similar to ours.

对，我们当然需要跟我们 mindset 相近的人。

I mean, well, it's okay if a person there, well, doesn't want to work in our style, so, well, we try as hard as we can, well, to explain to people what, well, is needed, so there'd be some kind of fit between us.

一个人不想按我们这种方式工作，也没问题，所以我们尽量把需要什么讲清楚，让双方之间有个 fit。

So.

就这样。

And so you'd take people with families, you're not against family people.

也就是说，有家庭的你们也要，不排斥有家庭的人。

We have a lot of family people.

我们这儿有家庭的特别多。

We have a lot of guys who have kids, who have babies there.

很多人有孩子，还有的孩子还是婴儿。

And family guys, well, I think it's the opposite, I mean family people, they're usually, well, very motivated, because there, obviously, well, the financial upside is very important, that, well, everyone wants, in the end, well, not just to become stars there and, well, everyone wants to make money, in the end and, well, and to become even more valuable professionals there on the market overall.

有家庭的——我觉得反而，有家庭的人通常动力特别足，因为财务上的 upside 对他们很重要，说到底大家想要的不只是成名，都想赚到钱，也想在市场上成为更值钱的专业人士。

So, yeah, we have quite a lot of guys with families and, well, I don't see a difference, any correlation there between family or non-family.

所以有家庭的兄弟我们这儿挺多，我看不出有家庭和没家庭有什么区别、什么相关性。

Maybe family people are even, like, more serious about a lot of things.

有家庭的人可能在很多事上反而更认真。

So.

就这样。

And uhm well as for openings, we're always looking for strong engineers.

岗位方面，我们一直在找强的工程师。

Uh back-end, front-end, uh infrastructure, uh we're looking for strong DevOps people.

后端、前端、基础设施，我们在找强的 DevOps。

Uh we, well, obviously, we're always looking for strong machine learning engineers.

我们当然一直在找强的 ML 工程师。

Uh we have, well, well, well, by now it's obvious, we have the strongest in the world.

现在已经很明显了，我们这支是全世界最强的。

Well, since, probably, the viewers should understand, we're top one in America.

观众大概得知道，我们在美国是 top one。

Video gene— generation startup.

视频生——视频生成创业公司。

Basically, and at the same time the market is mega-competitive, and our competitors raised rounds, some of them tens of times bigger than ours, but, well, some like five to seven times bigger than ours, and we beat them purely on speed, well, and on product quality.

而且这个市场竞争极其激烈，我们的竞品有的融的钱是我们的几十倍，有的是我们的五到七倍，我们纯靠速度、靠产品质量把他们打赢了。

So.

就这样。

And we're also looking, uh since we're, well, already becoming a real business, I mean we want to grow a lot as a business, because initially we were like a deep tech startup, we had a good valuation there, we raised good rounds as a deep tech startup, as, like, experts in machine learning, experts in video generation, and now we're, well, going to the next level, I mean we want to become some kind of successful business.

我们还在找人——因为我们正在变成一门真正的生意，我们想作为生意大幅增长，一开始我们是 deep tech 创业公司，估值不错，作为 deep tech 创业公司、作为机器学习和视频生成方面的专家融到了不错的轮次，现在我们要上一个台阶，想成为一门成功的生意。

So it's very important for us to put together a super-strong growth team, a growth department.

所以组建一支超强的增长团队、一个增长部门，对我们非常重要。

And we're actively looking for people right now.

我们现在在积极招人。

So, if you're an ambitious, uh, like, person who'd want to really grow specifically in distribution, uh, distribution is the foundation of any IT business, of any internet business — it's distribution.

所以，如果你有野心，想在 distribution 上真正长本事——distribution 是任何 IT 生意、任何互联网生意的根基，就是 distribution。

Any experienced founder will tell you that the first thing you need to think about is distribution.

任何有经验的创始人都会告诉你，第一件要想的事就是 distribution。

how you'll build distribution, and only second about the product.

怎么把 distribution 搭起来，产品只排第二。

Especially since products are pretty easy to build these days, like vibe coding, AI, so distribution is even more important.

尤其现在做产品挺容易的，vibe coding、AI 这些，distribution 就更重要了。

I mean if the cost of building some product is dropping very fast, then competition grows even stronger.

就是说，做一个产品的成本飞快下降，竞争就涨得更凶。

Those who can build products, and the competition there for proper distribution, for who knows how to do it, and for people who know how to build distribution, who understand how it works.

能做产品的人，还有对做对 distribution 的争夺、谁会做，对那些能搭起 distribution、懂它怎么运转的人的争夺。

It grows exponentially fast, just like the price of product development falls exponentially fast.

它在指数级上涨，就像产品开发的价格在指数级下跌一样。

So, if you're, well, ambitious, you want to grow in marketing, you want to buil—, well, build there, uh, plus, well, it's very advantageous to join us, because, well, we're already the top one AI video startup there, like I said.

所以，如果你有野心，想在市场营销上发展，想搭——想去做，另外，加入我们非常划算，因为我们已经是 top one 的 AI 视频创业公司了，像我刚说的。

I mean it's, well, you're taking on less risk.

也就是说，你承担的风险更小。

Not like when we were nobodies there, purely a deep tech startup.

不像我们当年还是无名之辈，纯粹一家 deep tech 创业公司。

Well, and on top of that we're already growing exponentially.

而且我们已经在指数级增长。

Literally, if you look at the charts there of our usage, of video generations, they're just exponentials there, right now is the perfect time, well, to join.

你去看我们使用量、视频生成量的曲线，就是指数曲线，现在是加入的最佳时机。

Plus, well, and moreover there, probably, there's, well, nothing close in the CIS, no companies like ours that actually build their own models there.

另外，独联体这边大概也找不出第二家像我们这样、真正自己做模型的公司。

That's very, well, I think, cool.

我觉得这挺酷的。

So I invite everyone to, well, I urge you to get in touch with us there, to join.

所以我号召大家来联系我们、加入我们。

And Arman's podcast, honestly, the nFactorial podcast really, well, boosted our company, because all these cool guys there, product engineers, ML engineers, uh, we managed to create a funnel thanks to this very podcast.

还有 Arman 的播客，说实话，nFactorial podcast 给我们公司加了很大一把力，因为这些厉害的产品工程师、ML 工程师，都是靠这个播客做出来的漏斗招来的。

I think the third time it should work too, uh, to gather into the funnel cool people who want to, like, get good at marketing.

我想第三次应该也能成，把想在市场营销上钻研的厉害的人聚进这个漏斗。

And, probably, well, like I was saying, the vibe is such that, well, it's a very stressful, aggressive, but very fun, incredibly interesting environment.

还有，像我刚讲的，那种 vibe 就是：压力大、很有攻击性，但特别好玩，一个极其有意思的环境。

And if you there, well, as a kid I read a whole lot of books on marketing.

如果你——我小时候读过特别多市场营销的书。

One of my

我其中一本

of my favorite books was marke— my favorite, one of my favorite books as a kid was Marketing Warfare.

最喜欢的书里有一本是《营——我最喜欢的，小时候最喜欢的书之一，是《营销战》(Marketing Warfare)。

It's, well, this bestseller, already a classic, about how back in the eighties corporations waged war with each other.

这是本畅销书，现在已经算经典了，讲八十年代那些大公司之间怎么互相开战。

Then all of that, well, got settled, well, obviously, there were probably lawsuits, and now there aren't such intense marketing wars anymore.

后来这些都平息了，中间大概打过官司，现在已经没有那么激烈的营销战了。

And, well, in it they look at marketing as war, literally.

书里是真的把营销当成战争来看的。

Like, they draw direct analogies, that, uh, obviously, whoever is at the top, and if a person is at the top, that means he's already built some kind of defensibility.

他们做的是直接类比：站在山顶的那个人，既然已经在顶上，就说明他筑起了某种防御壁垒。

And the hardest thing, well, the easiest thing is to be on defense, when you've, well, climbed up the mountain.

最难的——不对，最容易的是防守，前提是你已经爬上了那座山。

But attacking the defender, that's the hardest thing.

而去进攻防守方，这才是最难的。

Well, and it's the same in military affairs.

军事上也是一样。

Well, actually, if you're on defense, then you have better positions, and even more so if you're big, then on defense you have better positions than the small ones who are attacking you.

其实只要你在防守，位置就更好；如果你还是个大家伙，那你的防守位置比那些来攻你的小玩家好得多。

Uh, well, and there were a huge number of all kinds of military techniques in there, how they apply to marketing.

书里还讲了大量军事技法，以及怎么套用到营销上。

For example, there are, like, obviously, head-on clashes, there are flank attacks, and we too, for example, practice this often, because we don't have any huge budgets.

比如有正面硬碰硬，也有侧翼进攻，侧翼这套我们自己也经常用，因为我们没什么大预算。

With us, uh, well, if you take out the salaries of our creative team, well, that's basically our marketing budget, then, well, we, uh, the competitors, especially the Chinese competitors, they're very tough, they have a lot of money.

我们这边，把创意团队的工资刨掉，剩下的基本上就是全部营销预算；而对手，尤其是中国那几家，非常凶悍，钱特别多。

Well, so our biggest competitor there is the Kuaishou company.

我们最大的竞争对手就是 Kuaishou 这家公司。

That's around 25 billion dollars in market cap.

市值大概 250 亿美元。

It's the largest company in China, actually.

其实是中国最大的公司。

And it just, when you google some video startup there, like, I don't know, Runway, on Google Kling is always sitting at the top, because Kling just buys traffic.

你在 Google 上搜某个视频创业公司，比如 Runway，最上面永远挂着 Kling，因为 Kling 就是在买流量。

Like, they're just huge.

他们体量就是大。

And attacking them head-on, well, is unrealistic for us, we can't pull it off.

正面硬刚他们对我们不现实，扛不住。

So we, well, from Marketing Warfare too, ever since back then when I read it, and we have a lot of guys who also get this, our creative people, that guerrilla marketing exists, flank attacks exist, that there are various non-standard things you can do that actually work.

所以我们——从读完《营销战》那时候起，我们团队里很多人也懂这套，创意那帮人知道有游击营销、有侧翼进攻，有各种非常规、但真的管用的打法。

And, well, it's incredible, actually, super interesting.

这东西真的不可思议，特别有意思。

Plus we're a data driven company, meaning our processes are set up very seriously.

另外我们是 data driven 的公司，流程搭得非常认真。

With us, well, I mean, what I'm describing, these are processes, it's just, well, it's not that we've got chaos going on for the fun of it, it's just, well, that's what any Silicon Valley company looks like, not just ours, one that's successful, take the historical ones, like, I don't know, PayPal and so on.

我讲的这些都是流程，不是说我们图好玩天天一团乱——任何一家成功的 Silicon Valley 公司都长这样，不只是我们，你看历史上的，比如 PayPal 那些。

That's just what fast-growing companies look like.

高速增长的公司就是这个样子。

On the inside they look, well, very chaotic.

从内部看确实非常混乱。

But that's just a property of the company moving very fast.

但这只是公司跑得太快带来的副产品。

But at the same time the level of professionalism is at the very highest level, so with us everything is super data driven, the processes are super, uh, efficiently built.

同时专业水准是顶格的，所以我们一切都超级 data driven，流程搭得极其高效。

We, well, absolutely hate inefficient processes.

我们极度讨厌低效流程。

If they can be hacked, we always hack them.

能 hack 掉的，我们一定 hack。

If they can be automated, we always automate them.

能自动化的，我们一定自动化。

If we need to, uh, do a reach out there, well, manually it takes a day to write to thousands of creators, influencers there, then obviously we, well, will always sit down, just write the code for an agent with the help of GPT, always scrape there, do it all properly.

如果要做 reach out，手动给上千个创作者、influencer 写信得花一天，那我们肯定坐下来，用 GPT 写个 agent 的代码，把数据爬下来，规规矩矩做完。

And so that, if a normal company does a thousand reaches a day, we can right away do 100,000 a day.

普通公司一天触达一千个，我们直接能做到一天 10 万。

At the same time, to do it right, it's done, well, not entirely directly, because, well, you can end up in spam and so on.

但要做对，就不能完全直来直去，不然会掉进垃圾邮件之类的。

So these are fairly, well, complicated things, yeah, for example, like, I don't know, email marketing or just on so— on social media, uh, cold outreach, you have to do it in stages.

这些其实挺复杂的，比如邮件营销，或者在社交网络上做冷触达，得分阶段来。

We understand how the funnel works, that first you can write to the person, I don't know, like, if we want to get through to a person on LinkedIn, then the first step is to throw him a connection, that's the top of the funnel.

我们清楚漏斗怎么跑：比如要在 LinkedIn 上够到一个人，第一步先给他丢个 connection，这是漏斗的入口。

If it doesn't go through, you have to try to somehow write to him in the DMs.

过不去，就想办法私信他。

If that doesn't go through, then the next thing, you have to try to finish him off with an email.

还不行，下一步就用邮件再补一刀。

If that doesn't go through, you have to, uh, think through the further steps.

再不行，就得把后面几步想清楚。

And this works everywhere.

这套哪儿都管用。

And if you're data driven, you understand, for example, these creators reply to you, meaning they get in touch, then you can find look alike creators, similar to them, with the help of basic data science analysis.

如果你是 data driven 的，比如这批创作者回你了、联系上了，那用基础的 data science 分析就能找到 look alike 的创作者，跟他们相似的那批人。

And we do this and we get high conversions, high funnels.

我们就这么干，转化率很高，漏斗跑得很好。

And we, well, the fact that we grew so fast, of course, that's thanks to the efficiency of these data driven processes.

我们长得这么快，当然就是靠这些 data driven 流程的效率。

And, well, if you want to learn to do this professionally, and actually, well, there aren't that many companies in the world, because, well, we're very deep in the industry and we talk closely with everyone, we understand the average level in the industry and how efficiently, how professionally it's set up with us.

如果你想学会专业地做这件事——其实全世界这样的公司不多，我们扎在这行里很深，跟所有人都聊得很近，我们知道行业平均水平，也知道我们这套有多高效、多专业。

Well, in many companies it's not like that.

很多公司不是这样的。

At the same time we know the playbooks.

同时我们手里有 playbook。

We always, if we know about some company that, for example, right now is in some field similar to ours and they have 100 million ARR, we try to spend as much effort as possible and understand, reverse engineer, what it is they do that works and what doesn't.

只要知道哪家公司现在跟我们领域接近、又做到了 1 亿 ARR，我们就会花最大力气去搞明白、去逆向拆解他们哪些打法有用、哪些没用。

We bring in partners, we try to reach out directly.

我们会拉合作伙伴，也会直接找上门。

And for example, Shatan , our top analyst, who, well, basically is responsible for processes, for data driven processes.

比如 Shatan ，我们的顶级分析师，流程、data driven 流程基本上都归他管。

He's constantly on calls, he's constantly trying, uh, to get on a call with professionals at similar companies.

他天天泡在电话会里，一直想约同类公司的专业人士聊。

We're constantly trying to somehow make friends with them, constantly looking for them, writing to them.

我们一直想跟这些人交上朋友，一直在找、一直在写信。

you can always do some kind of exchange, experience, network exchange, contacts exchange, so that the company gets in touch and shares its experience, and we can give them something in return.

总能做点交换——经验交换、人脉交换、联系人交换，让对方愿意搭上线、把经验分享出来，我们也能回给他们点东西。

And so Shatan , he's constantly on calls.

所以 Shatan 天天在打电话会。

He has at least one call a day with some other company that we know, they're the best at LinkedIn marketing, they're better at Twitter marketing, they're better at Instagram marketing.

他一天至少跟一家我们知道的公司通一次——有的 LinkedIn 营销做得最好，有的 Twitter 营销更强，有的 Instagram 营销更强。

It could be anyone, it could be a crypto company, it could be an insurance company, it could be a financial one, it could be, like, the last time the guys got on a call, uh, with a company that's growing successfully, a BeReal competitor, meaning it's a social network, and how they move on TikTok.

可能是任何人，可能是加密公司，可能是保险公司，可能是金融公司；上一次同事们聊的是一家增长很猛的公司，BeReal 的竞品，社交网络，聊他们在 TikTok 上怎么打。

The guys are constantly researching, constantly want to collect playbooks, best practices.

同事们一直在做研究，一直想把 playbook、最佳实践攒出来。

And uh, in, well, in my view, uh, well, joining us — that's a super, well, way to grow as a specialist.

在我看来，加入我们，是成长为专业人士的一条特别好的路。

At the same time we don't have any, uh, any hard requirements.

同时我们没有什么硬性门槛。

Initially, when Higgsf—, when the company came together, uh, we didn't have this thing that a person has to come with some kind of, with some kind of achievements already.

最开始 Higgsfield 这家公司凑起来的时候，我们就没要求一个人必须已经有什么成绩。

With us, uh, the company is on the contrary kind of anti-credentials.

我们公司反倒是「反履历」的。

We never look at them when hiring.

招人的时候我们从来不看这些。

Meaning on the contrary we always want to understand the essence from first principles.

反过来，我们永远想从第一性原理去看本质。

Like, the person, well, does he really, well, want to do super ambitious things, to develop as a specialist.

这个人是不是真的想做超级有野心的事，想把自己练成专业人士。

And uh we, well, try to do it from first principles.

我们就是从第一性原理出发去做。

That's why our ML guys, well, obviously, none of them had experience training huge models and beating founders from Stanford who have tens of millions of dollars.

所以我们那帮 ML 同事，谁都没有训练超大模型的经验，更别说去打败手里握着几千万美元的 Stanford 创始人。

Well, nobody had that kind of experience.

这种经验谁都没有。

And we, I mean, from first principles just tried to understand whether the person, well, is capable of this or not.

我们就是从第一性原理出发，判断这个人到底做不做得到。

At the same time we took olympiad guys in physics, in math, we took guys like our Sultan, top machine learning engineer.

我们招物理、数学的奥赛选手，也招像 Sultan 这样的人，我们最顶的机器学习工程师。

He just, well, he graduated from Nazarbayev with a 4.0 GPA straight through, all the semesters.

他从 Nazarbayev 毕业，一路 4.0 GPA，每个学期都是。

Meaning he, well, just studied 4 years at 4.0.

四年全是 4.0。

Well, that's probably an indicator too and, well, and the most powerful technical knowledge.

这大概也算个指标吧，技术功底极强。

And that's how it is.

就是这样。

Uh we try from first principles, so it doesn't matter, if you, well, think that you just have a high IQ, and high, yeah, a high IQ matters, what matters, uh, very much, well, is capacity for work, stress resistance and, obviously, well, non-toxicity, because, well, our processes all happen in realtime, fast, so, well, we of course don't like toxic people.

我们从第一性原理出发，所以你要是觉得自己 IQ 高——高 IQ 当然重要，但更重要的是能扛活、抗压，还有不 toxic，因为我们的流程全是 realtime 高速跑的，我们当然不喜欢 toxic 的人。

it's very important for us, if a person is toxic but at the same time he's a top specialist, we still won't take him.

这对我们非常重要：一个人 toxic，哪怕他是顶级专家，我们也不要。

Meaning we, well, don't take toxic people.

就是不招 toxic 的人。

It's very important for us that, well, the culture in the company be more, kind of, uh,

对我们来说很重要的是，公司文化得更偏向那种，呃，

you've had to, you have had to fire someone, right?

那肯定，肯定也开除过人吧？

Yeah, of course.

是，当然。

We, we, yeah, we had, well, I think it's probably, well, normal that, well, it didn't suit everyone, meaning the culture, it's not like this right away, it kind of evolved, took shape.

我们确实有过，我觉得这挺正常的，不可能对所有人都合适；文化也不是一上来就这样，它是慢慢演化、长出来的。

what we've arrived at, it, well, it took some time.

走到今天这个状态，是花了些时间的。

And it didn't suit everyone.

不是所有人都适应。

And, yeah, we had to part ways with some people.

所以确实跟一些人告别了。

So.

就这样。

And, well, because why, well, work if there's no fit, meaning the person, meaning these were top specialists who, well, quickly found themselves some companies that suited them better.

没有 fit 的话干嘛还一起耗着呢——他们都是顶级专家，很快就在更适合自己的公司找到了位置。

They didn't fit because they, they didn't have the ability to work that hard or they were toxic.

他们不合适，要么是没办法那么拼，要么就是 toxic。

What was that down to?

是因为什么？

Well, of course, yeah, there were cases of toxicity too, but usually you can see it in the interview, whether the person is more or less toxic or not.

当然，也有 toxic 的情况，但一般面试就能看出来，这人大致 toxic 还是不 toxic。

Uh, and some kind of vibe check, whether he passes it.

还有 vibe check 过不过得了。

Uh, and also, [music] yeah, I mean, well, anyway, well, it's impossible, yeah, for the company culture to suit everyone.

另外，[音乐] 反正公司文化不可能适合所有人。

So, probably, right now I already understand this, I already said at the beginning, yeah, I already talk people out of it.

所以现在我算是想明白了，前面也说过，我现在都是劝人别来。

Meaning I spend a lot of time specifically explaining, meaning I try to name all the downsides, so that afterwards, when the person has joined, well, no surprises, so, yeah, no surprises, so that the person doesn't get upset, don't want to, well, waste anyone's, well, waste someone's time.

我会花很多时间专门解释，把所有缺点都摆出来，这样人进来之后没有意外——对，没有意外，别让人失望，也不想白白浪费谁的时间。

So.

行。

Great.

太好了。

It's really interesting to talk about your engineering, about how you actually build an image-to-video model.

特别想聊聊你们的工程,聊聊 image-to-video 模型到底是怎么做出来的。

Tell us about the stages.

讲讲分哪些阶段。

Well, again, you don't have to tell your your your secret, not to share your secret sauce, because maybe Pika, Kling or I don't know who else will be watching.

当然,你可以不讲你们的秘密,不用分享你们的秘方,说不定 Pika、Kling,或者不知道还有谁,正在看这期呢。

Uh, Runway there.

呃,还有 Runway。

Tell us what stages, uh, I mean the creation of such a SOTA, state of the art image-to-video model consists of.

讲讲做出这么一个 SOTA、state of the art 的 image-to-video 模型,要分哪几步。

Let's tell it in Kazakh, so they definitely won't find out.

咱们用哈萨克语讲吧,这样他们肯定听不明白。

Yeah.

对。

Well, by the way, yeah, about this espionage thing, actually, it's like that, yeah, it's possible, because we ourselves, for example, uh, well, it was more like product espionage probably, we were sending our employees to China, which is convenient.

对,顺便说说间谍这事儿,其实真有可能,因为我们自己就干过,嗯,我们那个大概更算产品间谍,我们把员工派去中国,这个方便。

Well, China is close by, there are direct flights.

中国近,有直飞。

For an interview, not for an interview, but just in China, well, to talk in general, uh, with creators, with the industry.

去面试?不是去面试,就是去中国,跟创作者、跟这个行业聊聊。

Plus just, basically, most Chinese apps are not available where we are.

还有就是,中国大部分应用在我们这边根本用不了。

Even if you take, uh, TikTok, then even if you download the Chinese TikTok, most of the functions will simply be unavailable by geolocation.

就拿 TikTok 说,你哪怕下了中国版 TikTok,大部分功能也会因为定位直接用不了。

And there, well, all of it is written pretty powerfully so that, well, you can't get around it through some VPNs and so on.

而且那套东西写得挺硬,你没法靠什么 VPN 之类的绕过去。

And there, even more than that, in different regions of China there are different features.

更夸张的是,中国不同地区的功能都不一样。

I mean it's granular even internally.

就是说在内部都做到了颗粒度。

That's why our guys fly over, to understand in general how the industry works, what's going on, because China, obviously, it's overtaking America very strongly, I mean it has already overtaken it.

所以我们的人会飞过去,就为了搞明白这个行业到底怎么转、都在发生什么,因为中国显然把美国甩得很远,或者说已经甩开了。

In fact, especially in consumer, China has overtaken by a lot.

事实上,尤其在 consumer 这块,中国领先很多。

I mean consumer is B2C, China right now completely sets the trends.

consumer 就是 B2C,中国现在完全在定义趋势。

And in fintech already, well, everyone knows it's been setting the trends for a long time.

fintech 就更不用说了,大家都知道早就在定义趋势。

So yeah, it's, well, quite possible that someone will be watching too.

所以对,确实很可能也有人在看。

So.

就这样。

And, well, actually, not much has glob—, if you take it globally, not much has changed since the last podcast, where I explained in real detail what training a big model, a video model, looks like.

其实整体看,跟上一期播客比变化不大,那期我特别详细讲过训练一个大模型、一个视频模型是什么样子。

Uh, probably we've gained even more expertise besides training base models.

