AGI 不是新范式,是当前范式 + 1-2 个大点子
"There still might be one or two things missing on top of what we already know works… my betting is about 50/50 if that's the case."
"I can't see a world in which we will sort of realize in a couple of years this was a dead end. Continual learning, long-term reasoning, some aspects of memory — these are still unsolved."
大规模预训练 + RLHF + chain-of-thought 已经被验证过太多次,不会被推翻;但真要到 AGI,还差一两个 idea,Demis 自己估的概率是 50/50。
大 context window 是 brute-force 工作记忆,不是真正的记忆
"Even if you can have millions of tokens, there's still a cost to looking it up and finding the right thing that's actually relevant for the specific decision you've gotta make right now."
"It's pretty brute force currently, and that doesn't seem right. We're trying to store everything in that — things that aren't important, things that are wrong."
真正缺的是 hippocampus 式的"挑出此刻有用的那一段"机制。100 万 token 看起来大,处理直播视频 20 分钟就用完了。
没有 continual learning,agents 永远是半成品
"Not having continual learning currently is one of the things holding back agents from doing full tasks. They don't adapt well with the context that you're in."
"That's the missing piece for them being really kind of fire-and-forget and they'll figure it out themselves. They need to be able to learn about the specific context."
现在的 agent 能干阶段性活,但你没法"扔过去就不管"——因为它学不到你 setup 的特殊性,要拿到 full general intelligence,这是必跨的坎。
RL 现在被低估,AlphaGo 那一套正在被重新启用
"A lot of the things we're doing today — all the leading models with thinking modes and chain-of-thought reasoning — are aspects of what was sort of pioneered with AlphaGo, coming back now."
"We're sort of re-looking at some of those old ideas at scale today in a more general way, including things like Monte Carlo tree search and other ways of augmenting the RL on top of the reinforcement learning we already do."
Reasoning 模型不是新发明,是十年前 AlphaGo / AlphaZero 的方法被重新拉到 foundation model 之上;未来几年的进展也会从这里来。
蒸馏看不到理论上限,大模型一年后落到 edge
"A year later after one of our leading Pro models or frontier models goes out, half a year later, a year later, you will have them in the really tiny almost edge models."
"You have to build the biggest models to have the frontier capabilities. But one of our biggest strengths has been distilling and packing that power into smaller and smaller models very quickly."
小模型 95% 的能力 + 1/10 价格 + 更快迭代,是 Google 服务"十几个十亿用户产品"的现实约束;同时也意味着端侧、机器人、隐私场景的临界点已经到。
Reasoning 还缺一个对自己思考过程的 introspection
"Sometimes it will consider a move, it will realize it's a blunder, but it can't find anything better, so it kind of goes back to that move and does it anyway. That shouldn't be happening in a precise reasoning system."
"It can solve gold medal problems in IMO — super hard — but on the other hand it can still make basic elementary maths errors if you pose the question in a certain way. There's something about almost an introspection about its own thought process that I feel like is missing."
"jagged intelligence(锯齿状智能)"——同一个模型能拿 IMO 金牌,但下棋会明知是 blunder 还落子;Demis 认为只差一两个 tweak 就能补上。
还没出现"vibe-coded"的全球热销大作,说明工具到人之间还有 gap
"We haven't seen a triple-A game that tops the app store charts that was sort of vibe-coded yet. Why haven't we seen a kid making a hit game that sells 10 million copies? That should be possible given the effort that's gone in."
"I can do a prototype of a theme park in half an hour now, which took me six months back when I was 17. It still needs craft and human soul and taste."
Demis 17 岁做 Theme Park 用了半年,现在他半小时能 prototype 一个;但市场上还没出现一个"小孩 vibe-coded 出来卖 1000 万份"的热销品——所以工具或流程上仍有缺口。
Move 37 不算创造,能不能让 AI 发明 Go 才算
"It's not enough to come up with Move 37. That's pretty cool, very useful, but can it invent Go? That's what I want — a system that can invent Go if you give it a high-level description."
"You give it: a game you can learn the rules of in five minutes but it takes many lifetimes to master, beautiful aesthetically, but you can play in an afternoon. Then I'd want the thing I get back is Go."
真创造的标准不是在已知空间里找到一步妙棋,而是从一段抽象描述出发反向构造出 Go 这个游戏本身。这是 Demis 给 AI 创造力划的硬线。
Edge 模型应该开源,因为它无论如何都"暴露"
"Our edge models — the things we want to use for Android and glasses and robotics — it's best that they're open models, because they're vulnerable anyway once you put them out on the surfaces. So they might as well be actually fully open."
"Gemma's been about 40 million downloads in just two and a half weeks. We also think it's important for there to be Western stacks on open source — a lot of the Chinese models are excellent and currently lead in open source."