嗯,除了训练基础模型,我们大概又多了不少积累。

We've also gained a whole lot of expertise specifically in post-training.

尤其在 post-training 上,积累了非常多。

I mean post-train is exactly about how the model is going to be used product-wise.

post-train 说的就是模型在产品里到底怎么用。

And if before, well, everyone was, yeah, at the foundation model stage, then now it all comes down to post-train, because there are getting to be a lot of models and they have to differentiate from each other.

以前大家都还停在基础模型这个阶段,现在全都归到 post-train 上,因为模型越来越多,彼此之间必须做出差异。

And so post-training, our expertise, the one we've leveled up in, it differentiates us a lot from the competitors.

所以 post-training 这块我们练出来的本事,让我们跟竞争对手拉开了很大差距。

That's why we were able to make a top offering for professional cinematographers, we were able to make a top top offering for VFX artists.

所以我们能给专业摄影师做出顶级方案,能给 VFX 艺术家做出顶级顶级的方案。

And right now we're making a top offering for, well, for avatars.

现在我们在给 avatar 做顶级方案。

It's mostly small and medium business that uses this, thanks to post-training.

用这个的主要是中小企业,靠的就是 post-training。

I mean we can, in a very short time, very optimally, uh, do the, uh, process.

就是说我们能在很短时间里,非常高效地把这个流程跑完。

For us it's exactly this business-slash-technical process, the way our post-training is set up.

我们这套 post-training 的搭法,本身就是个商业斜杠技术的流程。

And uh our machine learning engineers, well that's probably also the cool part, that, well, they're specialists of a very cool level.

还有我们的 ML 工程师,这可能也是牛的地方——都是水平非常高的人。

So, well, for example, a fun story.

举个好玩的例子。

So our ML lead "MLAS", and on top of that he's also a close friend of mine.

我们的 ML leadMLAS,同时还是我的好哥们。

We've been, well, friends since our teenage years, he's a mathematician and now a machine learning engineer.

我们从少年时候就是朋友,他是数学出身,现在是 ML 工程师。

And he, uh, when we realized that there's no data to deliver, to squeeze out the quality we want, we just realize that in open sources there's no data.

他……当我们发现根本没有数据能做出我们想要的质量,才明白公开数据源里就是没有这种数据。

For instance, for example, we want the person to be talking, an avatar, and at the same time there's a dolly zoom.

比如说,我们想要一个人在说话,是个 avatar,同时还有 dolly zoom。

But in films that's, well, considered — you can't do that, you can't shoot it that way, because a dolly zoom is always, well, academically it's supposed to show us some emotion, for example, well, a dolly zoom is such a strongly emotional moment, and nobody shoots dolly zooms with some lightning on top and with the person talking.

但在电影里这是不能干的,不能这么拍,因为 dolly zoom 一定是——按学院派的说法——要给我们展示某种情绪,dolly zoom 是情绪特别强的一个瞬间,没人会一边拍 dolly zoom 一边加闪电、还让人物说话。

I mean, well, it's just not done.

就是说,没人这么干。

We're like: "Well, what are we going to do?"

我们就说:那怎么办?

And he goes: "Well, we've got guys, well, really professional cinematographers".

他说:我们不是有真正专业的摄影师吗。

who are on our creative team.

就在我们创意团队里。

Well, let's go shoot it ourselves.

那我们自己去拍吧。

We go to, well, to Danya, to Danil.

我们去找 Danya,找 Danil。

He, well, he shot, well, a huge music video.

他拍过一个特别大的 MV。

He has all the gear, he has cameras that films are actually shot on.

设备他全有,有那种真拿来拍电影的摄影机。

Mhm.

嗯。

They just, all of this, the decision gets made in one day.

他们就……这一切都是一天之内定下来的。

They say: "We're doing a shoot".

他们说:开拍。

And our most important, well our, actually, we have a secret weapon — that's Seryoga.

我们最关键的那位,其实我们有个秘密武器——Seryoga。

He's the coolest prompt engineer, probably, in the world.

他大概是全世界最强的 prompt 工程师。

Well, I'm not afraid to say it, because we've just seen / been through a huge number of prompt engineers.

这话我敢说,因为我们见过/筛过的 prompt 工程师数量太大了。

And on top of that we know that competitors like Pika are trying to poach him from us.

而且我们知道,Pika 那些竞争对手一直想把他从我们这儿挖走。

Seryoga is straight-up our tank, just a beast, who can make any model work.

Seryoga 就是我们的坦克,纯野兽,什么模型他都能让它干活。

He's an unreally cool techie, unreally cool.

他技术强得离谱,离谱地强。

Well, he's just a total beast.

反正就是个野兽。

Well, in a word.

一句话。

Well, well we call him a research scientist, because he he reads all the papers on machine learning.

我们管他叫 research scientist,因为他把机器学习的论文全看了。

He's on top of all the latest papers on machine learning.

机器学习最新的论文他全跟着。

He knows everything that was at the latest conferences.

最近几场会议上有什么他都知道。

He knows everything about machine learning.

机器学习的事他全知道。

He isn't even close to lagging behind our top ML engineers.

跟我们最顶尖的 ML 工程师比,他一点都不落后。

And at the same time he's a prompt engineer.

而他还是个 prompt 工程师。

But in a past life Seryoga was a production designer.

但上辈子 Seryoga 是美术指导。

He shot music videos for Rammstein, for all of them.

他给 Rammstein,给那些人都拍过 MV。

And well he literally physically built all these sets, the ones he now assembles in AI.

而且这些布景当年是他一砖一瓦亲手搭出来的,现在他在 AI 里把它们又搭起来。

His his two universes collapsed into Higgsfield.

他的两个宇宙在 Higgsfield 撞到一起了。

Yeah, yeah.

对,对。

And we, yeah, two universes in Higgsfield.

对,两个宇宙在 Higgsfield 合到一块儿。

I mean he's both on the creative team and on the prompt engineer team.

他既在创意团队,也在 prompt 工程师团队。

That's where Ruslan is, bro.

Ruslan 也在那儿,兄弟。

And Ruslan was actually on the podcast, top three, I think, of the most-viewed podcasts.

Ruslan 之前正好上过播客,好像是播放量前三的一期。

And so there we are, Olzhik, Danil, Ziya, like, I say, we need to shoot.

然后我们几个,Olzhik、Danil、Ziya,我说,我们得自己去拍。

We're like, we go to Seryoga.

我们就去找 Seryoga。

We go to Seryoga.

去找 Seryoga。

Seryoga, you're a production designer, we need locations.

Seryoga,你不是美术指导吗,我们需要场地。

And, well, for the dataset, obviously, one location with one background won't do.

而且做数据集,显然不能只有一个场地、一个背景。

The dataset is, well, super important.

数据集超级重要。

Anyone who gets machine learning knows that the most important thing in a dataset is that it be diverse, because it'll learn one background and that's it, a person will upload their photo, the overfitted model will change the background to the one it was trained on.

懂机器学习的人都知道,数据集最重要的是多样,不然它把那一个背景学死了,用户传自己的照片,过拟合的模型就会把背景换成它训练时的那个。

We say: "Seryoga, here's the task, we want to organize a shoot tomorrow and at the same time have something like, well, a few dozen different locations".

我们说:Seryoga,任务是这样,我们想明天就组织一场拍摄,而且要有几十个不同的场地。

He says: "Okay, got it, I'm calling my contacts from my past life.

他说:好,明白,我给上辈子的人脉打电话。

We're looking for people, we call, well, industry professionals, they say: "Well, like, are you out of your mind?

我们找人,给行业里的专业人士打电话,他们说:你们疯了吧?

That doesn't happen".

没这种事。

Like any shoot, well, is planned 2 months in advance, and we also need actors who'll do the acting.

任何拍摄都是提前 2 个月排的,而且我们还需要能演的演员。

We call casting, casting directors, they say: "That doesn't happen".

我们给选角、给选角导演打电话,他们说:没这种事。

Like how, what kind of shoot is it that's, well, like, the next day?

什么叫第二天就开拍的拍摄?

And many of them think and, well, people so much don't believe that a shoot gets planned in one day, because that's the industry standard, a shoot is planned 2 months out.

很多人还想……大家根本不信一天就能排出一场拍摄,因为行业标准就是提前 2 个月排。

People don't believe that, well, it's done that way.

他们不信还能这么干。

And some of them think: "Oh right, there was a shoot scheduled.

还有些人想:对哦,是排过一场拍摄。

It, damn, slipped my mind.

我靠,给忘了。

Right, we did plan a shoot".

对,我们确实排过拍摄。

I mean people can't believe that someone really plans a shoot in one day.

就是说大家没法相信真有人一天就把拍摄安排出来。

So we're sitting there at night, Seryoga calls up his friend, the ones they shot music videos for, for Rammstein, for all the cool people.

我们大半夜坐着,Seryoga 给他朋友打电话,就是他们一起给 Rammstein、给那些大牌拍过 MV 的那帮人。

He just has huge sound stages.

他有特别大的摄影棚。

He's like: "Done, I'll get it ready".

他说:行,我来准备。

The guys, Ziya, are searching through the casting director, he just messages everybody as much as possible, they send over, well, I just didn't know how this works.

那边 Ziya 他们通过选角导演找人,他把所有人都问了个遍,人就发过来了,我之前根本不知道这事儿怎么运作。

They send over these actors, they all have portfolios.

他们把演员发过来,每个人都有作品集。

I mean we have to screen them, first of all, well, so that there's divers—, there's visual diversity, that matters too, so that all the people are different.

我们得先筛,首先视觉上要有多样性,这也很重要,所有人都得长得不一样。

Otherwise, if they're all of one look, then you upload your photo, and it'll turn you into, I don't know, an averaged Kazakh, for example.

不然要是长相都一个样,你传自己的照片,它就把你变成——我不知道——比如一个平均脸的哈萨克人。

So it's important, we say, we need enormous diversity, we need them to know English, I mean not just some random extras.

所以很重要,我们说,要极大的多样性,还得会英语,不能只是随便找的群演。

All night Olzhek, well those guys Zi—, they already, everything, the shoot is needed, the cameras have to be brought in, all of that.

整晚 Olzhek,还有 Zi— 他们,拍摄要的东西、摄影机都得运过来,全都得弄。

Olzhik, our machine learning engineer, sits down and just picks the actors.

Olzhik,我们的 ML 工程师,坐下来一个个挑演员。

He sits all night, this one will do, this one — this one just has a strong accent.

整晚坐着挑,这个行,这个——这个口音太重。

And that's it, overnight everything is completely staffed up, the next day and off they go "ne idut".

然后就齐了,一晚上人全配齐,第二天就开工。

And I didn't even suspect how hard shoots turn out to be, well, physically doing them.

我完全没想到,拍摄这事儿原来这么难,体力上这么难。

Turns out it's, well, for me it's, turns out, probably, for whoever is watching and knows, these are, well, obvious things, for doing a shoot.

原来这个——对我来说是这样,看的人里懂行的可能觉得这都是常识——要做一场拍摄。

It lasts 13-14 hours.

一场要持续 13-14 小时。

It's like called a shift.

这好像叫一个班。

For a shift people rent expensive equipment, gear, they set up the lighting.

一个班里,大家要租很贵的器材、设备,还要布光。

The cameras aren't iPhones, they're just huge cameras that get mounted on specialized rails.

摄影机不是 iPhone,是巨大的机器,要架在专门的轨道上。

To do these camera techniques, whole rigs get built.

要做这些运镜,得搭起整套结构。

For example, to do an Arc Left or an Orbit 360, you need to, well, build complicated rails in a circle and shoot.

比如要做 Arc Left 或者 Orbit 360,就得沿圆周搭一套复杂的轨道再拍。

And imagine, you need this whole rig, which, well, in a normal production, it gets assembled over a day and shot in one place, but for us we have to assemble it in one place, drag it to another place, then to a third.

你想象一下,这整套结构,正常制作里是花一天搭好、在一个地方拍完,而我们得在一个地方搭好,再拖到另一个地方,然后第三个。

The actors, it turns out, are also all on set for 13 hours, they have to be fed.

演员原来也要在现场待满 13 小时,还得管饭。

I didn't even suspect how hard this is to pull off.

我完全没想到这事儿有这么难。

And Olzhek, who basically owned all of this globally, who's a machine learning engineer, and the guys are just dropping stories, there's Olzhik in a cap, in headphones, going like, reshoot this, like a director.

而 Olzhek,整件事基本是他从头 own 下来的,一个 ML 工程师,大家还在发 story,Olzhik 戴着帽子、戴着耳机,说这条重拍,跟个导演一样。

So for one day he became a director.

他就当了一天导演。

Yeah, yeah yeah.

对,对对。

Yeah, he became a director.

对,成了导演。

And the actual professionals there, Seryoga, Z—, they they're running around next to him, and he's steering it so that everything, well, so the dataset would be perfect for ML.

而真正的专业人士,Seryoga、Z—,他们在旁边跑来跑去,他在那儿指挥,就为了让数据集对 ML 来说是完美的。

So.

就这样。

And, well, Oljik says afterwards: "Like, only now do I get what we're doing, how, like, well, valuable it is, because he he didn't expect how hard a real shoot is.

后来 Oljik 说:我现在才明白我们做的这事儿有多值,因为他没料到真实拍摄有多累。

It's very hard work.

这是非常累的活。

It's very much, first and foremost, for the technical team, the cinematographers, and for the tech team it's unreally, specifically physically, hard work.

首先是对技术团队,就是摄影师们,对技术团队来说,这是体力上难得离谱的活。

And the production designer — that's also unreally physically hard work.

美术指导也一样,体力上难得离谱。

Take Seryoga, he's mega jacked.

你看 Seryoga,他是个大块头。

Well uh and it turns out, well, you really do have to be a jacked guy to be a production designer.

而且原来,你真得练成大块头才能当美术指导。

If, well, you can't be a production designer and not be physically strong, because, well, real shoots, especially in our real—, our realities, there's no huge budget for you to have a lot of people, and the production designer hauls all of it himself, sets everything up.

你不可能又当美术指导又体力不行,因为真实拍摄,尤其在我们这边的条件下,没有大预算让你雇很多人,美术指导得自己扛所有东西、自己搭。

I mean it's an unreally, well, hard process.

就是说这是个难得离谱的过程。

And probably that's also why our guys from the creative industry got into our mode pretty easily, because, well, shoots are also work that really demands stress tolerance, very chaotic like that.

可能也因为这个,我们那些从创意行业来的人很容易就进了我们的节奏,因为拍摄本身也是特别需要抗压的活,特别混乱。

An actor can come up and say: "That's it, I, the way we agreed, I've shot my part and that's it, I'm going home".

现场演员可能过来说:行了,按说好的我拍完了,我回家了。

And meanwhile the shift is running, a huge number of people, a huge amount of setting all of this up.

可这边一个班还在跑,那么多人,那么多东西要搭。

Well, the shoot isn't fully done yet, and the actor can say: "Like, I'm leaving".

拍摄还没做完,演员就可能说:我走了。

And "Lili", the other way around, or someone else there.

Lili,或者反过来,还有别人。

Very hard work.

非常难的活。

And that's why, well, it was like that, our guys from the creative industry, and they're top people, they, well, meshed extremely easily with our engineers, well, in terms of work ethic.

所以我们这些从创意行业来的人,而且都是顶尖的,他们跟我们的工程师在工作作风上很容易就磨合上了。

So you you have to create your own data.

就是说你们不得不自己造数据。

And do you make synthetic data too, or or how?

那你们也做合成数据吗,还是怎么弄?

How?

怎么弄?

Yes, we, yes, we do make synthetic data, but, well, synthetic data always has, uh, well, some upper bound on quality, which, well, well, if the model that generates it can't do better.

对,我们做合成数据,但合成数据总有一个质量上限,就是说生成它的那个模型自己做不到更好。

And there's always an upper bound.

总有一个 upper bound。

Usually you can solve that with compute, and by the fact that, well, you can just make synthetic data, uh, you can take a mid-tier model, do a huge number of generations, select only the best ones, and you'll get some quality dataset there that you can train, for example, even that very same model on, and it works like a genetic algorithm, and it'll improve, but it's a very hard process, it can, well, also get worse.

通常可以用算力解决,还有就是,合成数据可以这么做:拿一个中等的模型,做非常多次生成,只挑最好的那些,这样能得到一份质量不错的数据集,再拿它去训练,比如说甚至就训练同一个模型,它像遗传算法那样起作用,模型会变好,但这个过程很难,也可能变差。

I mean there are a lot of factors there.

里面因素特别多。

Mhm.

嗯。

So.

就这样。

And each effect is a separate model.

那每个特效都是一个单独的模型。

Uh it's, well, a separate post-train, I mean the model is one, but the post-train is separate.

呃,这是单独的 post-train,就是说模型是一个,但 post-train 是分开的。

So.

就这样。

And, for example, to do a dolly, right?

比如说,要做 dolly,对吧?

And on the input you need 1.000 videos, 10.000 videos.

输入端需要 1.000 个视频,还是 10.000 个视频。

Is there some kind of—?

有没有一个——?

Well, that's a secret, actually, we can't, yeah, say.

这个其实是秘密,不能说。

And all these configurations are, well, a secret, yeah, that's a secret.

所有这些配置都是秘密,对,是秘密。

Well, what's interesting, I mean, sure, so when we

有意思的是,当然了,我们当初

we released this, there was no product in the world like ours.

我们把这个发布了，当时全世界没有像我们这样的产品。

Mhm.

嗯。

And obviously, competitors started copying us.

然后很明显，竞争对手开始抄我们了。

Namely Motion Controls, yeah, exactly these Motion Controls.

具体就是 Motion Controls，对，就是这个 Motion Controls。

There's one Israeli company, it's, well, pretty very big.

有一家以色列公司，体量相当、非常大。

It's in — its main business isn't video generation, it's a Consumer Application.

它的主营业务不是视频生成，是 Consumer Application。

Huge market cap there, it's the biggest company in Israel.

市值巨大，是以色列最大的公司。

And it copied from us, well, just one to one.

它就从我们这儿抄，一模一样地抄。

And, well, what else is funny, at our place, well, our guy there, Karim, he was the one coming up with these VFX effects.

还有更逗的，我们这边的 Karim，这些 VFX 特效就是他想出来的。

Mhm.

嗯。

And he came up with the name too.

名字也是他起的。

He made them up out of his head, some of them.

有些是他凭空想出来的。

And they even named theirs, yeah, like, I don't know, Wind to Face, yeah, the name Symbiote, yeah.

结果他们也这么命名，对，就那个，我也说不好，Wind to Face，对，还有 Symbiote 这个名字，对。

And he was the one inventing them, and they, well, even copied the names, which is just very, well, funny.

名字是他想的，他们连名字都照搬，这真的挺，怎么说，挺逗的。

And and here's what I realized, what's cool, is that I'd read this before in books about startups, but I hadn't understood it firsthand.

还有我想明白了一件事，挺有意思的——以前我在讲创业公司的书里读到过，但没亲身体会过。

What's interesting, well, in all the books about IT it's always written that it's very important to be first.

有意思的是，所有讲 IT 的书里都写，抢到第一非常重要。

The first one there often takes everything.

第一名往往通吃。

And, well, and it wasn't really clear why it works that way.

但一直没太搞明白为什么会这样。

And it really does work, because when they started releasing all of this, plus on top of that Canva —

结果它真的成立，因为他们开始把这些东西 release 出来的时候，再加上 Canva——

Mhm.

嗯。

Oh, they bought the company Leonardo for a big sum there.

哦，他们花了一大笔钱收购了 Leonardo。

And Leonardo AI is also, well, it's old-generation image generation, which also slowly started moving into video generation, but that's already Canva.

Leonardo AI 也是上一代的图像生成，也在慢慢往视频生成转，但那已经算 Canva 了。

Mhm.

嗯。

And so it turns out, Canva also copied from us one to one, stole our name, just copied it one to one.

所以你看，Canva 也把我们一模一样地抄了，把我们的名字偷走了，就是一模一样照搬。

And when they dropped it, obviously, the Israeli one is less popular, it's in narrow circles there, Canva is, well, obviously, a big company.

他们放出来的时候，很明显，那家以色列公司没那么出名，只在小圈子里有人知道，Canva 就不一样了，那是大公司。

People just come into the comments and start writing bad things about them, like, damn, aren't you ashamed, you just copied Higgsfield, and just all the comments.

大家直接跑到评论区骂他们，说，靠，你们不害臊吗，直接就把 Higgsfield 抄了，评论清一色都是这种。

Well, this is just a copy of Higgsfield.

说这就是 Higgsfield 的复制品。

That was very gratifying, that I understand that we're first, we kind of invented this trend, we, uh, people there, our users, because our renewal retention is very high, uh, meaning people don't cancel their subscriptions, on the contrary they use it.

这感觉特别爽——我们是第一个，这个风潮算是我们弄出来的，而且用户那边，我们的续订 retention 特别高，就是说大家不退订，反而一直在用。

And it was very gratifying when people would just come into the comments and go: "Damn, aren't you ashamed".

看到大家跑到评论区说"靠，你们不害臊吗"，真的很爽。

You just copied everything from Higgsfield.

"你们就是把 Higgsfield 全抄了。"

And, well, it's very cool to be first.

而且，做第一个真的很酷。

And right now I don't see any strong traction moment, like they copied us and they managed to somehow further, uh, I don't know, capitalize on it.

而且到现在，我也没看到什么强劲的增长势头——就是他们抄了我们，然后还能靠这个再往下，呃，我也不知道，变现。

They didn't manage to get a reputation there that they're cool at this, because — and then, well, here's what's interesting.

他们没能在这件事上攒出"我们很牛"的名声，因为——接下来才有意思。

And I'll quickly tell this too, that this Israeli company, it trained on our generations, I mean and they didn't even hide it.

再快速讲一个，那家以色列公司是拿我们的生成去训练的，而且他们连藏都不藏。

They open-sourced some part of it, uh, low quality, and they listed the datasets right there, and the datasets there are just screenshots from our site.

他们把其中一部分开源了，呃，质量比较低的那部分，数据集直接就列在那儿，而那些数据集就是从我们网站上截的图。

So the dudes didn't even, well, bother hiding it, because they copied everything from us.

这帮人压根懒得遮掩，因为他们什么都是从我们这儿抄的。

And the quality, of course, isn't as high as ours.

质量当然没我们高。

I mean they copied it there — a user, at first glance, will think it's the same thing, but if he starts generating, he'll see a big difference in quality.

就是说他们抄过去，用户第一眼会以为是一回事，但真开始生成，就能看出质量差距很大。

And in exactly the same way, uh, we

同样地，呃，我们

we went viral on TikTok, became a huge trend there, and we weren’t even ready for virality on TikTok at all.

我们在 TikTok 上爆了，成了那边的超级大趋势，可我们压根没准备好在 TikTok 上爆火。

Well, we found out what it means to be viral on TikTok.

反正，我们算是知道了在 TikTok 上爆红是什么滋味。

Which trend went viral?

哪个趋势爆了？

Uh, several trends.

好几个趋势。

Uh, but the biggest one is Eyes In.

但最大的那个是 Eyes In。

It’s the transition through the eye.

就是穿过眼睛的那个转场。

[music]

[音乐]

Like, that popular transition in music videos.

就是音乐 MV 里特别流行的那个转场。

Also all the cool rappers there use it in their videos.

那边的牛逼说唱歌手也都在自己 MV 里用。

It’s the transition through the eye.

就是穿过眼睛的转场。

A transition like that looks really striking.

这种转场看着特别炸。

And obviously, there Zika, Karim garbled name(s), the guys who really know music videos, they really wanted us to have this effect.

很明显，Zika、Karim 人名听不清 这几个特别懂音乐 MV 的人，非常希望我们能有这个效果。

They believed in it from the very start.

他们从一开始就信它。

Like, they knew that it would — Uh-huh. — go viral.

他们早就知道它会——嗯哼——爆。

And in the end it went viral after, like, I think, a month, maybe even a month and a half after the release.

结果它是在 release 之后大概一个月、甚至可能一个半月才爆的。

So it didn’t happen right away.

所以不是一上来就爆的。

People, yeah, started shooting on TikTok there, throwing in their selfie and doing it.

大家真的开始在 TikTok 上拍，把自己的自拍扔进去，就做出来了。

Putting really fun music on top of it.

再配上特别带感的音乐。

Uh, and, well, they edited it into all sorts of things.

然后他们把这个效果剪出各种花样。

You go into the eye, someone edited it so that it flies into the eye and then the universe.

钻进眼睛里，有人剪成飞进眼睛、然后飞进宇宙。

Someone edited it so that it goes into the eye, then flies out of the eye and something else comes out.

有人剪成进了眼睛，再从眼睛里飞出来，出来的是另一个东西。

There you go.

就这样。

And obviously, well, it became a worldwide trend.

很明显，这成了全球性的趋势。

Our servers there were on fire.

我们的服务器直接烧起来了。

It was a super stressful time.

那段时间压力大到爆。

It was April, yeah, I think.

是四月吧，我记得。

Yeah, yeah, it was the end of April.

对对，那是四月底。

And you went down for 13 hours there.

你们那次挂了 13 个小时。

Of May.

五月。

Yeah, yeah, yeah.

对对对。

Of May.

五月。

It was the end of May.

那是五月底。

And we, yeah, we went down for 13 hours.

对，我们挂了 13 个小时。

We didn’t understand what was going on.

我们完全搞不清出了什么事。

Like, we see that everything is growing, revenue is growing, and the number of subscriptions there is growing, the number of generations is just growing.

我们看到什么都在涨，revenue 在涨，订阅数在涨，生成数就是一直往上涨。

We open Cloudflare, the traffic there is just unreal, and then it goes down.

打开 Cloudflare，流量简直不像真的，然后就挂了。

Uh-huh.

嗯哼。

And, well, our engineers here — that’s Dias and Kuka, our backend guys, and our, we also recently got a new one, Abek “БКР” garbled, he only joined recently and immediately landed in brutal stress.

我们的工程师，就是 Dias 和 Kuka，我们的后端，还有最近刚来的 Abek “БКР”听不清，他刚加入就直接撞上这种高压。

We — They just sit down and, well, we realize, well, globally we didn’t have infra for traffic like that.

他们就坐下来，我们也明白了，整体上我们的 infra 根本扛不住这种量级的流量。

And basically, here with my cofounder, he’s super experienced, he says: “Well, Snap also went down when traffic grew several times over”.

我那位 cofounder 经验超级丰富，他说：“Snap 流量翻几倍的时候也照样挂。”

And basically, well, the guys are stressing, like, they’re responsible for this.

兄弟们压力很大，因为这事归他们负责。

And they, well, took it very personally, that they had written infra that went down.

而且他们特别往心里去，觉得是自己写的 infra 挂了。

I’m like, I call Alex, I say: “So, damn, here’s the situation”.

我给 Alex 打电话，说：“靠，情况是这样。”

And he says: “Well, congratulations”.

他说：“那恭喜啊。”

Well yeah, of course, he he Yeah, everyone in the office is sad, everyone is really just, well, there’s nothing fun going on there, everyone is just sitting there sad, like, what do we do.

是啊，他……对，办公室里所有人都蔫了，真的，一点欢乐劲儿都没有，全都丧着脸坐在那儿，不知道该干嘛。

There Dias says, his friend writes to him: “You guys went viral on TikTok, right?”

Dias 说，他朋友给他发消息：“你们在 TikTok 上爆了吧？”

And he writes back to him: “Like bro, we’re down, like, buried under the volume garbled”.

他回：“哥们儿，我们挂了，被这个量直接压垮了 听不清。”

And, well, everyone’s sad, we call Alex, we say: “Damn, we went down”.

大家都丧着，我们给 Alex 打电话，说：“完了，我们挂了。”

And he, all happy, goes and writes on Twitter: “We’ve got traffic there, the graph, posting the Cloudflare garbled one, we went viral there”, he’s super happy there.

结果他高兴坏了，跑去 Twitter 上发：“我们这边流量、这个图，Cloudflare 那张 听不清，我们爆了”，那边高兴得不行。

And well, and he says, since he’s, well, super seasoned, he’s seen a huge number of projects, and he says: “Not once in my whole career have I seen traffic grow 10x and prod not go down, not once”.

他见得太多了，看过的项目数不清，他说：“我整个职业生涯里，没见过一次流量涨 10x 而 prod 不挂的，一次都没有。”

And at Snap, he says: “It always went down”.

在 Snap，他说：“每次都挂。”

And he said right away: “Why did Snap go down?”

他当场就说：“Snap 当年为什么挂？”

We immediately started checking those things.

我们立刻就去查这些点。

Those were the bottlenecks that we fixed first.

这些正是我们最先修掉的瓶颈。

Uh-huh.

嗯哼。

And why Snap went down, he explained that right away.

Snap 当年为什么挂，他当场就讲明白了。

And I say, let me look right now at how much our traffic grew.

我说，我现在就看看我们流量涨了多少。

I look, we’ve got more than 20x.

一看，超过 20x。

So, and so obviously, the guys there, uh, sat down and just started fixing.

所以很清楚，兄弟们坐下来就开始修。

But, of course, the fixing, yeah, took quite a long time, because we fix one thing, it comes back up, then goes down, we fix another, it comes back up, then goes down.

但修当然花了不少时间，因为修好一个，它活过来，然后又挂；再修一个，活过来，又挂。

And we start calling all the top devops guys, there, at top startups, whoever’s reachable right then, and the time there is already, I don’t know, 12 at night there, night there.

我们开始给所有顶级 devops 打电话，那些头部创业公司的，当时能联系上的都打，那会儿已经，我不知道，半夜 12 点了。

We call, all the devops guys tell us: “You’re being DDoSed”.

我们打过去，所有 devops 都跟我们说：“你们被 DDoS 了。”

Traffic like that doesn’t exist.

这种流量根本不存在。

We send them Cloudflare screenshots there.

我们把 Cloudflare 的截图发过去。

All the devops guys tell us, everyone we called, you’re being DDoSed.

所有 devops 都这么说，我们打过的每一个人都说，你们被 DDoS 了。

This is unreal.

这不可能。

This is unreal traffic.

这流量不可能是真的。

We’re like: “Damn, what if competitors really did start DDoSing us, like, I don’t know, the Chinese?”

我们就想：“靠，万一真是竞争对手在 DDoS 我们呢，比如，我不知道，中国人？”

We’re like: “Cloudflare — a Solution Engineer, urgently; in parallel we go to AWS there — urgently give us the strongest Solution engineer on a call, we’ll figure it out”.

我们就说：“找 Cloudflare 的 Solution Engineer，加急；同时立刻找 AWS，赶紧给我们派最强的 Solution 工程师上会，一起查。”

The guys quickly split up.

兄弟们迅速分头行动。

And it was fun, literally the guys here, on one laptop an AWS Solution Engineer, on another laptop Cloudflare, on a third devops guys from other companies.

还挺有意思，就这么几个人，一台笔记本上是 AWS 的 Solution Engineer，另一台是 Cloudflare，第三台是别家公司的 devops。

Max, well, that kind of very tense atmosphere, really very tense, because, well, we understand that huge traffic is pouring in and the site is down, a very tense atmosphere.

气氛紧张到极点，真的非常紧张，因为我们清楚，巨大的流量正涌进来，网站却躺着，气氛非常紧张。

And Cloudflare, their engineer is like, looking, looking, looking, digging, digging, digging, then says: “Well, it’s unreal for this to be a DDoS, this is real traffic. These are real people, all real people”.

Cloudflare 那个工程师看啊看啊看，挖啊挖啊挖，然后说：“这不可能是 DDoS，这就是真实流量。这是真人，全是真人。”

Or it’s a very unreally, well, expertly planned DDoS, such that even we, Cloudflare, well, the top company in the world, can’t tell that these are real people.

要么就是策划得极其高明的 DDoS，高明到连我们 Cloudflare 这种全球顶级公司都分不出这是不是真人。

He says: “It’s 100% real people”.

他说：“100% 是真人。”

And TikTok meanwhile keeps growing even even more.

与此同时 TikTok 那边还在涨，涨得更凶。

People there are already starting to shoot.

那边已经有人开始拍了。

Well, because it’s like an exponential curve already.

因为这已经是指数曲线了。

We — million-follower bloggers are sending it to us, here they’re not just shooting the effect, million-follower bloggers are shooting tutorials on how to get in garbled and what to push it to.

百万粉博主给我们发过来，他们不只是拍效果，百万粉博主开始拍教程，教怎么进去 听不清、要调到什么程度。

They’re just sending it over.

就这么一条条发过来。

In the end, yeah, 13 hours.

最后，对，13 个小时。

And the AWS engineers helped us, big thanks to them for that.

AWS 的工程师帮了我们，非常感谢他们。

The last fixes already, us together with the AWS engineers, we fix everything.

最后那几个修复是我们跟 AWS 工程师一起做的，全部修完。

The site comes back.

网站回来了。

The very second it comes back, it’s as if it never went down at all, as if traffic never dropped at all.

它回来的那一秒，就好像它从来没挂过，流量根本没掉过。

Like, people were coming and hammering, refreshing, until it would work.

人们一直在那儿死命刷，刷到能用为止。

There was none of that at all, it’s as if for those 13 hours everyone was waiting, yeah, as if everyone was waiting, because, well, it was, personally I was very sad, because, well, personally I, well, had been dreaming, probably, since back in 2013, when I first started thinking about startups at all.

一点都没掉，就好像这 13 个小时所有人都在等，真的像是所有人都在等，因为——我个人当时特别难过，因为我大概从 2013 年就开始做这个梦了，从我开始琢磨创业那会儿。

Well, I was studying all those playbooks on how to go viral.

我研究过所有那些“怎么才能病毒式传播”的攻略。

Basically, well, since 2013 I’d been dreaming about it, yeah, about making a product that goes viral all over the world.

本质上，我从 2013 年就梦想着做出一个火遍全世界的产品。

And here’s that moment, and we just went down.

而这个时刻真的来了——我们却挂了。

And at the same time I myself had read a ton about how to build fault-tolerant distributed systems, but obviously, we didn’t expect it all to happen that fast, you can’t prepare for that, yeah, and it was very sad.

而且我自己读过一大堆怎么搭高可用分布式系统的东西，但很明显，我们没想到一切来得这么快，这根本没法提前准备，是真的很难受。

I thought, we’ll turn it on now and, well, we — I thought that, basically, we’d blown our exponential curve.

我当时想，我们这就把它开起来——我以为我们把那条指数曲线给搞砸了。

It turned out, no.

结果不是。

It turned out that the site comes back on and it’s just wave after wave.

结果网站一开，就是一波接一波。

And we’ve got, well, we’ve got top-tier infrastructure globally, especially GPU, we’ve got Grafanas set up there, we and our Anvar, the site just comes back up, he opens Grafana on a huge monitor, where our GPU is.

我们的基础设施全球来讲是顶级的，尤其 GPU，Grafana 都配好了，我们跟 Anvar——网站刚起来，他就在一台大显示器上打开 Grafana，看我们的 GPU。

And, well, Inference, it’s built on this kind of architecture, on queues, where a huge number of GPUs listen to one queue, into which a stream of generation requests comes, and the GPUs listen to them, and they autoscale, meaning they, well, the fuller the queue is, the more GPUs scale horizontally to unload the queue.

Inference 是搭在这么一套架构上的：基于队列，海量 GPU 监听同一个队列，生成请求以流的形式进来，GPU 监听它们，然后自动扩容——队列越满，GPU 就横向扩得越多，把队列消化掉。

And we just watch the Real Time graph start to grow.

我们就眼看着 Real Time 那条曲线开始往上走。

Like, it had been zero, the site was down after all.

之前是零，网站毕竟是挂的。

And we open the graph of the GPU infrastructure, there’s just zero, the number, literally zero, nothing at all.

我们打开 GPU 基础设施那张图，就是零，那个数字，字面意义上的零，什么都没有。

Like, before that there had always been some good numbers on average, we were happy about it, it never goes down, it always grows, and then a mega exponential curve, and then everything drops to zero.

在那之前数字一直挺漂亮，我们还挺高兴，从来不掉，一直在涨，然后是一条超级指数曲线，然后全部掉到零。

And he puts on some song, some, I don’t know, in the style of Vietnam flashbacks, some American one.

他放了首歌，说不上来，越战闪回那种调调，一首美国歌。

They crank up that song and the traffic just goes up.

歌一开，流量就直接往上冲。

It was, well, a really fun moment like that, when we’re all like that to that song, everyone in the office had been maximally tense for 13 hours.

那真是特别爽的一刻，我们全都跟着那首歌，而办公室里所有人已经紧绷了 13 个小时。

Before that, meaning 13 hours, but before that everyone had been working like a full day.

13 个小时是说这一段，可在这之前大家还上了整整一天班。

And then 13 hours of debugging and everyone is just watching the traffic start to grow.

然后是 13 个小时的 debug，最后所有人就看着流量开始往上涨。

It was, well, such a really fun moment.

那真是特别爽的一个瞬间。

Well, all the same I I’m a bit panicking that we’re losing the exponential curve.

不过我心里还是有点慌，觉得我们正在丢掉那条指数曲线。

I say: “We urgently need to come up with some kind of guerrilla marketing move to get it back”.

我说：“我们得赶紧想个游击营销的招儿，把它拉回来。”

I say: “Guys, come on, basically, let’s parse all the TikToks by by our generation, by this trend.

我说：“兄弟们，来，我们把所有跟我们的生成、跟这个趋势有关的 TikTok 全爬下来。

It’s not just parsing by hashtag there, because, well, not everyone tags the hashtag yet, uh, well.

不能只按 hashtag 爬，因为还不是所有人都打标签。

I ask, quickly write, write a quick parser.

我说，赶紧写，写个快的爬虫。

We parse the data, classi— we sort there from the freshest, latest ones and where there are the most likes.

我们把数据爬下来，分类——按最新的、点赞最多的排。

We split it across the whole company, across all the people.

然后分给全公司，分到每个人头上。

And everyone has to take on, I don’t know, 50-60 TikToks, every employee of the company.

每人认领，我不知道，五六十条 TikTok，公司每个员工都有份。

And we make an HXEL promo code, like, uh, what was it, “HXEL is back” garbled, likely a Higgsfield-branded code, something like that, basically.

然后我们做一个 HXEL 促销码，叫什么来着，“HXEL is back” 听不清，多半是 Higgsfield 相关的码，差不多这么个东西。

And we just, basically, leave it everywhere in the comments.

然后到处留，就留在评论区里。

And everyone just sits down right away and starts spamming that comment on all the TikToks.

所有人立刻坐下来，在所有 TikTok 底下刷这条评论。

This oh, this promo code, this promo code.

就是那个促销码，那个促销码。

Then when we, well, look the next day, it too started growing exponentially, and it started this, well, it definitely pushed the trend further, so that it, well, to make up for those 13 hours.

第二天再看，它也开始指数式往上涨，它确实把这个趋势又往前推了一把，算是把那 13 个小时补回来。

And what did the promo code give you?

那个促销码给什么？

It just gave, I think, three or four free generations, something like that.

就给三四次免费生成吧，差不多。

And the promo code gives free generations.

促销码给的就是免费生成。

Those who saw it — so some portion of them will generate.

看到的人里，总有一部分会去生成。

And there it comes down to numbers garbled, when there are a lot of generations.

生成一多，最后就是个数字问题 听不清。

The more of them there are, the higher the probability that some new viral clips will appear.

生成越多，冒出新爆款视频的概率就越大。

And that’s exactly what happened, once we, well, came back to life, we started processing all that traffic again, and the next day traffic grew hard by several more x.

结果真就这样：我们一活过来，重新开始接住这些流量，第二天流量又猛涨了好几倍。

And, well, the guys already understood what the infrastructure should look like, and we weren’t going down anymore.

兄弟们已经清楚基础设施该长什么样了，我们就再没挂过。

And over the following 3-4 days there were x’s in traffic again.

接下来那三四天，流量又是成倍地涨。

And, well, we already withstood that surge “раф” garbled.

这次我们扛住了这波冲击 “раф”听不清。

And that’s how we, well, understood what it means to go viral on TikTok.

我们就是这么明白了什么叫在 TikTok 上爆红。

Very cool.

太酷了。

Well, and you guys, uh, you mentioned GPUs there — you have a collab with AMD and TensorWave, right?

对，还有你们，呃，你刚提到 GPU，你们跟 AMD 和 TensorWave 有个合作，对吧？

I mean, I saw, uh, a tweet from AMD's official account saying that you guys are their partner.

我是说，我看到 AMD 官方账号发推，说你们是他们的合作伙伴。

Tell me about this collab.

讲讲这次合作。

Yeah, yeah, of course.

对，对，当然。

Well, it's no secret that in AI the hardware, the GPU, is the biggest cost, it's the, well, hardest part, where you need to have expertise, where you need to have, uh, access to GPUs, because, obviously, well, AI is, well, already a huge industry, everybody uses AI, I think, well, it's already become that for most people who, like, use the internet.

在 AI 这行，硬件也就是 GPU 是最大的成本，这不是什么秘密，也是最难的一块——你得有专业积累，得拿得到 GPU，因为很明显，AI 已经是个巨大的产业，所有人都在用 AI，我觉得大多数上网的人都已经是这样了。

Using AI has like already become a daily thing.

用 AI 已经变成每天的事了。

At the same time, well, one company in the world manufactures most of the chips there, TSMC, and that's such a bottleneck, so, well, there's this scarcity on GPUs, and there are only two big companies that, uh, produce general-purpose GPUs, uh, these are the former, well, gaming graphics cards, AMD and Nvidia.

同时，全世界大部分芯片只有一家公司在造，TSMC，这就是个瓶颈，所以 GPU 才这么稀缺；而造通用 GPU 的大公司总共就两家，就是以前做游戏显卡的 AMD 和 Nvidia。

Obviously, Nvidia is the leader and, well, we, I mean, we have a lot of expertise working with Nvidia, obviously, well, we use Nvidia GPUs both for training and for inference, but so that, uh, well, no, anyway the industries develop very fast, and we didn't want to miss AMD's growth either.

很明显 Nvidia 是老大，我们跟 Nvidia 打交道的经验很多，训练和 inference 都用 Nvidia GPU，但是，呃，反正这行业跑得太快，AMD 的增长我们也不想错过。

So I think it was the right move, uh, to figure out the infrastructure, how AMD graphics cards work.

所以我觉得那一步走对了：把基础设施摸清楚，搞明白 AMD 显卡怎么跑。

And what's interesting, sorry, we expected performance, well, more precisely, not perf, we expected that the engineering time to set up training or to set up inference on AMD GPUs would be much bigger.

有意思的是，不好意思，我们本来预期性能……更准确说不是性能，我们本来以为在 AMD GPU 上把训练或者 inference 跑起来，要花的工程时间会多得多。

We'd just, well, have to invest a serious amount of engineering time.

我们得实打实砸进去一大块工程时间。

And our company is small, we, well, can't afford that much time.

可我们公司小，耗不起这么多时间。

And for us that was a pretty risky bet, one that we decided to take.

这对我们是一次相当冒险的押注，我们还是决定赌了。

A lot of the credit there goes to my cofounder Alex.

这里很大一份功劳要记在我联合创始人 Alex 头上。

He did really, well, good bizdev with AMD.

他跟 AMD 那边的 bizdev 做得非常好。

He's, well, really a big professional.

他是真的很专业。

And how, well, American IT corporations work — he himself was, well, like, Director of Generative AI at Snapchat, there, a huge corporation.

美国那些 IT 大公司怎么运转——他自己就当过 Snapchat 的 Generative AI 总监，那可是超大的公司。

He knows how it works from the inside, how to do bizdev properly.

他知道里面是怎么转的，知道 bizdev 该怎么做。

He did an enormous amount of work.

他做了大量的工作。

Uh, he brought it to the point where we got the graphics cards and there was still a big risk that, uh, it might not work, but surprisingly almost everything fired up on the first try there, well, on the second, on the third, but not, like, not on the thirtieth, I mean, and we didn't spend, well, that much engineering time figuring out AMD, so, well, we were, of course, glad to run these things with them there, yeah, and then co-marketing, PR campaigns there, to show it off.

他一路推到我们真拿到了显卡，风险还是很大，可能跑不通，但出乎意料，基本上第一次就跑起来了，好吧，第二次、第三次，反正不是第三十次，也就是说我们并没在啃 AMD 上花掉太多工程时间，所以我们当然乐意跟他们一起做后面这些，做联合营销、公关活动，把成果亮出来。

And indeed there, uh, we tweeted that we're with AMD, AMD also wrote about us officially.

然后我们真的发推说我们跟 AMD 在一起，AMD 也官方写了我们。

And, uh, it went viral.

然后就爆了。

It's like, uh, there are, well, various subreddits on Reddit about the stock market, and all our tweets, our materials.

Reddit 上有各种聊股市的子版块，我们那些推、我们的素材全在上面。

Alex spoke at an AMD conference.

Alex 在 AMD 的大会上做了演讲。

All of it went really viral in that community of people who are bullish on AMD, I mean traders, investors.

这些在那群看多 AMD 的人里传疯了，就是那些交易员、投资人。

It went straight-up super viral.

是真的传爆了。

And, of course, well, we did affect the the stock, yeah, the, well, I, of course, can't straight up say that we affected the stock of a multibillion-dollar company, but the C-level at AMD were definitely very happy.

当然，我们确实影响到了股价，当然我不能直接说我们影响了一家几十亿美元公司的股价，但 AMD 的 C level 肯定非常满意。

I mean, the fact that, well, we're a small startup there, but obviously we're top in the industry, and people, well, tweets there that a top AI video startup uses AMD.

就是说，我们不过是家小创业公司，但很明显我们在这行是顶尖的，大家发推说，一家顶级 AI 视频创业公司在用 AMD。

That went really viral.

这个传得非常广。

So, also in the genre of interesting stories — we already talked about the fact that Elon Musk is one of your users, when did you discover that?

还有个有意思的故事——我们之前聊到过 Elon Musk 是你们的用户，你们是什么时候发现的？

Uh, we, uh, first we had a creator who, uh, made a video with us on Higgsfield.

我们最早是有个创作者，用 Higgsfield 做了个视频。

Ah, yeah, I remember.

啊，对，我想起来了。

It was a comparison with Sora, I mean he compared us with Sora.

那是跟 Sora 的对比，他把我们跟 Sora 比了一下。

Uh-huh.

嗯。

And, well, Sora is obviously already outdated technology, ours is way better.

而 Sora 显然已经是过时的技术了，我们强得多。

And yeah, Elon Musk liked it.

然后 Elon Musk 点了赞。

He's super excited, sends us a screenshot.

他特别兴奋，甩给我们一张截图。

Like, Elon Musk has already seen it, and the post really got millions of views there,

说 Elon Musk 已经看到了，那条帖子真的拿了几百万浏览，

if you want Musk to like your post.

想让 Musk 给你点赞的话，

Criticize OpenAI.

去喷 OpenAI。

Yeah yeah yeah.

对对对。

Yeah, criticize OpenAI.

对，喷 OpenAI。

And then, well, we were blown away there too

然后我们也是懵了

you didn't slaughter a sheep this time.

这次没宰羊吧。

Yeah, we didn't slaughter a sheep, but I remember, I'd just gotten home, they send me a screenshot that Elon Musk liked us, and I go to my wife: "Damn, Elon Musk liked us, I have to go to the office".

对，没宰羊，但我记得我刚到家，他们就甩截图给我说 Elon Musk 给我们点赞了，我跟老婆说：“卧槽，Elon Musk 给我们点赞了，我得回办公室。”

Well, we absolutely have to turn this into a news hook on Twitter.

这个怎么都得在 Twitter 上做成一个话题点。

I'm like: "That's it, I'm heading back, I'm calling everyone, come on, basically, let's get back to the office, let's make a news hook".

我说：“就这样，我回去了，我挨个打电话，大家赶紧回办公室，把这个话题点做出来。”

And this is like around 11:00 at night, right?

这大概是晚上 11 点吧？

Yeah yeah.

对对。

Yeah, it was, well, even later.

对，其实还更晚。

And we make the news hook.

然后我们就做这个话题点。

And then, yeah, and then there was already a series of posts that Musk liked.

接着就是一连串被 Musk 点赞的帖子。

And also, uh, he, uh, left comments, he left a heart there on our generation.

而且他还留了评论，还在我们的生成上留了个红心。

That was very nice.

这个真的挺开心的。

And then, well, I can't even go into further details, yeah, but yeah, I mean we, well, we, basically, got in touch with him, yeah, and everything's very cool, super.

再往后的细节我就不能讲了，不过对，我们跟他确实联系上了，一切都很棒，超好。

And here's just a layman's question, but, but in the TMZ genre, uh, what's the maximum your team, I mean the maximum your team has gone without sleep?

还有个特别外行的问题，不过是 TMZ 那种八卦风格的——你们团队最长多久没睡过觉？

Do you have some kind of internal record like that?

你们内部有这种纪录吗？

Well, for us, or maybe each employee separately probably has some personal record.

我们这边……可能每个员工各自都有自己的个人纪录。

What's yours?

你的是多少？

For me the record was probably more like when I had a mode for very many days where I worked a 20 to 24 hour workday and then 6 to 8 hours of sleep.

我的纪录大概是那段时间，很多天都是一个模式：一天工作 20 到 24 小时，然后睡 6 到 8 小时。

And that mode, it's, well, it's very unusual, you can, well, adjust to it.

这种模式很不寻常，但你是能适应的。

Well, of course, I don't recommend it to anyone.

当然，我不建议任何人这么干。

And in that mode you start losing, well, track of the days, and you don't get it, you're working, you're not, well, your day and night shift around, you don't understand, well, what time of day it is.

在这种模式下你会开始丢掉对日子的感觉，搞不清楚，你在干活，白天黑夜整个错位，分不清是什么时辰。

And I was in that mode, and so were our machine learning engineers, Sultan and Dilkhan.

当时我是这种模式，我们的 machine learning 工程师 Sultan 和 Dilkhan 也是。

I talked about them on the last podcast.

他们我上一期播客讲过。

And we really, well, «Январа лопса» — unintelligible ASR, and we really didn't understand, we're like working, and we step outside to get some air — night, then we come back in, as if we hadn't worked that much, we step out — day, we come back in again, work, step out — like, night again.

我们真的，呃，此处 ASR 乱码，我们真的搞不清了，就一直干活，出门透口气——夜里，回去接着干，感觉也没干多久，再出门——白天，又回去干，再出门——又是夜里。

And we, well, we really started to not understand, well, what was going on.

我们真的开始搞不懂到底怎么回事了。

Plus back then we didn't have as good an office as now.

而且当时办公室没现在这么好。

We didn't have windows there and so on.

那儿连窗户都没有。

We were just sitting in a coworking space, and there were people right next to us who, well, come in to work, work next to us there, leave, and we just didn't understand what was going on.

我们就坐在一个联合办公空间里，旁边有人来上班、在我们边上干活、然后下班走人，我们完全搞不清状况。

So there.

就这样。

Uh, and, well, everyone has a personal record like that, because, well, when there are some news hooks, the creative team just stays, that's it, we're like doing it, when something needs to be released there, the product team there, we're like we set Wednesday, release, we've already, uh, agreed with the creators that they'd all tag us, post about us, that we'd give them access that day.

而且每个人都有这种个人纪录，因为一有什么话题点，创意团队就直接留下不走了，就干呗；要发版的时候，产品团队那边，我们说定了周三发版，我们已经跟创作者们谈好，让他们都 @ 我们、发帖，那天给他们开权限。

And, well, at that point you can't back out.

到这一步就不能往回缩了。

There, maybe, some calls with partners, with investors have already been scheduled there, to show it, so that right after the release you start the call on a positive note.

可能还跟合作伙伴、投资人约好了几个电话，就为了发版之后马上带着好消息开会。

And the product team, well, always gives very optimistic deadlines, then, yeah, they crunch, they just work there a lot, well, like 30 hours.

而产品团队给的排期永远特别乐观，然后就是硬扛，连着干个 30 小时。

The last 2 and a ha— if you take the last 2 and a half months, why am I asking?

最近两个半——如果只看最近两个半月，我为什么问这个？

Well, the last 2 and a half months, if you take them, uh, what's your rough, I mean, mode?

就最近两个半月，你大概是什么作息？

Why am I asking?

我为什么问呢？

Just so I'll know when to text you.

就是想知道什么时候给你发消息合适。

Well, tell me about about that.

讲讲这个吧。

Well, there's no mode as such right now.

现在没什么固定作息。

Well, roughly.

大概说说。

Like, "I go to bed at 12 noon, I wake up in the evening".

比如“中午 12 点睡，晚上起”。

No, there isn't even that.

不，连这个都没有。

There isn't, right.

没有，对。

Well, I mean, you usually work all night.

也就是说你们通常整晚都在干活。

Well, sometimes we don't sleep for a day and a night.

有时候一天一夜都不睡。

Well, there isn't, yeah.

对，没有。

Got it.

明白了。

No day off?

休息日呢？

There's no day off yet either.

休息日目前也没有。

Well, we try, we try to make Sunday a day off, but it just, well, doesn't work out.

我们努力想把周日留成休息日，但就是做不到。

Among other interesting stories,

还有些有意思的故事，

uh, if we take the year twenty-three, that period when you were only just creating the company, one of the people who, uh, let's put it this way, with whom you had a conversation so that you together with — he would be your partner, that's Justin, Founder of Twitch.

呃，说到二三年，你们公司刚刚创立那阵，有一个人——怎么说呢，你们跟他谈过，想让他跟你们一起、做你们的合伙人，就是 Justin，Twitch 的创始人。

Tell us about that story.

讲讲这段故事吧。

Well yeah, that's a very, like, November of twenty-three, right, it's such a, well, for me it's a strange story, because, well, it's a person who is somewhere very far away, an OG, like, of Silicon Valley.

对，那是二三年十一月，这事儿对我来说挺奇怪的，因为这人在很远的地方，是硅谷那种 OG。

And here, probably, well, this is the skill of Alex, my cofounder, the fact that he, well, probably, well, he's like a fish in water in the Valley.

这里头大概是我联合创始人 Alex 的本事，他在硅谷如鱼得水。

I mean he, well, really approaches everything super thoroughly, like he tries to figure out how global business works there in the Valley.

就是说他做什么都特别扎实，会去搞明白硅谷的全球生意到底怎么运转。

And that's how he got to Justin Kan.

他就这么摸到了 Justin Kan。

I mean, Justin Kan found out about us.

也就是说，Justin Kan 知道了我们。

And this was still before.

而且这还是在那之前。

This was way, way before.

这可是老早老早之前了。

This was at the very beginning, right?

这是最开始的时候，对吧？

This, well, this was at the beginning, when we were like, we decided, we're going to do video generation, and we didn't even know that there were already competitors who had raised a lot of money.

这个，是在最开始，我们当时想，我们要做视频生成，我们甚至不知道已经有竞争对手，人家融了一大笔钱。

We were like, basically, we're going to generate video.

我们当时就想，反正我们要做视频生成。

And, well, he found out, and he liked the idea.

然后他知道了，他喜欢这个想法。

Well, obviously, some fired-up mega dudes want video generation there, and a working solution doesn't even exist yet.

很明显嘛，一帮打了鸡血的猛人想做视频生成，而能用的方案还根本不存在。

Well, we, obviously, well, I mean we were looking for angels, cool business angels who could.

我们呢，很明显，我们在找天使，找厉害的天使投资人，能够……

And well, that's how we got to him, but he, yeah, he wanted to become a cofounder.

我们就这么找上了他，但他，对，他想当联合创始人。

He says: "Like, I don't want to be just an investor, let me be a partner".

他说：“我不想只当投资人，让我做合伙人吧”。

Yeah, yeah, let me be a straight-up cofounder.

对对，让我直接当联合创始人。

I've got access to huge capital there and so on.

我有路子拿到巨额资本，等等等等。

Mhm.

嗯。

And we, yeah, yeah, indeed, well, the negotiations went far, we met with him many times, but still for us it was a risk.

我们，对对，谈判确实谈得挺深，跟他见了很多次，但对我们来说这终归是个风险。

Why didn't we go further and agree?

为什么我们最后没往下答应？

Because, well, for us it was still a risk that, well, we hadn't worked with him.

因为对我们来说风险在于，我们没跟他共事过。

Plus the guy has a whole lot of parallel projects going on.

再加上这人手上还有一大堆并行的项目。

And, well, well for us all of this was still kind of like, I mean Alex and I, that's it, we'd agreed, we'd committed that we're in it to the end, all, yeah, we're all in.

而且对我们来说这一切还是有点那个，我跟 Alex 已经说定了，我们承诺要干到底，all，对，我们 all in。

Alex left Snapchat, from a top position, with a huge salary there, uh, and we're like, we're all in, basically, that's it, we want to, well, really build a company in this space, that's it, we're committing.

Alex 从 Snapchat 离职，那是顶层职位，薪水高得吓人，呃，我们就是，我们 all in，就这样，我们想在这个领域实打实建一家公司，就这样，我们押上了。

Mhm.

嗯。

And for us this was kind of, well, we don't have information.

对我们来说这就是，我们没有信息。

He is anyway, well, on a completely different level.

他毕竟处在完全不同的层级。

And we hadn't gotten to that level yet.

那个层级我们还没走到。

Mhm.

嗯。

And for us it was unclear.

我们心里没底。

Mhm.

嗯。

Totally unclear how we would interact and so on.

完全搞不清我们要怎么协作，等等等等。

And so it was a hard decision, actually, because that's also, well, obviously, if a person like that joins, then you definitely close the capital question.

所以这其实是个艰难的决定，因为很明显，这样的人加入，你的资本问题就彻底解决了。

Mhm.

嗯。

Very easily.

轻轻松松。

I mean obviously, he's a close friend of Sam Altman himself, they've been at it since they were 18, the first batch there, yeah, 20 years already.

很明显嘛，他是 Sam Altman 本人的密友，他们从 18 岁起就在一块儿，第一批，对，都 20 年了。

I mean this is, well this is a real OG, these are like people who, well I mean these are the most influential people in the Valley right now.

这就是真正的 OG，就是那种人，现在硅谷最有影响力的那批人。

It's the kind of network where — there's a good article, Power Dynamics in Silicon Valley, where a dude describes how somebody really, well, a startup that people started really heavily strong-arming, uh, and, well, and possibly there were even legal grounds for it, and just one call to Paul Graham, and that's it, the issue was resolved.

这种圈子——有篇好文章叫 Power Dynamics in Silicon Valley，里面一个哥们儿讲到，有家创业公司被人往死里压，而且可能人家法律上还真站得住，然后就一个电话打给 Paul Graham，事情就解决了。

Paul just went to the company that had decided to sue him.

Paul 直接找上了那家决定要告他的公司。

He just said, well, like, well either you, like, stop all of it quickly today, or I simply destroy you as a company, and that's it.

他就说，要么你们今天赶紧全部停手，要么我就把你们这家公司直接干掉，就这样。

Basically, one call from Paul Graham.

总之，Paul Graham 一个电话。

I mean that's how, well, influential these people are.

这些人就是这么有影响力。

And, of course, well, it was scary, unclear.

当然，当时挺怕的，也看不清。

And, yeah, it was a hard decision to say no, but, well, and it's unclear whether it was the right decision or not, only the future will show.

对，拒绝是个艰难的决定，但也说不好这决定对不对，只有未来能给答案。

So there.

就这样。

But it was a very, like, surreal time.

但那真是一段很超现实的日子。

Well, well in general at Higgsfield, all our employees who watched Silicon Valley, they say: I just get flashbacks from that show.

总之在 Higgsfield，我们所有看过《硅谷》那部剧的员工都说：看得我全是那部剧的闪回。

Just every day, every day, like it's, yeah, Silicon, Valley, I mean and we went down, well, not for as long, like for 13 hours, but we also had it, when we went down because of super dumb bugs, of the level where they sat on a keyboard key over there, started fixing everything, rolling out super complex optimizations.

每天，每天都这样，就跟《硅谷》一样，我们也宕过机，虽然没到 13 小时那么久，但我们也有过，因为超蠢的 bug 宕机，就是那种有人一屁股坐到键盘按键上、然后开始各处抢修、上超复杂优化的水平。

We had the exact same thing.

我们碰上的一模一样。

The bug was, well, maximally like a YAML file, one line, some kind of error.

那个 bug 说白了顶多就是 YAML 文件里的一行，某个错。

Like, that kind of thing.

就这种。

Did you have calls with investors who actually weren't planning to invest, but were collecting info for for their portfolio companies?

你们碰到过这种投资人电话吗——对方其实压根没打算投，只是在替自己投的那些公司搜集信息？

That's totally, well, that's the thing, that's a very common thing, so you have to, yeah, research the partner, whether he invested in a competitor, and often you may not know, or he's choosing whom to invest in, or, well, or it's not public yet, that, yeah, it hasn't been announced yet.

这太常见了，这就是那个，这是非常普遍的事，所以要，对，去查这个合伙人有没有投过竞争对手，而且很多时候你根本不知道，或者他正在挑投谁，或者这事儿还没公开，对，还没官宣。

And usually, well, they always announce it only 2-3 months later, well, at minimum, sometimes they announce investments half a year later, and you don't even know.

而且通常都要等 2-3 个月之后才官宣，最少，有时候半年后才官宣投资，你压根不知道。

So usually, well, we always straight up try to ask, especially if we know, we straight up straight up ask what the call is even for, if there's, uh, a conflict of interest, if the company is in an adjacent area and, well, we wait for a normal explanation from the partner, as to why he even wrote to us, what this call is for, and, well, obviously, if he can't give a normal explanation, then that's, well, a totally, yeah, different conversation.

所以我们一般都会直接问，尤其是我们本来就知道的时候，我们就直接、直接问，这个电话到底为了什么，是不是有利益冲突，公司是不是在相邻领域，我们要等这位合伙人给出一个正常的解释，为什么他要来找我们，这电话是为了什么，很明显，如果他给不出正常的解释，那就完全是另一回事了。

Then there's no deep dive there and so on.

那就没有什么 deep dive 之类的了。

Uh.

呃。

Mhm.

嗯。

Let's talk about your pivot, because if I try to very briefly do it myself, right, the history of your releases — so you make the decision to do text, well, to work on video generation at the end of twenty-three, because you had a hypothesis that the year twenty-three is roughly the GPT1 era in the field of video generation, that is, it's 2019.

咱们聊聊你们的 pivot 吧，因为我自己来很简短地捋一遍你们的发布史——你们在二三年底做了决定，去做文本、呃去做视频生成，因为你们有个假设：二三年大致相当于视频生成领域的 GPT1 时代，也就是 2019 年。

Yeah, in in video generation you start working on it.

对，在视频生成上你们开始动手了。

Your first product — it was the product Diffuse, an app that lets you upload your, my photo and turn it all into a dancing video, which I then upload to TikTok.

你们第一个产品是 Diffuse，一个 App，能上传自己的、我的照片，把它变成跳舞视频，我再传到 TikTok 上。

Then your next product appears, that's probably already summer, fall of twenty-four, connected to the generation of series, a Marketplace, where uhh creators create their own series with us, where they themselves are the characters, right, and mini-series of the kind that are very popular in Asia, right, including in China.

然后你们下一个产品出来了，大概已经是二四年夏天、秋天，跟剧集生成有关，一个 Marketplace，创作者在你们这儿做自己的剧，他们自己就是角色，对，就是那种在亚洲很火的迷你剧，对，也包括中国。

And on the other side there are viewers, right, who who watch them.

另一边是观众，对，他们来看这些剧。

And then you get a parallel product, Real Magic, right, which lets you — what did it, what did it do, remind me?

然后你们又有了个并行的产品 Real Magic，对，它能——它当时是干嘛的来着，你提醒一下？

It's specifically for assembling long content, I mean specifically, I mean not short entertainment videos like Reels, but specifically content that's straight up ready for consumption.

它是专门用来拼长内容的，就是说不是 Reels 那种娱乐短视频，而是直接就能拿去消费的成品内容。

There.

就这样。

Uh, so it's straight up a kind of multi-episode series.

呃就是说，这算是好几集的连续剧。

You could do it in Real Magic, well, you can do it.

在 Real Magic 里可以做，嗯，可以做。

There.

就这样。

Mhm.

嗯。

And then on March 31st of twenty-five you do a release, uh, yeah, I mean of your state-of-the-art video model.

然后二五年 3 月 31 日你们做了一次发布，呃，对，就是你们那个 state-of-the-art 的视频模型。

So it turns out this was preceded by a certain pivot, when you made the decision to go not into consumers, right, just, uh, the way Pika and others, Sora, were doing it, but to go into prosumers, right, professional users, to make incredibly high-quality video for editors, clip-makers and so on.

这么看，在这之前有过一次 pivot，你们决定不去做消费者，对，不像 Pika 还有别人、Sora 那样，而是去做 prosumer，对，专业用户，给剪辑师、做短片的那些人做质量高得离谱的视频。

How was that decision even made?

这个决定到底是怎么做出来的？

Yeah, well this

对，嗯这个

a very, like, interesting evolution, probably.

这算是一段挺有意思的演变吧。

I mean, like I said at the very beginning, we were coming together as this kind of deeptech startup.

就像我一开始说的，我们当初是按一家 deeptech 创业公司攒的班子。

I mean at the start, well, we had a strong imbalance.

一开始，我们团队严重失衡。

Most of our engineers - they're machine learning engineers.

我们大部分工程师都是机器学习工程师。

We had this one guy, Almaz, he was the iOS engineer.

只有一个人，Almaz，是 iOS 工程师。

And we had, well, Kuka joined, our backend guy, he's the only one.

后来 Kuka 加入，我们的后端，就他一个。

Dias, our current backend guy as well, he also, well, joined as a founding engineer at the start, but we hired him specifically for data engineering, specifically for ML tasks and then, well, he moved over to product-end tasks.

Dias，现在也是我们的后端，他也是最早以 founding 工程师身份加入的，但我们招他是专门做数据工程、专门做 ML 那块，后来他转去做产品端的活儿了。

He's originally a C++ guy from Yandex and mm and the product wasn't like for the first several months the product wasn't well exactly a product in the form of, like, a mobile app or a website that wasn't like our main focus.

他原本是 Yandex 的 C++ 工程师，而且头几个月，产品——真正意义上那种手机 App 或者网站——并不是我们的主要方向。

I mean we always pitched the models, we pitched our infrastructure there, our machine learning, expertise, uhh and specifically the mobile app and then the website that we were making, that was more like it was just a playbook.

我们一直是拿模型去 pitch，拿基础设施、机器学习、技术积累去 pitch，至于手机 App 和后来做的网站，那更像只是个 playbook。

It needs, it needs to just exist and to be shown.

就是得有这么个东西，能拿出来给人看。

So you were planning to sell your technology through an API?

所以你们当时打算通过 API 卖自己的技术？

We, well, we looked at a lot of different options.

我们看过很多不同的路子。

I mean we, obviously, early stage, we looked at different options, and we, well, we didn't know for sure how the business model would look in the end.

我们当时明摆着还在 early stage，各种方案都看过，也确实不知道最后商业模式会长成什么样。

I mean we're like deeptech.

我们就是搞 deeptech 的嘛。

Uh-huh.

嗯。

Uh, Diffuse, what's good, we were the first who released dancing people, it went viral.

Diffuse 这边，好处是，我们是第一个 release 出「让真人跳舞」的，火了。

Of course, not on the scale that we ended up, well, going viral now, uh, relatively we went viral.

当然规模远不如我们后来、现在这样爆，那时候只是相对地火了一下。

And it turned into this kind of small trend.

后来变成了一个小小的潮流。

Well, it's just that back then the quality still wasn't that high.

那会儿质量还没那么高。

And when we pitched, I mean it was this clear story, I mean for the investors, that there's deeptech expertise, there is, uh, well, here are guys with deeptech expertise, plus there's my cofounder Alex, who has expertise in, well, in Snapchat, a social network, so a product that's, uh, well, more like B2C, was, well, understandable to investors.

我们 pitch 的时候，故事讲得很顺：有 deeptech 的积累，有一帮带着 deeptech 经验的人，再加上我的联合创始人 Alex，他有 Snapchat、社交网络那块的经验，所以做一个更偏 B2C 的产品，投资人是听得懂的。

Well, obviously, they invested in us at, like, when we were still at the pre-seed stage, that, well, something interesting will probably come out of it, and Diffuse started picking up good metrics, showing them, and that helped us close the round.

他们投我们的时候我们还在 pre-seed 阶段，赌的就是这帮人大概能做出点有意思的东西；后来 Diffuse 的数据开始好看，拿得出手了，这帮我们把那一轮谈成了。

I mean Diffuse, well, really did help us close the round.

Diffuse 确实帮我们把那一轮融资 close 掉了。

And what's good, we really very much also already back then, well, it was clear that the team works fast, because we released Diffuse, we were the first who did the dances.

而且好在，那时候就已经很明显——这个团队做事快，因为 Diffuse 是我们 release 的，跳舞这功能我们是第一个做出来的。

And then Viggle too, well, it was viral last year.

然后 Viggle 也是，它去年火过。

It's not based on neural nets, it's more 3D technology there.

它不是基于神经网络的，更多是 3D 技术。

And Viggle a month later published its own solution for dances and went viral.

Viggle 晚一个月发了自己的跳舞方案，然后爆了。

And since it was based on 3D there, and there at that moment the quality was better than on neural nets.

因为它走的是 3D，当时那个质量比神经网络路线更好。

And Viggle, it went viral harder than we did.

Viggle 比我们火得还猛。

But since we just did it faster, earlier, we closed the round successfully.

但我们就是做得更快、更早，所以那一轮顺利 close 掉了。

And just if we had, for example, dragged it out by only a month later, and that was possible, because, well, of course, you want a good product, one that's higher quality, that breaks less there and so on.

要是我们哪怕只往后拖一个月——这完全有可能，因为谁都想做个好产品，质量更高、更不容易崩之类的。

And that was the first lesson then, that you have to release as soon as possible, faster.

这就是第一课：能多早 release 就多早，越快越好。

Very important, well, time to market, basically, is very important.

很重要，time to market 真的非常重要。

And ours is, well, super fast.

我们这块超快。

In the end it's in our company's DNA that time to market there is close to zero, we release as fast as we possibly can.

到最后这成了公司的 DNA：time to market 接近于零，能多快 release 就多快。

And yes, that was a turning point, because if Viggle had released and we came after Viggle, then it would have been way harder for us.

这确实是个转折点，如果 Viggle 先 release、我们跟在它后面，那我们会难得多。

I mean that's the kind of lesson, that you have to release fast.

这就是那个教训：必须快速 release。

So there.

就这样。

And then we, well, we tried to understand how, well, what the future would look like.

然后我们试着想明白，未来会长成什么样。

Uh, obviously, yeah, there were a lot of these prosumer apps.

那时候明显有一大堆 prosumer 应用。

And back then at that moment it was still very hard to go at the prosumer, because these companies like Polaps were still around, the ones that have already shut down, and startups, like Hyper, and others, Moonvalley, and they raised huge rounds from cool funds there, like, uh, well, from the top ones there like Khosla Ventures and so on, huge investments there, tens of millions of dollars.

那个时间点想切进 prosumer 还是很难，因为像 Polaps 这种公司当时还在——现在已经关门了——还有 Hyper、Moonvalley 这些创业公司，他们从很厉害的基金那儿融了巨大的轮次，都是 Khosla Ventures 这种顶级基金，投的钱是数千万美元级别。

And they just started subsidizing prosumers.

然后他们就开始直接补贴 prosumer 用户。

I mean they just gave prosumers free generations in Discord.

就是在 Discord 里白送 prosumer 免费生成。

And back then it was fashionable to do all of this in Discord, because back then, well, the success case that a company on Discord, has a huge ARR there.

那会儿流行把这些都放在 Discord 上做，因为当时的 success case 就是：一家开在 Discord 上的公司，ARR 高得吓人。

And they just subsidized GPUs, burned investors' money, and for us, well, to compete on the prosumer market against them, that would have been, well, maximally like suicide, because, well, we couldn't subsidize.

他们就是在补贴 GPU，烧投资人的钱，而我们要在 prosumer 市场上跟他们硬碰硬，那基本等于自杀，因为我们补贴不起。

We were barely, barely there, well, somehow making our finances, the economics, add up, because GPUs are very expensive, and we don't have such big investment amounts.

我们自己都是紧巴巴地、勉强才把财务、把账算平，因为 GPU 太贵，我们融到的钱又没那么多。

So there.

就这样。

That's why we were looking at different options, we tried to look far into the future.

所以我们才去看各种不同的路子，试着往很远的未来看。

And, well, obviously, the future - it's about long content generation.

很明显，未来属于长内容生成。

And we, well, as R&D, we decided to just figure it out.

我们就当 R&D，决定把这事儿彻底搞明白。

Like, basically, let's understand how content is going to be generated in the future.

说白了就是：先搞清楚未来内容会怎么生成。

That was the beginning of twenty twenty-four, now it's twenty twenty-five.

那是 2024 年年初，现在是 2025 年。

Everything that we back then, well, predicted, worked out, how we understood how the industry would work, we already see that 100% of it is coming true.

我们当时预测的、推演出来的那一整套——我们对行业会怎么运转的理解——现在看，100% 都在应验。

We understood that on social media, well, it would all start with this kind of content, the thing that's now called brainrot.

我们判断，社交媒体上这一切会从那种内容开始，就是现在叫 brainrot 的东西。

By the way, Seryoga our prompt engineer, he's essentially one of the authors of brainrot, because he was making brainrot when nobody got it.

顺便说，我们的 prompt 工程师 Seryoga，某种意义上就是 brainrot 的作者之一，因为他做 brainrot 的时候还没人看得懂。

He was making some kind of very strange generations there that went viral on TikTok.

他当时做的那些生成特别怪，在 TikTok 上全火了。

Like jacked Squidward, dancing to weird music, dancing cats went viral.

比如肌肉版 Squidward，跟着怪音乐跳舞，跳舞的猫，全火了。

He's like the author of brainrot, the father of brainrot in general.

他算是 brainrot 的作者，简直是 brainrot 之父。

And we understood what the industry would look like, it all matched up 100%.

我们对行业会长成什么样的判断，100% 全对上了。

the way we thought from first principles.

跟我们从第一性原理推出来的一模一样。

And we started building this technology of long video generation.

于是我们开始搭这套长视频生成的技术。

And we built it and then we started thinking, what's next.

搭出来之后，我们开始想下一步干什么。

And the hardest thing in long generation - that's consistency between frames, right?

长生成里最难的就是帧与帧之间的一致性，对吧？

Yes.

对。

I mean you remember, right, this thing that we discussed back in February of twenty twenty-three with Igor Tsai, right?

你还记得吧，2023 年 2 月我们跟 Igor Tsai 聊过的那个事儿？

When the frames, right, at that moment, that was two and a half years ago, there's no, there's no continuity between the frames.

那会儿的帧，也就是两年半以前，帧和帧之间根本没有连贯性。

The eyes are green first, then purple.

眼睛一开始是绿的，后面变紫的。

Yeah, how, how did you solve this problem?

那你们是怎么、怎么解决这个问题的？

In the end, yeah, well, we solved it by the fact that we started having expertise in post-training, that we learned to train models that can generate frames exactly consistently.

最后我们是靠积累起 post-training 的能力解决的，我们学会了训练那种能真正一致地生成帧的模型。

And we really have huge expertise in this at this point, which we even still, well, I think, the time just hasn't come yet.

这块我们现在确实积累极深，深到我们甚至还没有……我觉得吧，只是时机还没到。

And since the question, purely an economic question, and generations right now are pretty expensive, and for even a regular prosumer it's, well, very expensive to generate like 5 minutes, 10 minutes of content content costs quite a lot.

因为这纯粹是个经济问题——现在生成还挺贵的，哪怕普通 prosumer，生成 5 分钟、10 分钟的内容，成本也相当高。

Back then we still, well, didn't understand that, uh, prices, well, we thought that LLM prices are falling and on video generation prices should also fall just as fast.

那时候我们还没想明白价格这事儿——我们以为 LLM 的价格在降，视频生成的价格也该一样快地降下来。

But obviously, there are a lot of things there that don't depend on us, like geopolitics there and so on, corporate wars there .

但明摆着，有很多事不由我们决定，比如地缘政治之类的，还有企业间的战争 。

And to this day generating long consistent video, it still costs a lot, and the economics don't add up yet.

到今天，生成长的、一致的视频还是很贵，账目前算不平。

And we, well, we're just waiting for the moment, since our expertise, well, well, I, of course, don't know for sure how it is with competitors, but I do, well, know for sure that we have the most powerful expertise in this.

我们就是在等那个时机，因为我们的积累——当然，竞争对手是什么水平我不敢说，但我确定我们在这块的能力是最强的。

In my view, it seems to me, we have one of the world's best expertise in generation

在我看来，我们在生成这块的能力是世界上最好的之一

long videos.

长视频。

Uh, and we realized that selling users the generation of long videos, well, well, at the beginning we still didn't get this, we, well, now we've got it, but we roughly understood, possibly, it's too expensive, and for the economics to work out, well, these have to be not just fun videos for social media, these already have to be videos that people are ready to pay for directly, the way it looks in real productions.

我们意识到，把长视频生成卖给用户——一开始我们还没想明白，现在才想明白，但当时大概也感觉到，可能太贵了，要让账算得过来，那就不能只是社交媒体上好玩的视频，得是人们愿意直接掏钱的视频，像真实影视制作里那样。

The thing that shooting a series there — that costs a lot of money, shooting a film — that's even more expensive.

拍一部剧集，成本非常高；拍一部电影，更贵。

animated ones, there the budgets are huge too.

动画的，预算同样巨大。

And we thought: "What if we try to go into the market, make an AI Netflix, that is, literally make an application, a streaming platform, where we ourselves will generate the content and people will watch it, pay for it, consume this content."

我们就想："要不试试进这个市场，做一个 AI Netflix，就是真的做一个应用、一个流媒体平台，内容我们自己生成，人们来看、来付费、来消费这些内容。"

And we, uh, really did start doing this.

然后我们真的开始做了。

Uh, we realized that, uh, you can make an AI Netflix where people will watch AI generated content.

我们意识到，可以做一个 AI Netflix，人们在上面看 AI 生成的内容。

That's a kind of, well, fairly, well, ambitious bet, to build an AI Netflix, because, well, Netflix is a huge company.

这算是个挺有野心的一注，要造 AI Netflix——毕竟 Netflix 是家巨无霸公司。

It's unclear whether the market is even ripe for this or not, whether people are ready to watch AI content, because often it's clear, people over there on social media watch AI content, it racks up hundreds of millions of views.

市场到底成熟没成熟，不清楚；人们愿不愿意看 AI 内容，也不清楚——因为很多时候这是明摆着的，大家在社交媒体上就是在看 AI 内容，动辄几亿播放。

Uh, and back then it was totally close, unclear.

而那时候，这事完全 接近、说不清楚。

Nobody even knew how to make 15 seconds of more, more or less quality, watchable stuff.

那时候没人能做出哪怕 15 秒还算过得去、能看下去的东西。

And we're all super-ambitious, obviously, like, we'll make it, uh, AI Netflix, that's it, we sit down.

我们一帮人野心爆棚，当然了，心想：我们就把它做出来，AI Netflix，就这样，坐下开干。

Like HD, how do you do this?

然后就是：HD，这东西到底怎么做？

Well, we're engineers, let's reverse engineer it.

我们是工程师嘛，那就反向工程呗。

At that point we still didn't have a creative team.

那时候我们还没有创意团队。

We had prompt engineers and our ML engineers.

我们有的是 prompt 工程师，还有我们的 ML 工程师。

We're like: "Let's reverse engineer films and series".

我们就说："来，把电影和剧集反向工程一遍。"

We download content and start, uh, doing data science, that is, analyzing with the help of multimodal models.

我们把内容下下来，开始做 data science，就是用多模态模型去分析。

And for training video models it's very important to have this expertise in multimodal LLMs.

而训练视频模型，非常关键的一点就是在多模态 LLM 上有这份积累。

That is, we didn't just use them, multimodal LLMs, we straight up trained them.

也就是说，我们不只是拿多模态 LLM 来用，我们是直接训练它们。

We, we have our own multimodal LLM models that are really great at labeling, uh, frames from films and scenes from films.

我们有自己的多模态 LLM 模型，特别擅长给电影的画面和场景打标注。

Very high-quality, precise, so as to understand it straight up directly, because, well, a frame or a shot, uh, well, a film, it essentially consists of a shot list.

质量很高、很精确，能直接把它读懂——因为一个画面，或者说一个 shot，电影本质上就是由一份 shot list 组成的。

Every shot has a camera angle, and there it's what composition, what's happening in the frame, the emotion, uh, the actor's line.

每个 shot 都有机位角度，还有构图是什么样、画面里在发生什么、情绪、演员的台词。

Well, that's already like metadata, uh, a shot list, so that's the film's metadata, but a shot list, it doesn't exist, doesn't exist in digital form, well, at least not in public access.

这就已经算元数据了，shot list 就是电影的元数据，但 shot list 并不存在、并不存在电子版，至少公开渠道拿不到。

And shot lists, directors usually make them somewhere on paper for themselves during, like, during production.

shot list 通常是导演拍摄的时候自己在纸上写的。

That is, they plan out what the film is going to look like.

也就是他们规划这部电影会长成什么样。

They make a storyboard.

他们会做分镜。

Usually it's drawn straight by hand.

通常是直接手绘出来的。

Then they make a shot list of how the cameras will need to be set up on the shoot and so on.

然后做 shot list，写清楚现场机位要怎么摆等等。

And we, well, we understand that, okay, a film has metadata, that's the shot list.

我们就明白了：OK，电影有元数据，那就是 shot list。

We roughly understand what a shot list looks like.

我们大概知道 shot list 长什么样。

We need to train a multimodal LM that can, you give it some content, it makes you a shot list.

我们得训一个多模态 LM，你喂它一段内容，它给你出一份 shot list。

And the idea is that if we get shot lists, we'll then be able to generate shot lists and then train it the other way around, from shot to video.

想法是：只要我们拿到 shot list，之后就能生成 shot list，再反过来训练从 shot 到视频。

And then from each shot we'll be able to generate video.

然后每个 shot 我们都能生成一段视频。

glue that together, and you end up with a long video.

拼接起来，你就得到一段长视频。

And there's a lot there, I mean in directing there are a lot of rules.

而这里面规矩特别多，我是说导演这门手艺里规矩特别多。

There's the concept of blocking, that's how a scene should be properly built compositionally, there's the concept of continuity, that's how the frames, so that there's a feeling that the characters are in one space, that the actions are connected to each other.

有个概念叫 blocking，就是一场戏在构图上该怎么正确搭建；还有个概念叫 continuity，就是画面之间要让人感觉角色处在同一个空间里、动作彼此接得上。

So, for example, well, the coolest thing is, well, those Hong Kong martial arts films, Jackie Chan, where continuity is solved by it just being one long shot and the fight just goes in one shot, and it looks really cool.

比如说，最牛的就是香港武打片，Jackie Chan 那种，continuity 的解法就是一整个长镜头，打戏就一条 shot 拍完，看着特别爽。

Uh, but to shoot something like that, you just need professionals.

但要拍成那样，非得是真正的专业人士不可。

And, uh, the popcorn kind of industry, more capitalist, in Hollywood it, well, shoots differently.

而好莱坞那种爆米花的、更资本化的工业体系，拍法不一样。

There the process has to be more, uh, not tied to some cool kung fu artist.

那边的流程更要求不绑死在某个厉害的功夫演员身上。

That's why those Hollywood fights, action, they're very, well, really, well, hard to watch.

所以好莱坞那些打戏、动作戏，说真的，很难看下去。

Every second a new cut.

每一秒一个新的 cut。

Every second a new cut.

每一秒一个新的 cut。

Every second something new.

每一秒都有新东西。

If you watch closely, you can even notice a lot of continuity errors, stuff that doesn't connect.

仔细看的话，甚至能发现一堆穿帮，前后接不上。

Continuity — it's very important that this feeling is there.

Continuity——这种感觉必须在，这非常重要。

For example, do you like the channel Every Frame a Painting?

比如说，你喜欢 Every Frame a Painting 这个频道吗？

There used to be a channel like that.

以前有这么个频道。

You haven't watched it?

没看过？

Uh, I don't remember them by name, but when I was diving into this, I watched a lot of these YouTube channels, figured out how it works, because, well, in fact nobody knows at all what directing looks like.

名字我记不住，但我当时钻进去的时候看了好多这类 YouTube 频道，弄明白它是怎么运作的——因为说实话，根本没人知道导演这门活儿到底长什么样。

Well, like, what blocking, continuity is, we just don't know.

就是说，blocking、continuity 是什么，我们压根不知道。

Mm-hm.

嗯。

shot list and so on.

shot list 等等。

And, well, the simplest thing is just a dialogue.

而最基础的，就是一段对话。

That means you have to set up two cameras.

那就得架两台机器。

There one camera, for example, on me, the other on Arman.

比如一台对着我，另一台对着 Arman。

There.

就这样。

Well, like right now here in the studio.

就像现在在录音棚里这样。

And two cameras have to be set, set up.

两台机器必须摆好、架好。

And there has to be that «figure eight», the shot-reverse-shot.

而且要形成那种"正反打"。

There's also this technique, over the shoulder, where the other person's shoulder needs to be visible, so the viewer understands that they're facing each other like that.

还有个手法叫 over the shoulder，就是要让对面那个人的肩膀入画，好让观众明白他们是面对面的。

And composition matters there too.

构图在这儿也很重要。

If, for example, one is more dominant over the other, then he has to be shot from above, the second one from below.

比如其中一个人比另一个更强势，那他就得从上方拍，另一个从下方拍。

And there's a huge number of such unspoken rules.

这种不成文的规矩多得吓人。

Have you, have you seen that famous video where, well, one president of one of the big countries, there are two microphones, and one stands a bit higher, the other a bit lower, and he plays with it to set his own higher.

你看过那个有名的视频吗——某个大国的总统，那儿摆着两个话筒，一个稍高一个稍低，他就在那儿摆弄，非要把自己那个调得更高。

You don't remember that video?

不记得那个视频？

That reminded me about this, like, how to put it, power dynamic, right?

这让我想起那个，怎么说，权力关系的博弈，对吧？

Cool.

有意思。

So yeah, well, we didn't know about this.

所以说，我们当时不懂这些。

I mean, like, if, well, like, well, Spielberg, right, I knew that he's some cool director or whatever, and the other way around, they're so overhyped that people actually hate on them.

就是说，比如 Spielberg，我知道他是个很牛的导演；反过来，有些人被吹得太狠，反而挨黑。

Well, for example, Tarantino, right, many don't like him, well, on TikTok it's specifically the zoomers who don't adore Tarantino.

比如 Tarantino，很多人不喜欢他，TikTok 上尤其是 Z 世代 并不 迷 Tarantino。

Why don't zoomers like Tarantino?

Z 世代为什么不喜欢 Tarantino？

For the violence or for what?

嫌暴力，还是别的什么？

Or, I don't know, for some reason, the way I understand it, he's too cringe for them, basically.

我也不知道，反正照我理解，他对他们来说太尬了。

Well, I didn't really go into the details, but for them for some reason he's considered cringe, it seems.

我没细究，但在他们眼里他不知怎么就被当成很尬，好像是这样。

because, well, there's this meme that a guy mansplains Tarantino films to his girlfriend.

因为有那么个梗：男生给女朋友大讲特讲 Tarantino 的电影。

Of course, there is such a meme.

当然，是有这么个梗。

There.

就这样。

And, well, I didn't understand why they're cool or not cool.

我当时也不明白他们凭什么牛或者不牛。

And then I, well, I understood, uh, and a professional film critic consulted for us, uh, Miras, he's actually a professor at KIMEP, he teaches film studies there, as I understand it.

后来我懂了——有位专业影评人给我们做顾问，Miras，他其实是 KIMEP 的教授，据我了解在那儿教电影学。

So he consulted for us for several months, explained things to us, he himself studied to be a film critic in England, that is, he understands it very well.

他给我们做了好几个月顾问，一点点讲；他自己是在英国学的影评，所以吃得非常透。

And so he showed us that Spielberg, here he shoots one long shot, and you literally, you kind of see that one character walks in, the camera is on him, he goes up, the camera moves, he's always the center of attention, then some object appears and it's right in the center, you literally make a point and everything is always at that point.

他给我们看：Spielberg 拍一个长镜头，你真的看得出来——一个角色走进来，镜头对着他，他往上走，镜头跟着，他始终是注意力中心；然后出现某个物件，正好在正中间，你就是在定一个点，一切永远落在这个点上。

And then that object gets taken by someone else or someone walks in with a gun.

然后这个物件被另一个人拿走，或者有人拿着枪走进来。

And this line is always held.

这条线始终被拽住。

some kind of arcs and mise-en-scène.

一些弧线什么的，还有场面调度。

Yeah yeah.

对对。

Yeah.

对。

There.

就这样。

Yeah, by the way, yeah, we don't know what mise-en-scène is.

对，顺便说，我们根本不知道什么叫场面调度。

He explained it to us.

是他讲给我们听的。

Well, that's how little we understood.

我们就无知到这个程度。

And, well, it's, well, it's shot really cool technically, that these are complex shots, that it's dynamic, he like runs up the stairs.

而且这在技术上拍得非常牛——都是复杂的 shot，很有动感，他就那么冲上楼梯。

So if you go rewatch it, though I forgot the name of the film, and so he runs up like that, and the camera moves.

你们要是回头重看——虽然片名我忘了——就是他那样往上冲，镜头跟着动。

That's, well, really coolly shot.

那真的拍得牛。

And so Miras explained all this to us, we didn't understand this.

Miras 把这些都给我们讲了，我们原来根本不懂。

We, well, we did a really huge deep dive.

我们做了一次特别大的 deep dive。

And then we did a very deep deep dive into soap operas, because right now that's the fastest growing market of content consumption.

后来我们又对肥皂剧做了一次非常深的 deep dive，因为它现在是内容消费里增长最快的市场。

And apps with mobile soap operas, uh, these mobile apps where they get streamed, in terms of revenue, in terms of mobile revenue, they've already overtaken Netflix.

而那些手机肥皂剧的应用——就是播放它们的手机应用——按 revenue、按移动端 revenue 算，已经超过 Netflix 了。

The Chinese came up with this.

这是中国人想出来的。

In China 60% of watched content is these soap operas, mobile, vertical.

在中国，人们看的内容里 60% 就是这种肥皂剧，手机上的、竖屏的。

And this gradually started moving over to America.

这股风逐渐开始传到美国。

In America it's on the order of less than 2% market penetration.

在美国，市场渗透率大概还不到 2%。

That is, it's still a huge market.

也就是说市场还大得很。

And obviously, we were very interested in it.

显然，我们对它非常感兴趣。

Plus it's not Spielberg, generating Spielberg level, it's like generating the level of soap operas, which get shot for 500 episodes anyway.

而且这不是要生成 Spielberg 级别的东西，是生成肥皂剧那个级别的——那种本来就一口气拍 500 集的。

And we really, well, armed with all this knowledge, the reverse engineering, the technology, we understood that, well, at the very least soap operas can be made with the help of AI.

我们真的带着这一整套知识——反向工程、那些技术——明白了：至少肥皂剧是 能用 AI 做出来的。

And back then nobody understood this at all, when we said that we generate literally several minutes of content, close to an hour of consistent content, well, nobody believed it.

而那时候根本没人懂这个，我们说我们能生成好几分钟、接近一小时的一致性内容，没人信。

It like caused a wow.

大家的反应就是"哇"。

And next we needed to understand, well, whether people really want to watch.

接下来我们得搞清楚，人们到底想不想看。

And we made the app HERA.

于是我们做了 HERA 这个应用。

And here too, well, I'm showing that we have super-fast delivering.

这里也一样，我想说的是我们交付速度超快。

The guys, Kuka, Dias, Almaz, uh, they just, well, got together and say: "Basically, in 2 weeks we'll write Netflix, well, specifically the product streaming, the app, so that everything works fast, flies".

那几个小伙子，Kuka、Dias、Almaz，他们就那么凑一块儿说："我们两周就把 Netflix 写出来，就是产品化的流媒体应用，一切都要跑得飞快。"

We just did it in 2 weeks.

我们就两周做完了。

I'm like: "Damn, well, good luck".

我心想："我靠，那祝你好运。"

Yeah, yeah, well damn, that's hardcore.

对对，我靠，太狠了。

We, well we're like, back then in 2 weeks we'll also raise consistency 3x and that's it.

我们当时想的是，两周里还能顺手把一致性再提 3x，就完事了。

But they really did do it in 3 weeks, and they also, uh, before that we did everything very chaotically.

但他们真的三周就做完了，而且他们还——在这之前我们做东西非常混乱。

Well, product development, a whole lot of bets, fast development, and here they're like: "Uh, basically, for 2 weeks you won't touch us, we'll properly design a normal architecture and so on".

产品开发嘛，一堆 bet，开发很快；而这回他们说："这两周你们别来烦我们，我们要好好把架构设计出来等等。"

And they really did make a very high-quality app.

他们真的做出了一个质量非常高的应用。

And that, I mean, you need to understand product, specifically the product that you can poke at, use.

这就是说，你得懂产品，就是那个你能上手戳、能用起来的产品。

And even the content streaming, that's not even 50% of the real product.

而内容流媒体那部分，连真正产品的 50% 都算不上。

More than 50% of the product is competent product analytics.

产品里超过 50% 是靠谱的产品分析。

That's competent infrastructure for A/B testing.

是靠谱的 A/B testing 基础设施。

If it's B2C, then everything is based on product analytics.

如果是 B2C，那一切都建立在产品分析上。

And uh you release, uh, and you release it, well, at first on random.

然后你 release，一开始是随机放量。

B2C.

B2C。

Well, it's a story about funnels, and you need, well, to make the funnels super-optimal.

这是关于漏斗的事，你得把漏斗做到极致优化。

And very fast iteration of funnels.

还要非常快地迭代漏斗。

Uh, for example, well, literally in the app there are onboardings.

比如说，应用里就有 onboarding。

These are different, you can make 100-500 different variants of onboardings.

这能做出 100-500 种不同的 onboarding 版本。

Different variants of, uh, content ranking, how it shows the content, what it shows at the top.

还有内容排序的不同版本——内容怎么展示、什么放在最上面。

Uh, everything depends.

一切都取决于这些。

The size of the CTA buttons, Call to Action buttons, uh, what color they should be.

CTA 按钮、Call to Action 按钮的尺寸，还有该用什么颜色。

And at first we didn't understand this, that is, we learned all of this on the fly, in the diffuse, we, we learned all of it.

一开始我们不懂这些，都是边跑边学，在一团混沌里，全是这么学会的。

The thing is there's this, we even have a straight-up internal concept, lettuce green.

我们内部甚至有个专门的说法，叫"生菜绿"。

That is, you have to make, if it's B2C and you want high conversion, you have to lower your aesthetic, uh, preferences a little bit and make more lettuce-green colors, more of those network-y ones.

就是说，如果是 B2C、你想要高转化，就得把自己的审美偏好往下压一点，多用生菜绿这种颜色，更 网感 一点的。

And we even have this meme — it's vibrating buttons that literally vibrate so that the person definitely taps them.

我们甚至还有个梗，就是会震动的按钮，真的在那儿抖，好让人一定去点它。

Uh, you need to show competent paywalls.

还得展示靠谱的 paywall。

Uh, and to find a competent one, you need to A/B test dozens of different ones.

而要找到靠谱的那一个，得 A/B 测几十种。

And that is, all of this goes into the 2 weeks of product.

这些全都压在那两周的产品里。

And a huge amount of work was done.

这里面的工作量巨大。

The guys are super professionals.

这帮人超级专业。

Almaz leveled up unbelievably in product analytics.

Almaz 在产品分析上进化得不可思议。

And HERA really did come out as a product.

HERA 真的作为一个产品做出来了。

Next, so when we started doing our first marketing, looking for people who understand marketing, and I

接下来，我们开始做最早的市场推广，找懂营销的人，然后我

I personally started, well, digging into it together with Almas, since he leads the product, I — the two of us started, like, really deeply, we decided to figure out what B2C products look like in general, how to build them properly at all.

我自己也开始钻研，跟 Almas 一起，因为产品是他在带，我——我们俩就开始往深里挖，决定搞清楚 B2C 产品到底长什么样，到底该怎么做才对。

And, uh, we just started looking for companies who understand this, taking consultations from them.

我们就直接去找懂这行的公司，向他们买咨询。

and just speedrunning how to competently build B2C products, their marketing.

就是速通「怎么把 B2C 产品和它的营销做漂亮」。

And, well, we started to understand that the funnel starts already with the ad, that is, the whole Netflix story, well, obviously, betting on us sitting down and writing some superhit right away with AI, that's, well, unrealistic, because, well, so far we've only built the technology, uh, well, and we can't do production yet there, like the Hollywood studios, which, well, that's a different business, finding a hit there.

我们开始明白，漏斗从广告那一步就已经开始了，也就是说 Netflix 那套路子，很明显，指望我们坐下来用 AI 直接写出一个超级爆款，这不现实，因为我们目前只是把技术做出来了，我们还做不了内容制作，做不到好莱坞片厂那样——那是另一门生意，得在那里面找爆款。

And we understood that we need to move data driven.

我们明白，我们得按 data driven 的方式往前走。

That is, we need to build a Data Driven funnel, starting from the ad, which the user—, that is, it was, well, not organic marke—, not organic marketing, but Paid marketing, that is, and this is called user acquisition marketing.

就是说我们得搭一条 Data Driven 漏斗，从广告开始，用户看到的那条——也就是说，那不是自然流量的营——不是 organic 营销，而是 Paid 营销，这个叫 user acquisition marketing。

Paid user acquisition marketing.

Paid user acquisition marketing。

And to figure it out, there's this huge industry there, there are companies there, huge unicorns, these are huge cashflow businesses.

要搞懂它——这背后是个巨大的行业，里面全是公司，全是巨型独角兽，都是现金流极大的生意。

And that's when I understood that B2C is not even remotely a roulette, as it turns out.

那时候我才明白，原来 B2C 压根就不是什么轮盘赌，差得远。

It's, actually, you can, well, really build huge cash machines there, if you just know all the playbooks.

其实在那里你真的能造出巨大的现金机器，只要你知道全部的 playbook。

On the contrary, there's no lottery there at all, the way it was, for example, a long time ago.

恰恰相反，那儿根本没有彩票这回事，不像很早以前那样。

A modern B2C Paid User Acquisition, uh, app is, essentially, data driven, huge funnels that you make add up, you make the unit economics add up, you make the finances add up.

现代的 B2C Paid User Acquisition 应用，本质上就是 data driven，是巨大的漏斗，你要把它算平，把单位经济模型算平，把财务算平。

Everything turns into a big financial model for you, which you then understand how to scale.

所有东西都变成一个大的财务模型，接下来你就知道怎么把它放大规模。

and you make out of — out of it a mega cash machine.

然后把它做成一台超级现金机器。

There, essentially, the notion of Product Market Fit isn't even really applicable there in that classic Y Combinator sense.

在那个领域里，Product Market Fit 这个概念——Y Combinator 那种经典意义上的——其实根本不太适用。

The laws there are completely different, it's a completely different world there.

那里的规则完全不一样，是一个完全不同的世界。

There are companies with a billion-dollar ARR there, that make some kind of diet apps there, which have clones there, there are thousands of them, but they're not — and it's not because they built some cool product there, some super cool, I don't know, UI there in Steve Jobs style, not even close.

有些公司 ARR 上十亿，做的就是那种减肥 App，克隆品有成千上万个，但他们靠的不是——不是因为做了什么牛的产品、什么超牛的、我也说不好、Steve Jobs 风格的 UI，完全不沾边。

the whole product, starting from the person seeing that ad, ending with how he paid and how he then also renewed the subscription the next time.

整个产品，从一个人看到这条广告开始，一直到他怎么付款、下一次又怎么续订。

It's a huge Data Driven funnel.

这是一条巨大的 Data Driven 漏斗。

And a huge number of scientists, data analysts work at these companies.

这些公司里养着大量的科学家、数据分析师。

Uh, and for me this was, well, a big discovery.

这对我来说是个不小的发现。

I didn't understand that it's fully Data Driven.

我之前不懂这行是完全 Data Driven 的。

And in B2C companies I didn't understand that the most important expertise is Data Science and Data analytics.

在 B2C 公司里，最重要的能力是 Data Science 和数据分析——这一点我之前不懂。

And yes, after that product analytics.

对，再往后才是产品分析。

I didn't understand this.

这个我以前是不懂的。

And so, having figured all this out, we hired the top specialists there that you can find in Kazakhstan, guys, mega awesome ones, uh, analysts, marketers who get User Acquisition Marketing, uh, who subsequently, later, well, for Higgsfield too were able to figure out the organic side garbled, and now we're going to set that process up.

把这些都摸清之后，我们把哈萨克斯坦能找到的顶级人才招了进来，一帮特别猛的家伙，分析师、营销人，全都吃透了 User Acquisition Marketing，后来他们也帮 Higgsfield 把自然流量那一块搞明白了听不清，现在我们正要把这套流程搭起来。

But back then this was acquisition marke—, this was for HERA, right?

但那时候这套 acquisition marke——这是给 HERA 做的，对吧？

Yes, this was for HERA.

对，是给 HERA 做的。

And we built our first funnels and just started measuring data driven.

我们搭出了第一批漏斗，开始 data driven 地做度量。

And in HERA there are both real series and AI generated series.

HERA 里既有真人拍的剧集，也有 AI generated 的剧集。

And how many series were there at — at the peak?

那巅峰的时候里面有多少部剧？

HERA's.

HERA 的。

At the peak, damn, I can, yeah, dozens,

巅峰的时候，我靠，我可以说，几十部，

which you generated yourselves, right?

都是你们自己生成的，对吧？

Uh, some — some part we generated ourselves, some part we bought.

有一部分是我们自己生成的，还有一部分是买来的。

Uh, there's a large number, mm, it's really a big industry where you can buy them, these series.

这类剧的量特别大，这真的是个成规模的行业，能在里面买到这些剧。

New series are expensive, the ones that are exclusive on top of that, they cost several hundred thousand dollars there.

新剧贵，尤其是带独家的，一部要几十万美元。

Non-exclusive series, which, for example, are in other apps, they cost much less there.

非独家的剧，比如别的 App 里也有的那种，就便宜得多。

You can get them in batches there at various prices, for low-quality ones you can even buy for a few thousand dollars, for more high-quality ones, for a couple dozen thousand dollars.

可以成批买，价格分档，低质量的几千美元就能拿下，质量好一些的，两三万美元。

And a typical series is like 100 episodes, each one a minute long.

典型的一部剧差不多是 100 集，每集一分钟。

Yeah.

嗯。

Yeah.

嗯。

And there I — there's video there, there's audio there, there are subtitles there, right, like normal material, right, there's everything you need there, there's some hook for the next episode, right, and these companies that, well, overtook Netflix in, uh, in revenue, that's not production, that's, that is, well, that's purely big B2C funnels like this, data driven marketing machines, financial cashflow machines.

里面有视频、有声音、有字幕，跟普通片子一样，该有的都有，还留着通向下一集的钩子，而那些在 revenue 上超过 Netflix 的公司，做的不是内容制作，纯粹就是大体量 B2C 的这种漏斗、data driven 的营销机器、财务上的 cashflow 机器。

And we tried, yeah, to understand how they work, and we managed to figure it out.

我们试着弄懂它们是怎么转起来的，还真弄懂了。

And we built the funnels, and before that we started looking at, uh, watch-through rates AI versus non AI.

我们搭好了漏斗，在这之前就已经开始看完播率了，AI 对比 non AI。

And then our first signal, which we understood, was that, uh, there's no difference.

然后我们读到的第一个信号就是：没有差别。

People with — people, the target audience there, it's mostly adult women, like 45 plus, 50 plus.

人们——目标受众主要是成年女性，45 岁往上、50 岁往上。

For them there's basically no difference what to watch.

对她们来说，看什么其实无所谓。

If they start watching, their watch-through funnel will be the same, as for — for the series it'll be the same as for a non-AI one ной.

她们只要开始看，完播漏斗就是一样的，AI 剧的完播漏斗跟非 AI 的ной一模一样。

And if you put a paywall in the same place there at, I don't know, at the fifth episode, then the conversion percentage will be identical for those who buy and those who — for both the AI series and the non-AI one уно, if you put the PW in that same place.

如果把 paywall 卡在同一个位置，比如说第五集，那付费转化率的百分比是一样的，AI 剧和非 AI 剧уно都一样，只要 paywall 摆在同一个位置。

And that's, well, our signal that you really can disrupt the market with AI generated content.

这就是我们拿到的信号：确实可以用 AI generated 内容把这个市场掀了。

And an even stronger, uh, thing, a more powerful learning that we found out on top of this, is that the creatives, specifically the ad creatives in the ad network, and we, well, use Meta, TikTok there, the big ad networks.

还有更猛的一点，我们额外学到的一条更有分量的 learning，就是创意素材——具体说是广告网络里的广告创意素材，我们用的是 Meta、TikTok 这些大广告网络。

I can also tell you in detail about them, how they work on the inside.

它们内部怎么运作，我也可以细讲。

And, uh, ad creatives are literally, well, an ad video that you — when you put on Reels or TikTok there, it comes up for you.

广告创意素材说白了就是一条广告视频，你打开 Reels 或者 TikTok 的时候，它就刷出来了。

And its metrics, they can even exceed the metrics of a real, uh, piece of the series, because you have more control.

它的数据甚至能超过真人拍的剧集片段的数据，因为你的控制力更强。

If you generate — with real content you have to shoot it, you shot it once and then only at the editing level can you control it there, somehow set up a hook, somehow properly, mm, into those 5, 7-3 seconds there when the user is watching, somehow competently control the video like that, edit it so that the user actually stops, watches it through.

如果是生成的——实拍内容你得先拍，拍完那一遍之后就只能在剪辑这一层控制了，想办法插个钩子，在用户观看的那 5 秒、7 秒、3 秒里恰当地处理好，把视频控住、剪对，让用户真的停下来、看完。

Then you need him to click through.

接着还得让他点进来。

that's CTR, and CPM is so that, well, a large number of people see it across a certain whole network there, so that there's a big click-through conversion, uh, from the video onward into the App Store, and then CPA, so that the person actually buys, and then you need to count his retention so that we can understand the ROI of this ad video.

这就是 CTR；CPM 则是要让足够多的人在整张网络里看到它，让视频往下到 App Store 的点击转化率够高；再往后是 CPA，得让人真的掏钱；然后还要算他的 retention，这样我们才能算出这条广告视频的 ROI。

Uh, you can understand the ROI directly.

ROI 是可以直接算出来的。

And the cool B2C companies, they, everything, they even factor people's salaries into this economics entirely, to cut off.

厉害的那些 B2C 公司，什么都算，连员工工资都全部并进这套账里，为了话被截断

they don't know likely "in order to know" the ROI, literally of one single ad video.

好算出疑为「为了知道」单独一条广告视频的 ROI。

Its whole ROI, when the payback will come, there are benchmarks there, metrics, there's a good one there, these companies there, they work on a long влонun horizon.

它完整的 ROI、什么时候回本，这里都有 benchmark、有指标，有一套好的标准，这些公司做的都是长线влонun。

Even 12 months is considered a good payback in this market.

在这个市场上，12 个月回本都算好的。

Strong players, the cool ones, they get payback in 6 months.

强的玩家、猛的那批，6 个月就回本。

Well, huge, the bigger the scale, the more the payback stretches out, up to 12 months too.

规模越大，回本周期越往后拖，能拖到 12 个月。

That is, you showed the user, uh, an ad, and he only pays back for you within 12 months.

就是说你给用户看了一条广告，这个用户要 12 个月才把成本挣回来。

That's how, well, the B2C mega-giants work.

B2C 那些超级巨头就是这么玩的。

And, uh, that's how, well, the market is set up.

这个市场就是这么个结构。

And we understood that we get all these numbers, these metrics, really good ones, because the entry point is video.

我们发现这些数字、这些指标我们都拿得很漂亮，因为入口就是视频。

And if it, well, is low-quality, then all the metrics further down will be bad, bad.

视频质量一旦不行，后面所有指标就全都难看，特别难看。

And we understood that our AI crea— beats non AI creative.

我们发现我们的 AI crea——跑赢了 non AI creative。

And here for us this was a huge insight, a huge learning, that AI creatives are better, because there's more control.

这对我们是个巨大的洞察、巨大的 learning：AI 创意素材更好，因为控制力更强。

And coming back to how the ad networks work, uh, the ad networks in the big social networks are huge machine learning blackbox algorithms.

回到广告网络怎么运作这件事：大社交平台里的广告网络，就是巨大的 machine learning blackbox 算法。

Uh, they work on the basis that they, uh, these are auctions, machine learning auctions, where your ads, all the advertisers' ads, compete among themselves for the slot заместо to be shown to the user.

它们的运作基础是拍卖，machine learning 拍卖，你的广告和所有广告主的广告互相竞价，抢那个展示给用户的位置заместо。

And Meta as the provider, its job is, uh, to set the most optimal price, uh, of the auction, so that everyone comes out winning.

Meta 作为平台方，任务是把拍卖价格定到最优，让所有人都赢。

Well, and first of all, of course, Meta.

当然，第一个赢的肯定是 Meta。

Meta will be the one winning.

Meta 会是那个赢家。

If the right people see the right ad, the ad will pay back, so advertisers will earn and will invest more money further into Meta.

对的人看到对的广告，广告就能回本，广告主就能赚钱，就会继续往 Meta 投更多钱。

Well, buy more ads from Meta.

也就是从 Meta 买更多广告。

Meta actually itself gives out low-interest loans.

Meta 其实自己就在放低息贷款。

If you're a big B2C business, you can Data Driven, you show that you're buying tens of millions of dollars a month of traffic there.

如果你是大体量的 B2C 生意，你可以拿 data driven 的数据出来，证明你每个月买几千万美元的流量。

You can come to Meta, they'll even open a credit line for you, and you can take money from Meta and pour it back in and earn.

你可以直接去找 Meta，他们甚至会给你开一条授信额度，你从 Meta 拿钱，再把钱砸回去买量，然后赚钱。

That is, it's a huge industry, it's very far from the Silicon Valley industry.

所以说这是个巨大的行业，跟硅谷那一套离得非常远。

That is, these are businesses, this is more like private equity businesses.

这些生意更像是 private equity 那一类。

There you go.

就这样。

And they're also financed differently.

它们的融资方式同样也不一样。

And all of this, we spent 2 months figuring all of it out.

这一整套东西，我们花了 2 个月才全都摸透。

And the machine algorithms, uh, the advertiser himself can set what he wants to optimize.

机器算法这边，广告主自己可以设定他想优化什么。

And Meta's machine Learning Blackbox will optimize this itself.

然后 Meta 的 machine learning blackbox 就会自己去优化这个目标。

Those are just as huge machine learning models there, comparable to an LLM there — those are really, well, huge machine learning models in size.

那些同样是超大的 machine learning 模型，体量能跟 LLM 相提并论——真的是尺寸巨大的 machine learning 模型。

And they, uh, you can, for example, set: "I want to optimize the number of installs".

你可以比如这样设定：「我要优化安装量」。

And then you just pass this objective function straight into the machine learning algorithm.

然后你就把这个 objective function 直接扔进 machine learning 算法里。

It starts learning, it starts — first there's an exploration stage there, a learning stage, literally at Meta, which lasts, uh, a number, some number of steps.

它开始学，一上来先有 exploration stage、learning stage——Meta 那边字面上就是这么叫的——会持续一定数量的 step。

You can look right in the documentation, uh, what does this mean?

你可以直接翻文档，看这是什么意思。

And it learns.

它就这么学。

At first it shows at random to different audiences.

一开始它随机投给不同人群。

That's literally like an exploration stage.

这就是标准的 exploration stage。

Then exploitation.

然后是 exploitation。

It finds the audience and that's it, it starts optimizing.

它找到人群，就这么定了，开始优化。

And you can tune it for whatever you want.

你想按什么目标调都行。

You can tune it literally for money, for, uh, revenue, for the number of purchases there.

可以直接按钱来调，按 revenue，按购买次数。

If you, for example, you have some hot-selling product and you want to sell it in large quantities, you can optimize so that Meta learns, uh, to sell a large quantity of this product.

比如你手上有个畅销货，想走量卖出去，那就可以这么优化，让 Meta 学会把这个货大批量地卖掉。

Or else you can, uh, teach it to optimize money directly for you, for your business.

或者你也可以教它，让它直接给你的生意优化钱。

That is, if you've built all this correctly, data driven, then it turns simply into a cash machine.

就是说，只要你把这一整套按 data driven 搭对了，它就变成一台现金机器。

Meta — Meta's machine algorithm optimizes by itself, finds the audience by itself.

Meta——Meta 的机器算法自己做优化，自己找人群。

You don't need at all — uh, whereas earlier the industry looked like you needed a data analyst who'd sit down, or a marketer there, well, a long time ago it was fashionable there, uh, imagine who your client is, and he's middle-aged there, he probably likes this.

完全不需要——以前这行是这样的：你得有个数据分析师坐下来算，或者一个营销的人，很久以前流行这么干：「设想一下你的客户是谁，他中年，他大概喜欢这个」。

Well, now all of that is, well, that's last century's marketing there.

现在这一套全是上个世纪的营销了。

Now, on the contrary, you need to give Meta as little as possible — no bias at all about who your client is, because Meta better — its machine learning algorithm will itself understand better who your client is.

现在恰恰相反，你要尽量别给 Meta 任何关于「你的客户是谁」的偏见，因为 Meta 更清楚，它的机器学习算法自己就能更准地判断谁是你的客户。

And you maybe never even thought that this is our — could be our potential client.

你甚至可能压根没想过，这个人会是我们的潜在客户。

And it will, if you, for example, optimize revenue, then it will directly show it straight to people who are of that income level, who at this given moment, because it knows everything about the users.

比如你优化的是 revenue，它就会把广告直接投给收入水平对得上的人、此刻正处在那个状态的人，因为它对用户什么都知道。

And it will — it literally knows, right at this moment this person wants to buy this product, and it will show him this ad.

它字面意义上就知道：此时此刻这个人想买这个产品，然后就把这条广告推给他。

And that's just unreal.

这简直离谱。

And if you know how to set this up, then you can build huge cash machines.

只要你会调这套东西，就能造出巨大的现金机器。

That is, well, and we understood it on ourselves, once we understood how it works, then we, well, we also built a machine, of course.

我们自己就体会过，等我们搞懂它怎么转之后，当然也造了这么一台机器。

We, like, we released, and we immediately, in less than a month, ran up to a million ARR

我们一 release，不到一个月就直接冲到了一百万 ARR

…the field, right?

…这个领域，对吧？

Yes.

对。

I mean, we just understood how this works.

就是说，我们搞明白了这套是怎么转的。

At the beginning, like, the D2C stuff there was random, we mostly just went viral, somehow it all happened by accident.

一开始那个 D2C 完全是瞎打，我们主要是靠爆，一切都是撞出来的。

Here we just built it data driven, and it turned into a cash machine.

这次我们是照 data driven 搭起来的，结果变成了一台印钞机。

Millions, payback.

百万级，回本周期。

Uh, uh, well, I'd need to double-check the latter, but it was less than a year.

呃，后面那个数我得再核一下，不过是不到一年。

And we were like: "Damn, this can be scaled, this is cool".

我们心想：「靠，这东西能规模化，太牛了。」

And, well, and coming back again to the learning that, about the machine auction, I mean it does everything for you, if you know how to build this.

再回到那个心得，就是机器竞价那套——只要你会搭，它就全替你干了。

And several companies, well, several there, of course, there are a lot of companies like that in the world, but not such a huge number.

有几家公司——说几家，当然，世界上这种公司不少，但也没那么多。

If small business out there knew how to build this, such a data-driven process, it could make money too.

小生意要是也会搭这套 data-driven 的流程，一样能赚钱。

It's just clear, for some small business out there, well, the person just buys ads on Meta somehow at random.

很明显，那些小生意，人就是随手在 Meta 上买点广告，瞎买。

But to build a system like that, well, that takes a whole lot of great engineering time.

而要搭出这么一套系统，得砸进去大量顶级工程师的时间。

Well, literally, like, Alzhan who I talked about, he was building the infrastructure there together with Shatan, our data analyst.

就拿我刚说的 Alzhan 来说，那套基础设施是他跟我们的数据分析师 Shatan 一起搭的。

I mean this is huge data engineering, it's a lot of integrations, and you have a lot of, uh, strea-, streamings of data and you need to make a huge data engineering infrastructure to match all of it up with each other.

这是巨大的 data engineering，要接一堆集成，数据流一大把，得做一套庞大的 data engineering 基础设施，把这些东西全对上。

Because if it's the iPhone, Apple, that's the only, well, where the good money is, Apple has very strong privacy, and Apple doesn't let you attribute the user directly, and you can't, you can't hook the whole funnel.

因为如果是 iPhone、Apple——只有那儿才有像样的钱——Apple 的隐私管得极严，Apple 不让你直接归因到用户，你就没法把整条漏斗串起来。

And this is where super high IQ data engineering begins.

这时候就轮到 super high IQ 的 data engineering 上场了。

Well, without a great data engineer with a high IQ there, it's impossible to figure this out.

没一个高 IQ 的牛数据工程师，这块根本搞不明白。

So we got lucky that, well, we found Shatan, who right now actually leads organic for us, but back then at that moment he was, he was actually a physicist.

我们运气好，找到了 Shatan——他现在管我们的 organic，但当时他其实是个物理学家。

Uh, from Nazarbayev University.

从 Nazarbayev University 出来的。

And he and I were together in a libertarian community, and that's also where we hired our first founding ML Ops guy, Anvar.

我跟他都在同一个自由意志主义社群里，我们第一个 founding ML Ops Anvar 也是从那儿招来的。

I mean, by the way, Anvar is also a 2016 graduate of the incubator.

对了，Anvar 也是孵化器 2016 届的。

Yes, Anvar is, well, right now a superstar in infrastructure.

对，Anvar 现在是基础设施领域的超级明星。

I mean about, about him, a top engineer of the top labs, uh, they have a very high opinion of Anvar, back when we would sit down for deep dives.

顶级实验室的顶级工程师对 Anvar 评价极高，当年我们坐下来做 deep dive 的时候就这样。

Anvar, a really strong impression on the biggest people.

Anvar，给那些最大牌的人留下的印象特别深。

Well really, well, well like the top ones in the Valley there, top managers, who also do research, so he made a really strong impression on them.

真的，就是硅谷那些顶级高管，还自己做 research 的那种，他给他们的印象特别深。

And so it turns out, they were building this whole infrastructure, the engineering, and the guys who were in marketing set up the right media buying.

所以就是他们把这整套基础设施、工程那块搭起来，marketing 那边的人把买量调对。

I mean it's pretty hard to set all of this up, but once it's, once it's all set up?

这一整套配起来挺难的，但一旦全配好了呢？

It turns into a cash machine where you can just put money in, build a financial model, and it pays back.

它就变成一台印钞机，你只管往里放钱，搭好财务模型，它就回本。

and it scales, and it scales predictably.

而且能放量，还是可预测地放量。

And the companies that learned to do this, uh, in the end they have the same infrastructure.

而那些学会这么干的公司，最后基础设施都长一个样。

Everyone arrives at the same solutions.

所有人都走到同一套方案上。

Everyone has roughly, like, Big Query, Google, like Click House, uh, AWS Red Shift set up properly.

大家差不多都是 Big Query、Google，或者 Click House、AWS Red Shift，都配得挺规范。

Everyone has, uh, Tableau BI on top of it there, everyone has roughly the same dashboards.

上面再盖一层 Tableau BI，大家的 dashboard 也都差不多。

Everyone has the same funnels that they test.

测的漏斗也都一样。

Everyone has everything the same.

什么都一样。

And the only differentiator for all these companies, uh, and there are, well, a lot of companies like that, well, even in one field, they start competing with each other, because, well, it's really just money out of thin air, it's of course cash machines.

而这些公司唯一的差异点——这类公司真不少，哪怕同一个赛道里也一堆，他们开始互相卷，因为这真的就是凭空印钱，当然是 印钞机器。

And, uh, they all have the same infrastructure, the top ones, the ones with hundreds of millions of ARR there or tens of millions of ARR, if it's a new market.

头部那批人的基础设施全一样——那些几亿 ARR 的，或者新市场里几千万 ARR 的。

The only differentiator is the creative teams that shoot the ads themselves, because at this point, well, well, you need to shoot in different ways.

唯一的差异点是创意团队，他们自己拍广告，因为到这一步，拍法得各不相同。

I mean, uh, and everyone kind of finds their own audience through these ads.

每家都靠这些广告找到自己的受众。

Some go more into UGC and shoot UGC, hire a large number of UGC actors there, some build their own studios.

有的更往 UGC 走，就拍 UGC，签一大堆 UGC 演员；有的自己开摄影棚。

And the differentiation is at the level of the creatives.

差异化就发生在创意素材这一层。

And here we realize, we, since we can generate video, we can generate anything at all, this expertise of ours disrupts, well, this advertising business really heavily.

而我们清楚，既然我们能生成视频，我们什么都能生成，我们这身本事能把广告这门生意狠狠颠覆一把。

And here we realize that, well, here we really can, well, disrupt a lot.

我们清楚，这块我们是真能大幅颠覆的。

And actually, well, we see that long-term.

而且说实话，我们看的是长期。

I think we'll come back to this again anyway, because it's, well, a huge business, a huge market, and we have huge expertise in it.

我觉得我们早晚还会回到这块，因为这是一门大生意、一个大市场，而我们在这上面积累了巨大的专长。

There you go.

就这样。

And do you remember when you made the decision that you were shifting all your efforts over to the current image-to-video, right?

那你还记得吗，你们是什么时候决定把所有精力都转到现在这个 image-to-video 上的？

Yes.

对。

Yes.

对。

And then, when we, well, we've got millions of ARR, HERA is growing, we understand how to do this.

后来呢，我们手上有几百万 ARR，HERA 在涨，我们也知道该怎么干了。

We built the infrastructure in 2 months.

我们用 2 个月搭起了那套基础设施。

Usually it's like, well, everyone would tell us, usually it's like $5M, you can come into this market starting from $5M.

一般是这样——大家都会跟我们讲——一般得有 500 万美元，这个市场得从 500 万美元起才能进。

And you have to have experts who've done it before, like a team of superstars in the B2C market, who are actually, well, on top of that not Silicon Valley people usually, I mean you have to look for them all the way out there in China, in Cyprus, around the world.

而且你得有做过这事的专家，一支 B2C 市场的超级明星团队，而这些人通常还不是硅谷的人，得满世界去找，中国、塞浦路斯都得找。

We were like: "Damn, well we're engineers, we can reverse-engineer anything at all, we'll reverse-engineer it".

我们心想：「靠，我们是工程师啊，什么东西都能逆向出来，那就逆向。」

Well and in the end we pulled it off.

结果还真让我们搞成了。

Mm-hm.

嗯。

And then, uhh, next comes the question of scaling.

然后接下来就是放大规模的问题。

I mean this market, if we're still talking about AI Netflix and competitors pour in traffic, well, they buy traffic for hundreds of millions even.

这个市场——我们说的还是 AI Netflix，竞争对手 也在灌流量，他们买流量一买就是好几亿。

There's this concept of high season, when at New Year's, for example, people very, well, eagerly buy, uh, subscriptions, and in High Season you can scale the marketing budget 10-20x.

有个说法叫 high season，比如新年那阵子，人特别愿意买订阅，High Season 里营销预算能放大 10-20x。

If you know how to do it, you can scale it, and on those days you get the most money.

只要你会玩，就能把它放大，那几天你赚得最多。

High Season, that's certain specific dates.

High Season 就是那么几个特定日子。

and the companies, uh, there are huge ones.

而且有的公司体量非常大。

And plus you need to raise for this, uh, big money, so you can pour in traffic and so you can grow like a venture company, because, well, we're a startup, we're a purely venture startup.

另外还得为这事融一大笔钱，才能往里灌流量，才能像一家拿风投的公司那样长，因为我们是创业公司，纯粹走风投路线的创业公司。

We have investors who expect a venture story from us.

我们的投资人期待的是一个风投级别的故事。

And here what happens to us is, uh, we don't understand how to move forward, because this doesn't, doesn't fit a venture story at all.

于是我们这儿就卡住了，不知道下一步怎么走，因为这事儿跟风投的故事根本对不上。

If we want to do HERA, we have to go into private equity and raise huge rounds, literally huge ones there, in order to start doing AI Netflix.

如果我们要继续做 HERA，就得去找 private equity，融特别大的轮次，真的特别大，才能开始做 AI Netflix。

Uh and obviously, uh, raising private equity for a venture startup is completely, well, quite a very risky path, and it was unclear how we should move.

而很明显，一家风投型创业公司去融 private equity，这条路相当冒险，我们不知道该怎么走。

And in the end, yes, we had to, well, and in fact, probably, this, if you go do it, it's still, probably, more performance marketing than, uh, technical, yes, more of that, like, we exactly, when we spent 2 months on this, we even sat our ML engineers down there to build infrastructure, the ones who, well, know how to make top models.

最后确实只能这样——其实你真要去做这事，它说到底更像 performance marketing，而不是技术活儿；我们那 2 个月就是这么过的，甚至把会做顶级模型的 ML 工程师按在那儿搭基础设施。

And actually, well, anyway there, engineers making top models, if you're, well, just venture, well, you, you there, we know a huge number of examples of companies that became unicorns in a very short time and, uh, to go out to the market there and try to raise $500M right after the seed there, that's pretty, well, or you have to have a huge network in private equity.

而且说实话，能做顶级模型的工程师……如果你只是走风投这条路，我们知道一大堆例子，有些公司很短时间就成了独角兽，而刚拿完种子轮就去市场上想融 5 亿美元，这相当……要么你在 private equity 里得有极大的人脉。

Mm-hm.

嗯。

or, well, it'll be, yeah, unclear.

要么就，是啊，说不清。

And on top of that the cost is high, since we see that startups that are way weaker than us technically are already raising good rounds, Series A,

而且成本很高，因为我们看到，技术上比我们弱得多的创业公司，已经在融到不错的 Series A 了，

series B.

series B。

And we're sitting there looking at teams that deliver way worse than us.

我们就这么看着那些交付比我们差远了的团队。

Mhm.

嗯。

Products way worse than ours.

产品比我们差得多。

Time to market speed way worse, slower than ours.

time to market 的速度也差得多，比我们慢。

Raising big rounds.

却在融大额的轮次。

And we're like: "Well, something's not right here."

我们就想：“这不太对劲吧。”

And we're like, well, we've got a lot of jokes, like, about our competitors, like, they only exist because at that moment we got distracted with HERA, because the moment we showed up, that was it, and all their users started switching over to us.

我们内部有一堆拿竞争对手开的玩笑，说他们能存在，纯粹是因为我们那阵子被 HERA 分了心，因为我们一冒头，就到此为止了，他们的用户全开始往我们这边转。

And, well, they don't even come close to releasing as fast as we do.

而且他们发版的速度，跟我们根本不在一个量级。

So yeah.

就这样。

And we, yeah, decided, well, made the decision.

然后我们，对，就决定了，做了这个决定。

Do you remember roughly when that decision was made?

你还记得这个决定大概是什么时候做的吗？

Well, it was happening over a long time here, right?

这事儿拖了挺久才成形的，对吧？

It was happening over a long time, yeah, gradually.

拖了挺久，对，是一点点来的。

Uh, because there's also, well, possibly both a good and a bad trait.

因为还有一个特质，可能既是优点也是缺点。

So generally, basically, with us, with a lot of the guys, like with me, with Almaz, with the boys, with the engineers, uh, it's that when we see something new, super interesting, we're like: "Damn, we gotta reverse-engineer it, we gotta figure it out."

整体上，我们这儿很多人——我、Almaz、兄弟们、工程师们——只要看到什么新东西，特别有意思的，就会想：“靠，得逆向出来，得搞明白。”

And so when we found out about these B2C cash machines, that such a thing even exists, we're like: "Damn, this is so interesting, we gotta figure out how this works.

所以当我们知道有 B2C 印钞机这种东西、世上居然有这种玩意儿，我们就想：“靠，这也太有意思了，得搞明白它怎么运作。

Let's figure it out."

咱们搞明白它。”

That's a whole separate company, really.

那完全是另一家公司了。

Yeah yeah yeah.

对对对。

Let's figure it out.

咱们搞明白它。

That's where we learned about payments.

我们就是在那儿摸清了支付。

Well, that's a totally separate thing, how it goes in payments, there's a whole separate funnel there that you can build smartly, hack.

支付那块完全是另一码事，里面有一套单独的漏斗，可以很讲究地去搭、去 hack。

And payment services are a whole separate world, how it's all built.

支付服务本身就是另一个世界，它那套构造。

Like Airwallex, there's a company that recently became a unicorn, super fast-growing.

比如 Airwallex，这家公司最近成了独角兽，增长超快。

Like on 20VC, for example, the founder was on.

比如 20VC，创始人上过那期节目。

Like it disrupted Stripe, disrupted PayPal.

它把 Stripe 颠覆了，也把 PayPal 颠覆了。

And we used Airwallex there and all the other services.

我们当时就在用 Airwallex，还有其他所有服务。

We even met founders there from Stanford who are also building payments stuff "payймо" — payments / Paymo, and we're like: "Damn, we started swapping ideas there, you could actually do it super cool data-driven, infrastructure-wise, right."

我们甚至在那儿认识了几个 Stanford 出来的创始人，他们也在做支付 "payймо" — payments / Paymo，我们就想：“靠，我们就开始互相换想法，基础设施层面完全可以做得非常 data-driven，特别酷，对吧。”

And basically, we got distracted hard.

结果我们，说白了，被带偏得很厉害。

In the end, yeah, we had to, well, come back.

最后，对，只能，嗯，收回来。

We're like: "Damn, basically we got distracted, let's do this, well, something's off.

我们就说：“靠，我们跑偏了，来吧，这事儿有点不对。

Weak teams and weak products out there are raising "делают организт" garbled huge rounds.

外面那些弱队、弱产品都在融 "делают организт" 含糊 巨额轮次。

Basically we need to, uh, well, show "кост" garbled, likely "класс" how it's supposed to be done, right.

说白了我们得，嗯，露一手 "кост" 含糊，疑为 "класс"，让人看看这事该怎么做，对吧。

Yeah.

对。

And we're like: "That's it, we sit down and do it."

我们就说：“行了，坐下来干。”

And how much time did it take to build it, to, well, kind of get to the release on March 31st in the end?

那做出来花了多长时间，最后走到 3 月 31 日 release 那天？

Uh, well, from that moment, I think, the whole product was built in about, like.

从那个时间点算，我觉得，整个产品大概是在，就是。

Like Higgsfield AI, the first version, the one that, well, took off right away, yeah, uh, about 4 weeks total.

Higgsfield AI，第一个版本，就是一上线就爆了的那个，对，总共大概 4 周。

Awesome.

牛。

I mean we sat down, and we weren't releasing anything, we completely shut ourselves off from the world.

就是我们坐下来，什么都不发，彻底跟外界断了。

Mhm.

嗯。

We were completely full focus.

我们完全 full focus。

By the way, we had a regimen back then, we all agreed on it, that we don't get distracted by calls, or investors, or anyone.

顺带一提，我们当时定了个作息，所有人都说好了：不被电话打断，不被投资人打断，谁都不行。

Full focus.

Full focus。

We come in, like, everyone comes in, I don't remember, of course, whether everyone agreed on 6:00 in the morning or 7:00 in the morning.

我们来，所有人都来，我当然记不清了，是约的早上 6:00 还是早上 7:00。

And everyone actually came in like that.

而且大家真就那么来了。

Everyone, all 4 weeks everyone comes in at the same time, we just grind, we hold the regimen.

所有人，整整 4 周所有人都同一时间到，就是死磕，守住这个作息。

7 days a week, right?

一周 7 天，对吧？

7 days a week we hold the regimen.

一周 7 天，守住这个作息。

From 7:00 in the morning till midnight.

从早上 7:00 到半夜。

Well, until, yeah, we realize that we need to, well, go to sleep, because performance already dropped and that's it.

一直干到，对，意识到该去睡了，因为 performance 已经掉下来了，就这样。

And in 4 weeks we, that's when it was awesome for us sentence garbled.

我们 4 周里，那阵子真的挺猛 此句含糊。

So back when we were looking for a killer designer for HERA, that's when we found Madi.

当初给 HERA 找厉害设计师的时候，我们找到了 Madi。

Without Madi the product definitely wouldn't have worked out, because design was a huge deciding factor.

没有 Madi，这产品肯定做不成，因为设计起了决定性作用。

And plus there's, well, the company culture too.

再加上还有公司文化。

So, well, obviously, everyone says you gotta talk to users, do customer interviews and so on, but here we're more, probably, on Peter Thiel's side.

当然，所有人都说要跟用户聊、要做用户访谈等等，但这一点上我们大概更站 Peter Thiel 那边。

Peter Thiel, on the contrary, is of a different opinion, that you don't need to talk to users.

Peter Thiel 恰恰相反，他的看法是：不需要跟用户聊。

He says the opposite, like, great companies were created by people close to the autistic spectrum, who basically don't like talking to people.

他反过来讲，伟大的公司都是接近自闭谱系的人造出来的，那种压根就不喜欢跟人打交道的人。

And we literally did exactly that.

我们就是照这么干的。

We didn't talk to anyone at all, no custdev of any kind, we didn't do anything.

我们完全没跟任何人聊，什么 custdev 都没做，什么都没做。

Before that we did the opposite and, actually, for some reason it didn't work.

在那之前我们是反着来的，说实话，不知为什么就是不管用。

We basically decided to sit down internally and do everything internally, the cool way, the way we like it.

我们干脆决定关起门来，全在内部做，按我们喜欢的、够酷的方式做。

We've got this killer designer Madi, who— Well, by the way, Madi isn't just design, he can code too, he can do prompt engineering too, and he knows ComfyUI, I mean he's just a beast.

我们有个很强的设计师 Madi，他——对了，Madi 不只是设计，他还会写代码，会做 prompt engineering，还懂 ComfyUI，就是头野兽。

And uh Almaz, his taste is super high, he's a seriously professional photographer, he's also a DJ, I mean he's, well, a really stylish dude.

还有 Almaz，他的品味极高，是相当专业的摄影师，还是 DJ，就是个很潮的家伙。

And obviously, well, Madi, Almaz, well, the product's gonna be stylish no matter what.

那明摆着，有 Madi、Almaz 在，产品怎么都会很潮。

Plus we've also got Ziya, Kairim, Danil and the guys from the creative industry.

再加上还有 Ziya、Kairim、Danil，还有创意行业出来的那些人。

ML prompt engineer there, Seryoga, Rus, Marat, I mean, quality is not in question there, infrastructure is already all dialed in, the product is dialed in.

ML prompt engineer 那边有 Seryoga、Rus、Marat，质量根本不是问题，基础设施早就调顺了，产品也调顺了。

And, well, as soon as we released on March 31st it was just an instant hit, basically, the product instantly became a hit, people immediately started buying subscriptions.

然后我们 3 月 31 日一 release，立刻就是爆款，产品瞬间就火了，人们马上开始买订阅。

And what's interesting, the people who bought on day one, they have huge retention.

有意思的是，第一天买的那批人，retention 高得吓人。

I mean we just built internally a perfect product that right away has huge day one retention.

就是说我们关起门来做出了一个完美的产品，一上来 day one retention 就极高。

The ones who bought on, well, huge revenue retention specifically from the people who came from day one.

那些买了的人，revenue retention 极高，尤其是第一天进来的那批。

And on the contrary, the people who bought on the first day are like total mega-users.

反过来说，第一天就下单的那批人，简直就是一群 mega 用户。

They have some kind of unreal retention, even relative to other days.

他们的 retention 离谱得不真实，跟其他日子比也是。

They're like super power users.

就是些超级 power user。

How did they find out that you were releasing?

他们怎么知道你们要 release 的？

Ah, Twitter.

啊，Twitter。

I mean, you made some kind of announcement

就是说你们发了个预告

and we immediately, like, so our whole marketing team, the one that was building all this data-driven infrastructure for the marketing, we've got a separate stream, we're like, every day.

我们马上就，我们整个市场团队，就是搭那套 data-driven 营销基础设施的那批人，我们单独开了一条线，每天都在推。

We have to build a Twitter playbook.

我们必须做出一套 Twitter playbook。

Because we understand that all our competitors are on Twitter.

因为我们清楚，所有竞争对手都在 Twitter 上。

We're just like, let's do this data-driven, scrape everything, figure it out, build a graph of how Twitter works, because nobody, I had never even been on Twitter before and nobody on the team had ever used Twitter at all.

我们就说，来，用 data-driven 的方式把这些全爬下来，搞明白，建一张图，看 Twitter 到底怎么运转，因为没人——我以前根本不玩 Twitter，团队里也从来没人用过 Twitter。

And we're like, we're just reverse-engineering Twitter, we're scraping data analytics there, and start studying everything super hard.

我们就是在逆向 Twitter，一边爬数据做分析，一边开始疯狂研究。

We've got Nargiz, Ali — those are our marketers, product marketers.

我们有 Nargiz、Ali，他们是我们的市场，产品市场。

They start analyzing everything that Sultan scrapes.

他们开始分析 Sultan 爬回来的所有东西。

We're trying to reconstruct the playbook, and Sultan, he figures out, like a purely professional ad guy, how to do the messaging and the copy right.

我们试着把那套 playbook 还原出来，Sultan 则用纯专业广告人的路数，琢磨消息点和文案该怎么写才对。

But we, the whole prep— so, development is going on and in parallel the preparatory work for marketing is going on.

但我们，整个准备——就是，开发在推进，同时市场那边的准备工作也在推进。

And we, well, start studying there what Product Led Growth, uh, what PLG looks like, the playbooks.

我们开始研究 Product Led Growth 是什么，PLG 长什么样，那些 playbook。

how PLG properly, how the frameworks work.

PLG 到底该怎么做，那些框架怎么运作。

We start studying all of that, we sync up too.

这些我们全都开始研究，也一直在同步。

Every, like, Almaz

每个，那个，Almaz

he's like: "I'll study it like 2 hours a day today, I'll be reading blogs about PLG there, how to sell, how SLG / «L Grows» works.

他说:“我今天开始每天学两小时,去读 PLG 的博客,学怎么卖,学 SLG 怎么运作。

Then we sit down, and he gives everyone this presentation, right, like, here are these playbooks, this is how it's done, and everything has to be connected, how the funnel also looks different.

然后我们坐下来,他给所有人讲了一套演示,就是有哪些 playbook、该怎么做,而且所有环节都得串起来,漏斗长得也不一样。

And everything worked in this mega-coordinated way like that.

整套东西就这么高度协同地跑起来了。

At the same time, uh, everyone was collecting datasets, I mean everyone in the company was collecting datasets for post-training, and filtering them — the ML guys spun up these automated systems to label everything.

同时,数据集是所有人一起收的,全公司都在为 post-training 收数据集;过滤这块,ML 那帮人搭了自动化系统来打标。

And like our marketers, they were sitting there, the graphic designers, the designers, everyone, literally the first version of the product wasn't just, well, the marketers doing marketing — the marketers also took part in the product itself, because some of the effects were put together by, for example, Ali, our marketer, he literally went and did it, basically that effect is his.

我们的市场同学也一起坐着,还有平面设计师、设计师,所有人——产品第一版根本不是市场的人只管做市场,他们也直接参与产品本身,因为有些效果就是他们攒出来的,比如我们的市场同学 Ali,他真就自己上手做了,那个效果基本就是他的。

There you go.

就这样。

And uh well that's how smoothly everything was worked out.

整个配合就是这么顺。

And, well, and, probably, also the fact that we have our own top prompt engineers in-house — Ruslan, Seryoga and Marat, three top prompt engineers.

还有,可能也因为我们内部有自己的顶级 prompt 工程师——Ruslan、Seryoga 和 Marat,三个顶级 prompt 工程师。

And we were making the product for them, basically, internally we were showing it to our own prompt engineers, they gave feedback, and that's how we iterated fast.

我们本质上就是在给他们做产品,内部先拿给自己的 prompt 工程师看,他们给反馈,我们就这么快速迭代。

And that's why when we released, right away, basically, it became a hit without all those iterations, like, talking to users.

所以我们一发布就直接成了爆款,根本没走那些迭代、找用户聊天那一套。

Yeah.

对。

Yeah.

对。

By the way, very good point, interesting point.

顺便说,这个点很好,很有意思。

Indeed, if you go down the path of like lean startup, customer development, you often end up creating some kind of lowest common denominator, right, if you listen to everyone and do everything you're told, you end up like in that joke, the committee that tries to build a horse and ends up making a camel, right, because it listens to everyone and tries to accommodate everyone.

确实,要是走精益创业、customer development 那条路,做出来的常常是某种最小公约数——你要是谁的话都听、别人说什么就做什么,就会变成那个段子里的委员会,想造一匹马,最后造出一头骆驼,因为它谁的话都听,想迁就所有人。

Uh, well there was a question, actually, related to the competitive landscape, but it feels to me like we've more or less already covered it, right, there's, there's Runway, there's Pika.

本来还有个问题是关于竞争格局的,但我感觉我们差不多已经聊过了,有 Runway,有 Pika。

Uh, well, the only thing, I do have one on competitors.

不过关于竞品,我还真有一个问题。

So Veo, they dropped this speak feature of theirs.

Veo 不是刚放出他们那个 speak 功能嘛。

You dropped yours after them.

你们是在他们之后放的。

So it turns out you built the product after the market did.

这么说等于你们是跟在市场后面做的产品。

So how is adoption going for you right now?

那你们现在的采用情况怎么样?

Well, it's only a day old, right, so it's too early to say, right, still, yeah, we shipped it on Friday, uh, no, on Thursday, right, 2 days, yeah.

它才上线一天,现在下结论还太早,是周五发的——啊不,周四,两天。

Uh, well, adoption is good — for a new product a pretty big percentage of generations right now is Speak.

采用情况不错,作为新产品,现在相当大比例的生成都是 Speak。

And we didn't expect that, because, well, we wanted to release the product as fast as possible.

我们没想到,因为我们只是想尽快把产品发出去。

And it differs from Veo 3, uh, first of all in control.

它跟 Veo 3 的区别,首先在控制力。

I mean Veo 3 is text-to-video.

Veo 3 是 text-to-video。

And, for example, uh, in Veo 3 you can't create your own avatar so that it looks the way you want, so that it has the voice you want, and you also can't do more than 8 seconds consistently, because it doesn't let you create your own character.

比如在 Veo 3 里你没法造一个自己的 avatar,让它长成你想要的样子、配上你想要的声音,而且超过 8 秒也做不到一致,因为它不让你创建自己的角色。

Veo 3, however you—, whatever you write, it'll obey it, but not, well, not, well, it's impossible to describe things perfectly with a prompt, with text — for example, you describe in one prompt, for example, Arman, like, a curly-haired Kazakh in a shirt.

Veo 3 你怎么写、你写什么,它都照做,但是……用 prompt、用文字根本没法描述得完美——比如你用一句 prompt 描述 Arman,说是个穿衬衫的卷发哈萨克人。

like Anvar's outward— oops, Arman's outward features, but you generate three videos and in the three different videos Arman will look different each time.

把 Anvar 的外——哦,Arman 的外形特征写进去,可你生成三条视频,三条里的 Arman 长得都不一样。

It'll be three different people, just similar to each other.

那会是三个不同的人,只是彼此有点像。

Speak is different in that you can straight up make AI influencers, I mean consistent ones, that you can — and uh that's how the industry works, that people create whole AI influencers, uh characters that can post, run social media, uh, or if it's B2B, then it can be some kind of AI, uh, copy of the boss, who talks to the employees, or like, uh, of a real person who does employee training.

Speak 的不同在于,你可以直接做 AI 网红,就是有一致性的那种,你可以——行业现在就是这么玩的:大家会造出一整个 AI 网红、一个角色,能发帖、运营社交账号;要是 B2B,那可以是老板的 AI 分身,给员工讲东西,或者是负责员工培训的那个真人的分身。

There you go.

就这样。

And we have more control in that respect, it comes out as more of a professional tool.

这方面我们的控制力更强,出来更像一个专业工具。

And that's probably, yeah, the main difference.

这大概就是主要区别。

Mhm.

嗯。

Talking about the geographic — the geographies of your users, I saw that you have a lot of users in Japan.

说到用户的地理——地域分布,我看到你们在日本用户很多。

What other countries are there?

还有哪些国家?

And how, how does usage differ depending on geography?

不同地区的使用方式有什么差别?

Yeah, for us, if you take specifically the fans, like the users who literally write huge guides, like PDFs, documents on how to use Higgsfield AI, prompt instructions, that's, yeah, and that's the Japanese, they just make these huge ones covering all our effects, just everything in detail.

要说真正的死忠粉,就是那种会写巨长指南的用户——PDF、文档,讲怎么用 Higgsfield AI、prompt 怎么写——那就是日本人,他们写的东西特别长,把我们所有效果都覆盖到,事无巨细。

What is it for?

写这些是给谁看的?

It's for prompt engineers who are far removed, for example, from special effects, or far from the professional production that only exists in the industry.

给那些离特效很远、或者离行业内部才有的专业 production 很远的 prompt 工程师看的。

Guides on how to use this, what it's needed for.

教你怎么用、这东西能干嘛的指南。

Just in Japanese.

全是日语写的。

And they write blogs, articles there, and Japanese bloggers shoot really cool YouTube videos in Japanese on how to use «кл».

他们还写博客、写文章,日本博主还拍 YouTube 视频,拍得特别好,用日语讲怎么用 «кл»。

Uh, well, that's really cool.

这太酷了。

And Korean users are also very active, especially on social media.

韩国用户也非常活跃,尤其在社交媒体上。

They do a lot — I mean if the Japanese ones write these blogs, uh, record tutorials on YouTube.

他们量特别大——日本用户是写这种博客、在 YouTube 上录教程。

Korean users, they record a lot of content specifically on Twitter, on, uh, on Threads, the guys told me, and on Instagram, I mean on social media, on how to use it in detail.

韩国用户则是大量在 Twitter、Threads——同事跟我说的——还有 Instagram 上产内容,就是在社交媒体上细讲怎么用。

Uhh, and then the standard ones for us, of course, well, number one is obviously the United States, America, then, like, the United King— Kingdom, uh, Germany.

然后就是常规的那些,第一名当然是 United States,美国,然后是 United King——Kingdom,再是德国。

A lot of users, a lot of users in Brazil, especially from TikTok.

巴西的用户非常多,非常多,尤其来自 TikTok。

Brazilians, well, they're these very stylish dudes, just very, well, a young nation, lots of stylish young people there, fashionable, and they used us through TikTok a lot, and obviously there are users in Southeast Asia too, because, well, big populations and there was huge traffic.

巴西人特别时髦,而且是个很年轻的民族,时髦、潮的年轻人特别多,他们大量通过 TikTok 用我们的产品;东南亚当然也有用户,人口基数大,流量特别大。

And uh, well, it seems, uh, here's the interesting part, a lot of users, specifically good ones, paying ones, in Saudi Arabia, I mean Saudi, well, we — I mean we didn't expect it, a country that, well, somehow they didn't even think about it.

还有个有意思的:沙特阿拉伯的用户特别多,而且是优质的、付费的,就是沙特,我们——我们真没想到,这个国家,之前压根没往那儿想过。

And when we were looking at the analytics, in some periods Saudi Arabia was in the top three by paying users,

我们看后台数据的时候,有些时间段沙特阿拉伯能排进付费用户前三,

and those are creators, right?

那些人是创作者吧?

Yeah, yeah, those are creators.

对对,是创作者。

Saudi creators. «Кеторы соуска»

沙特的创作者。«Кеторы соуска»

Very interesting.

很有意思。

But another thing I like is that for you every update is a new ad campaign.

我还喜欢一点:你们每次更新都是一次新的广告活动。

I mean every new update you do is an occasion for you to do something else, something else interesting across all the social networks.

就是你们每做一次新更新,都成了一个由头,能在所有社交平台上再搞点别的有意思的东西。

It's like what Apple does once a year.

这就像 Apple 一年做一次。

You do it every week, right?

你们每周都做一次,是吧?

Yeah.

对。

Almost every day, right?

几乎天天做,是吧?

You probably do, that's, that's cool.

估计是天天做,这真的很酷。

That's our playbook.

这就是我们的 playbook。

And in general, Sequoia has an accelerator, damn, I forgot what it's called.

还有,Sequoia 不是有个加速器嘛,靠,我忘了叫什么名字了。

It's very trendy right now.

现在特别火。

And uhh in that accelerator they, mm, they have, uh, an article about AI and startups and how they see startups.

在那个加速器里,他们有篇文章讲 AI 和创业公司,讲他们怎么看创业公司。

And they divided startups into three categories.

他们把创业公司分成了三类。

And the first category, I think, was startups that have some kind of mega deeptech advantage thanks to huge capital, like, well, they raised like 500 million, a billion, and they can sit in stealth building some technology.

第一类,我记得是那种靠巨额资本吃到某种超级 deeptech 优势的创业公司——就是融了 5 亿、10 亿美金,能躲在 stealth 里闷头做技术的那种。

Then the third one I forgot, you can look it up in the blog.

第三类我忘了,可以去博客里看。

And the second one is when it's very high, well, a boiling broth, when it's a highly competitive environment, for example, LLMs or video generation, tons of products, tons of releases, and everyone is trying to deliver faster than everyone else.

第二类是烧得滚开的一锅汤,高度竞争的环境,比如 LLM 或者视频生成,产品特别多,release 特别多,所有人都在比谁 deliver 得更快。

And this playbook, it's super, well, for me, when I read it, it became super clear to me.

这套 playbook 太牛了,当时我一读,一下就全明白了。

And, well, and there they write out the reasons why you can win and why you can lose.

而且里面还写了为什么能赢、为什么会输。

And to lose — that's because, well, you just lose the head-on competition.

输,就是因为,就是纯粹在正面竞争里输掉。

And that's exactly what the market looks like.

市场就是长这样。

I, and there's the same thesis from Ben Horowitz, for example.

我,Ben Horowitz 也有一模一样的论断。

Ben Horowitz thinks that there are a ton of companies, they're all alike, they're all doing something similar, nobody understands what.

Ben Horowitz 认为公司多得要死,彼此都长得差不多,都在做类似的事,谁也说不清具体在做什么。

And his, uh, his thesis is that it's like embryonic development, that all embryos look alike, so right now all early-stage companies look alike, they're all competing with each other.

他的论断是,这像胚胎发育——所有胚胎看上去都一样,所以现在所有早期公司彼此都像,全在互相竞争。

Then as the market grows, the industry grows, a new world appears, uh, the economy will look different, some of those embryos will survive and they'll grow up and become some kind of huge products that we don't even expect at this point.

等市场长大了、行业长大了,新世界出现了,经济形态也变了,这些胚胎里会有一些活下来,长成我们现在压根想象不到的巨型产品。

And we look at it roughly the same way, I think, like, well, in 5 years Higgsfield and, well, well, probably we'll be doing some, well, I hope some totally unreal things that I can't even imagine right now.

我们大概也是这么看的,5 年后 Higgsfield 会怎样,我希望我们在做的是现在我根本想象不出来的那种事。

And so it turns out, uh, all the companies are alike, and the winner is whoever just wins the competition.

所以说到底,所有公司都长得一样,谁赢下竞争谁就赢。

And, well, we kind of got lucky that we have these super ambitious guys, engineers of the highest world-class level, and we, well, really just move faster than anyone on the market.

我们算是运气好,团队里这帮人野心极大,是世界最顶级的工程师,我们确实就是全市场跑得最快的。

And the problem is that you can't release on just any day.

问题在于,不是哪天都能 release。

I mean, ideally you need to release.

就是说,理想情况下才该 release。

Well, in the ideal case, yeah.

嗯,最理想的情况,对。

Ideally it's Monday, because the whole week is ahead of you.

最理想是周一,因为整整一周都在后头。

And March 31, when we did our first release, that was a Monday.

3 月 31 号我们第一次 release,那天正好是周一。

And Monday morning by, by San Fran, 9 a.m. San Francisco time, that's the best time, yeah.

周一早上,按旧金山时间早上 9 点,那是最好的时间点,对。

And then obviously, well, if you released on Monday, then there's no point releasing on Tuesday.

然后很明显,周一 release 了,周二再 release 就没意义了。

You can't do one or two releases back to back in a day.

不能一天里连着来一两个 release。

And the next release is better done on Wednesday, Thursday.

下一个 release 最好放周三、周四。

But Thursday isn't that good anymore, because the week is ending, and Saturday-Sunday is already bad.

但周四已经没那么好了,一周快结束了,周六周日就更糟。

Well, there'll be very few sales.

销量会非常少。

So the beginning of the week is ideal.

所以周初最理想。

And it turns, well, you're sitting on those working days where the real peak is, well, Monday, Tuesday, Wednesday, and we have way more ideas and releases than days.

于是你就守着这么几个工作日,真正的高峰就是周一、周二、周三,可我们的点子和 release 比日子多得多。

And we have a huge backlog right now.

所以我们现在积了一个巨大的 backlog。

Of releases, I mean, yeah, we've piled up a huge backlog of releases, because we can't and, well, it's also a shame, I don't know, to release it on a Saturday, and we have to, well, pile up the backlog.

就是 release 的 backlog 堆得特别高,因为我们没法发,而且也舍不得,比如放到周六去 release,只能这么攒着。

Mhm.

嗯。

So yeah.

就这样。

And really, yeah, we're, well, thinking about how, basically, how, how to, yeah, work with this.

我们确实也在想,这事到底该怎么处理。

And I think a really interesting point is that, uh, well, there was, well, this topic is still popular, the seven, yeah, the Seven Powers of Hamilton Helmer, right, the concept, yeah, of competitive advantage, right, like economies of scale, economics, like, network effect, counter-positioning, switching costs, branding, cornered resource, process power.

我觉得有个特别有意思的点,就是到现在还挺流行的那套东西,Hamilton Helmer 的 Seven Powers,对吧,竞争优势那个概念——规模经济、经济那个网络效应、反定位、转换成本、品牌、独占资源、流程壁垒。

And it's as if in startups, and your startups, for example, the competitive advantage is speed, right, all these, let's say, powers that we named, that fits better the fifty-year-old companies, right, that have already had time to gather some, some assets, to sort of, uh, dig a deep moat around their castle, right?

而在创业公司这边,比如你们这种,竞争优势就是速度——刚才点的那几种力,更适合五十岁的老公司,人家有时间攒下资产,在自己城堡外面挖出一条深深的护城河,对吧?

And speed for, what's it called, what is the asset, what is the competitive advantage of a startup — it's its speed.

而速度,怎么说来着,创业公司的资产是什么、竞争优势是什么——就是它的速度。

A company that does like seven releases a day, oops, a week, naturally, in a year it will have 365 releases.

一家一天做七个 release,哦,一周七个,那一年下来自然就是 365 个 release。

A company that does one big release a year.

另一家一年做一个大 release。

Well, there's kind of no question here at all, no competition.

这根本没什么可比的,不存在竞争。

And your speed really amazes me.

你们的速度是真让我叹服。

That's why, uh, I hadn't looked at Pika in a long time, hadn't looked at Runway and the others in a long time.

所以我很久没看 Pika 了,也很久没看 Runway 那些了。

I just looked now, while you and I were talking, at their examples.

刚才跟你聊天的时候我顺手翻了一眼他们的样片。

And like, of course, it's night and day.

那当然是天差地别。

And it's only been like 2 months since your release.

而距离你们那次 release 才过了 2 个月。

I think more and more users will be on your side.

我觉得会有越来越多用户站到你们这边。

And speaking about — do you have a Discord, where you're with the users at all, or not that actively?

说到——你们有 Discord 吗,会在里面跟用户混,还是没那么活跃?

Yeah, yeah, we have a Discord.

有,有,我们有 Discord。

The only thing is, well, we have very few resources to actually maintain it, but we, yeah, always try to dig through it.

唯一的问题是我们没什么人手真正去维护它,但我们一直在尽量把它清干净。

We have our young employees Madiyar and Mazhay two names run together in the ASR, the two of them somehow manage, on top of the operational marketing processes, to also clear out Discord, support users, uh, do refunds and so on.

我们有两个年轻同事,Madiyar 和 Mazhay 两个名字在 ASR 里连读了,就他们俩,除了市场那边的日常运营,还能顺手把 Discord 清一清、给用户做 support、做退款这些。

So if we talk about, well, abstracting away from your field, right, uh, video generation, how do you see the landscape at the moment overall?

那如果跳出你们这个领域,视频生成,你现在整体怎么看这个版图?

globally, like if we're talking about speed, I think the industry has changed a lot over the year, how everyone looks at this in general.

从全局看,如果讲速度,我觉得这一年行业变了很多,大家整体是怎么看的。

I mean, obviously, the first big breakthrough was DeepSeek.

很明显,第一个大突破是 DeepSeek。

If you even look at how investors used to treat it, obviously, but everyone there always thinks that Silicon, Silicon Valley is, basically, that's where all the coolest, smartest people are, we're like the absolute most, like 'San Francisco is back', everyone was writing that, right, that's it, basically, we're coming back.

你哪怕只看以前投资人的态度就知道了,大家总觉得硅谷嘛,最牛、最聪明的人全在那儿,我们最厉害,"San Francisco is back",所有人都在写这个,对吧,行了,我们回来了。

And then DeepSeek comes along and just destroys everyone, just wipes the floor with them.

然后 DeepSeek 一来,把所有人按在地上摩擦。

Well, purely engineering-wise, ML, AI optimizations, DeepSeek did what, well, neither OpenAI nor Anthropic, nobody had done.

纯从工程、ML、AI 优化上讲,DeepSeek 做到了 OpenAI 和 Anthropic 谁都没做到的事。

Way faster, way more efficient, way cooler, with a smaller budget.

快得多、效率高得多、牛得多,预算还更少。

And, well, for example, John Schulman, he gave a lecture at Stanford where he broke down DeepSeek, and he was presenting it as, well, as a new thing.

比如 John Schulman 在斯坦福讲过一课,专门拆 DeepSeek,而且他是当成一个新东西来讲的。

Which obviously means OpenAI didn't have these innovations.

这显然说明,OpenAI 内部没有这些创新。

No matter how you slice it, OpenAI says that internally we're ahead of everyone, but we watch John Schulman's lectures, he, well, you can see it, he talks about it, since he's at Anthropic now, right, I think.

怎么说都好,OpenAI 嘴上说我们内部领先所有人,但我们看 John Schulman 的讲座就看得出来,他讲的就是这个,他现在在 Anthropic 嘛,我记得是。

Oh no, he's already at Mira Murati's, right, it turns out, yeah, that's a total superstar team over there.

哦不对,他已经去 Mira Murati 那边了,对,那边完全是全明星阵容。

And so, I mean, well, the first one is DeepSeek, then, well, also, if you look at our competitors who raised huge rounds, cool guys there, researchers from DeepMind, take Haiper for example — they've already, well, gone bankrupt, shut down; then there's Pika, well, well, I don't want to, of course, talk really badly, right, about competitors, but there's just a general opinion in the industry about where we are and where Pika is, specifically in terms of generation quality, product quality, visual quality and so on.

所以第一是 DeepSeek;再往下,看看我们那些融了巨额轮次的竞争对手,都是牛人,有 DeepMind 出来的研究员,比如 Haiper——已经破产关掉了;还有 Pika,我当然不想说竞争对手的坏话,但行业里就是有个共识:我们在哪儿、Pika 在哪儿,具体到生成质量、产品质量、画面质量这些。

Demi Guo, when she came to Kazakhstan in 2015, she won silver here at the IOI.

Demi Guo 2015 年来哈萨克斯坦的时候,在这儿拿了 IOI 银牌。

And she probably didn't even suspect that out of this country someone would give her such big compet-, competition, right.

她大概根本没料到,会有人从这个国家杀出来,给她这么大的竞、竞争。

Yeah, yeah.

对,对。

The thing is — well, and the attitude is different now, in that if, basically, in 2023 it was super trendy, CBC, if you're a researcher, if you have a PhD, Stanford, Berkeley, that's it, here's 80 million dollars for you, go disrupt.

而且现在的态度也变了——2023 年那会儿超级流行,CBC,你是研究员,你有 PhD,斯坦福、伯克利,行,给你 8000 万美元,去 disrupt 吧。

Mhm.

嗯。

You went through MIT, geniuses there.

MIT 出来的,那儿都是天才。

it went by, well, literally like 2 years, and everything is different.

结果过去,也就 2 年,全变了。

Like, obviously, if you hand it to researchers who have neither product expertise nor that startup culture of delivering and so on, well, it doesn't work.

很明显,你把钱交给一帮研究员,他们既没有产品经验,也没有创业公司那种把东西 deliver 出去的文化,这就是不 work。

And for us, uh, probably the benchmark company — well there's, basically, there's Cognition, there's Cursor, right, similar products, but completely, totally different, uh, philosophies.

对我们来说,标杆公司大概是这样——有 Cognition,有 Cursor,产品类似,但哲学完全、彻底不一样。

And for us Cursor is like a huge beacon, totally.

而 Cursor 对我们来说就是一座巨大的灯塔。

Someone we admire, someone we want to repl-, to repeat, to replicate.

我们佩服他们,想复、想复刻他们。

A company that showed how you're supposed to do GenAI right, and at the same time the company doesn't even build its own model.

这家公司示范了 GenAI 到底该怎么做,而它连自己的模型都不做。

I mean just super fast shipping, tons of releases, and they pick all the low-hanging fruit that exists, because there's a huge number of companies that probably even the viewers haven't heard of, that raised hundreds of millions of dollars.

就是纯粹的超快 shipping、极多 release,把世上所有低垂的果子全摘了;因为还有一大堆公司,观众可能听都没听过,却融了好几亿美元。

For example, yeah, Magic.

比如 Magic。

Poolside, Augment Code, that's the Sutter Hill Ventures one, I mean, and probably the viewers, even programmers, engineers, haven't heard of the third company, but they raised, well, if you add them all up, it comes out to more than a billion dollars.

Poolside、Augment Code,就是 Sutter Hill Ventures 投的那家,第三家公司观众里就算是程序员、工程师大概也没听说过,可他们全加起来融了超过 10 亿美元。

And where are they, and where is Cursor?

他们在哪儿,Cursor 又在哪儿?

I mean there's Windsurf, also a magnificent company too.

还有 Windsurf,也是家了不起的公司。

And recently a podcast came out with the founder on 20VC.

前段时间他们创始人上了 20VC 的播客。

I recommend everyone watch it.

强烈建议都去看。

And these companies, they show how to properly build a GenAI startup, the right playbook.

这两家公司示范了 GenAI 创业公司到底该怎么做,正确的 playbook 是什么样。

Only speed, and they pick all the low-hanging fruit, they don't go building some cyborg in stealth like Cognition, which, well, then nobody needs.

只拼速度,把低垂的果子全摘掉,而不是像 Cognition 那样在 stealth 里造个什么赛博人,造出来根本没人要。

And they just, well, the guys damaged their own reputation.

结果这帮人把自己的名声也搞坏了。

And — and what do you think, I mean, the fact that they have a pile of money in the bank, will it let them hold out, make a great product, or is the battle already won?

那你怎么看,他们银行里躺着一大笔钱,能不能撑到做出一个好产品,还是说这仗已经打完了?

I think the battle isn't won.

我觉得这仗还没打完。

We're still early stage anyway, and I think, well, they just need to transform.

毕竟大家都还在 early stage,我觉得他们就是得转型。

Mhm.

嗯。

So yeah.

就这样。

And if they, well, do transform, then of course, well, they can win, because the guys have super high IQ, they're all cool dudes there.

如果他们真转过来了,当然还是能赢的,毕竟那帮人 IQ 超高,个个都是狠角色。

Uh, but yeah, it's specifically the playbook that Cursor showed — how it's supposed to be done.

但对,正是 Cursor 用它那套 playbook 示范了该怎么做。

Even Cursor's marketing, it showed how to do GenAI marketing right.

连 Cursor 的营销都示范了 GenAI 的 marketing 该怎么做。

How they did it — they just hired engineers who were Twitter influencers.

他们怎么做的?就是专挑那些本身是 Twitter 网红的工程师来招。

I mean they literally just went around to all the engineers who were Twitter influencers, hired them onto the team so that they'd work for Cursor and constantly post about Cursor.

就是字面意义上把 Twitter 上有影响力的工程师一个个找过来,招进公司,让他们给 Cursor 干活,顺便天天发帖聊 Cursor。

And it's just this cool organi-, just guerrilla growth-hack marketing, while some other guys over there were trying to push it through B2B somehow, through Microsoft somehow, to push it from the top, to build some schemes, Cursor just, well, guerrilla marketing just won.

这就是很牛的自然……纯粹的游击式增长黑客打法;别人还在琢磨走 B2B、走 Microsoft、从上往下推、搭各种体系的时候,Cursor 就靠游击式营销赢了。

And, well, really cool.

真的挺牛。

one hundred percent organic, just, well, they tore up the market.

百分之百自然流量,直接把市场撕开了。

And Windsurf, the same thing.

Windsurf 也一样。

Windsurf, uh, they're actually a lot like us on Twitter, if you look at them.

Windsurf 在 Twitter 上其实跟我们很像,你去翻一下就知道。

And what do they do on Twitter?

他们在 Twitter 上干什么?

Our playbooks are similar, actually.

我们两边的 playbook 其实很像。

And if, well, you start comparing — interesting sentence cut off.

你要是去比一比——挺有意思的句子没说完。

So yeah.

就这样。

So for us, yeah, these two companies are the benchmark for how to do it right in GenAI garbled.

所以对我们来说,这两家公司就是标杆,GenAI 该怎么做才对此处 ASR 含糊。

Just quickly pick the low-hanging fruit.

就是快,把低垂的果子一个个摘掉。

What is the low-hanging fruit in, well, for example, in codegen, in your view, that Windsurf and, and Cursor picked up,

那在你看来,Windsurf 和 Cursor 摘掉的低垂果子,比如在 codegen 里,具体是什么,

just all the little things.

就是各种小细节。

Like literally little things at the level of some optimizations that take like 5 dash even 7 seconds of time.

字面意义上的小事,就是那种要花 5 到甚至 7 秒的地方,做点优化。

Like, I don't know, moving some copy-paste code around, moving something somewhere, navigation between folders, working with comments, code review.

比如复制粘贴的代码搬来搬去,把某个东西挪个位置,文件夹之间的跳转,处理注释,code review。

Mhm.

嗯。

All these things are just small features that the guys delivered very fast.

这些东西全都只是小 feature,而那帮人 deliver 得极快。

Mhm.

嗯。

We literally just recently started another season of the startup, well, the incubator, the Incubator, where about 100 developers will be building their own apps in 10 weeks.

我们前不久刚开了新一季的创业,就是那个孵化器 Incubator,大概 100 个开发者要在 10 周里做出自己的 app。

And what recommendations do you have for them?

你对他们有什么建议?

What can they do in, in such a short period?

这么短的时间里,他们能做出点什么?

Well yeah, right now the best

对,现在最好的

it's the time, of course, to build apps, because, well, the speed of development has dropped likely: the cost/barrier of development has dropped.

现在当然是做 app 的时候，因为开发的速度降下来了 他大概是想说开发的成本/门槛降下来了。

If what used to be, well, you needed a whole team of, like, senior engineers there, now it's one engineer, who in any case has to know engineering well, and without that there's no way.

以前需要一整个团队的资深工程师，现在一个工程师就够了——但这个人无论如何得把工程这块吃透，没这个免谈。

Well, for example, like, for example, like a backender, he doesn't need a frontender anymore.

比如说，一个后端工程师，他已经不需要前端工程师了。

If he's a good backender, he knows how to write code, then he'll write the front-end together with Claude.

他要是个好后端，会写代码，那前端他就跟 Claude 一起写出来。

Infrastructure he'll also write together with Claude.

基础设施他也是跟 Claude 一起写。

like some processes there, like business processes there, how to do payments, how to set all this up, that too he can do with Claude.

一些流程、业务流程，怎么做 payments，怎么把这些都配起来，他也能跟 Claude 一起搞定。

Well and there with ChatGPT and with the rest.

还有 ChatGPT，以及其他那些。

So one professional now can, well, cover a huge number of, well, things.

所以现在一个专业的人能顶掉一大堆活。

You can even see it in our company, the fact that, well, nobody had experience there.

这个在我们公司就看得到——我们这儿根本没人有那方面的经验。

Uh, that is, you have to understand that we now have 1,000 GPUs, as many as 200 garbled GPU count.

呃，你得知道，我们现在有 1000 张 GPU，足足 200 GPU 数字这里 ASR 乱了。

Mhm.

嗯。

That's, well, huge, well, I never knew.

这量很大，我以前根本不知道。

I remember, when I had, when we had 64, I came onto the podcast with Arman ASR: 'к карману' and, and it seemed to me that we were, basically, super cool.

我记得我们还只有 64 张的时候，我来上 Arman 的播客，当时觉得我们已经牛得不行了。

That is, now we have thousands of GPUs, and at the same time we have three infra engineers who had never had experience with GPUs before, that is just, well, high-IQ, charged up.

现在我们有几千张 GPU，而管这些的就三个基础设施工程师，他们以前从来没碰过 GPU——就是 IQ 高、劲头足。

There's Anvar, uh garbled names, he too made a very, very cool impression on top researchers and top managers in the Valley.

比如 Anvar，呃 这里名字听不清，他也让硅谷那些顶级研究员和高管印象非常非常深。

when we were doing deep dives, and he's the best graduate of RFMSh of his year there, a math olympiad guy, optimizes neural networks.

当时我们在做 deep dive，他是 RFMSh 那一届的最佳毕业生，数学奥赛出身，做神经网络优化。

And our third junior ML infra engineer, Arman, he's actually an iOS engineer.

我们第三个 junior ML 基础设施工程师 Arman，他其实是 iOS 工程师。

That is, he — we hired him as an iOS engineer.

我们当初是当 iOS 工程师把他招进来的。

Then we shut down, well, paused HERA and I say: "Well, that's it, you're going to be an ML infra engineer".

后来我们把 HERA 关了——是暂停了——我就说：「行了，你以后就是 ML 基础设施工程师。」

And 2 months, that's how long he's been with us, or 3 months, well, that is, a really tiny bit, and he's already an ML infra engineer, he is.

他来了才 2 个月，或者 3 个月，就这么点时间，他已经是 ML 基础设施工程师了。

That is, well, without him we wouldn't have managed to scale there.

没有他，我们根本扩不上去。

That's how we scaled, because 24 hours by 7 coverage, always onsite.

我们就是这么扩起来的，因为 24 小时乘 7 天全覆盖，永远 onsite。

Uh, oops, on call, that is, infra.

呃，说错了，是 on call，就是基础设施那边。

I remember how, when we had only just started going viral on TikTok, and I just see in Grafana a huge clogged-up queue of people who want to generate a video for themselves.