Google 的开源策略不是慈善,是策略上的统一:既然 nano 级模型一定要部署到设备上、必然能被逆向,那索性全开,顺带在西方开源生态里占住位置。
推理永远不会便宜到等于零 — Jevons 悖论
"I'm not sure inference will ever be essentially free. I think we'll just end up using whatever we can get our hands on — millions of agents, swarms of agents working together. All of that will use up any inference that's available."
"Even if energy is essentially zero with fusion or superconductors, there'll still be the physical creation of the chips. There'll be some bottleneck for at least the next few decades, so there'll still be rationing on the inference side."
"等推理免费"是个错觉。能耗也许能解决,但芯片产能撑不住"几百万 agent 同时跑"。所以效率永远是一个生产力杠杆,不是过渡阶段的临时痛点。
AlphaFold 模式可复用:大组合空间 + 清晰目标函数 + 数据/模拟器
"If the situation can be described as massive combinatorial search space — the more massive, the better — and you have a clear objective function, and enough data and/or a simulator that can generate in-distribution synthetic data, then with today's methods you can go a long way."
"Drug discovery is the same way. There is a compound out there that would solve this disease — if one could only find it. As long as the laws of physics allow it, the only question is how to find it efficiently."
不要再问"AI 能不能解决我的领域";问"我的领域能不能写成这三件事的组合"。能,就开做。
AGI 时间表 2030 + deep tech 是 10 年项目 = 必须假设 AGI 在中点出现
"If you start off on a deep tech journey today, you have to just consider AGI appearing in the middle of that journey. It's not bad necessarily, but you have to take that into account."
"What will the AGI system do with it? One thing I see happening is Gemini, Claude, or one of these general systems making use of AlphaFold-like specialized systems as tools."
不要规划一个"AGI 出来就废"的产品;要规划一个"AGI 出来反而能被它当 tool 调用"的产品。这个判断决定了 deep-tech 项目能不能活过下半场。
"You've been thinking about AGI longer than almost anyone. When you look at the current paradigm — large-scale pre-training, RLHF, chain of thought — how much of the final architecture for AGI do you think we already have, and what's fundamentally missing right now?"
"你思考 AGI 的时间几乎比任何人都长。看当前这套范式——大规模预训练、RLHF、chain-of-thought——你觉得 AGI 的最终架构里我们已经握住了多少?最根本缺的是什么?"
"The components that you just mentioned, I'm pretty sure, will be part of the final architecture for AGI. They've come such a long way now, and we've proven out so many things about what they can do. I can't see a world in which we will sort of realize in a couple of years this was a dead end."
"你刚才说的那几样,我很有把握会是 AGI 最终架构的一部分。它们一路走到现在,我们已经验证过它们能做的事太多。我无法想象两年后我们会回头说'这条路是死胡同'。"
"But there still might be one or two things missing on top of what we already know works. Continual learning, long-term reasoning, some aspects of memory — these are still unsolved. And how to get the systems to be more consistent across the board."
"但已知能 work 的这堆东西之上,可能还差一两样。Continual learning(连续学习)、long-term reasoning(长期推理)、某种意义上的 memory——这些还没解决。还有怎么让系统在所有维度上保持一致(consistency)。"
"It might be that the existing techniques can just scale up to that with some incremental innovation, or there's still one or two big ideas left that need to be cracked. I don't think it's more than one or two, if there are out there. My betting is about 50/50."
"也可能现有方法做点增量创新就能 scale 到 AGI;也可能还差一两个大点子要破。如果有,我估计也不超过两个。我个人下注是 50/50。"
"Of course at Google DeepMind we work on both."
"当然在 Google DeepMind,这两条路我们都在做。"
"Working with a bunch of agentic systems, the wildest thing to me is to what degree it's the same weights ev- over and over. So this idea of continual learning is so interesting, because right now we're sort of cobbling it together with duct tape — these dream cycles at night and things like that."
"我用各种 agentic 系统时最震撼的一点是——它的 weights 居然全程不变,反复同一份。所以 continual learning 这个概念现在格外有意思,因为我们目前就是用胶带糊起来的——比如所谓的'夜里 dream cycle'之类。"
"It's pretty cool, the dream cycles. We used to think about this with consolidation, with episodic memory. Actually, that's what I studied for my PhD — how the hippocampus works and integrates new knowledge gracefully into the existing knowledge base."
"Dream cycle 这个比喻挺好。我们以前在 consolidation(记忆固化)、episodic memory 这套框架里就讨论过。其实我 PhD 研究的就是这个:海马体(hippocampus)怎么把新知识平滑地整合进已有的知识库。"
"The brain does that amazingly well — during sleep, especially things like REM sleep, replaying back episodes that are important so that you can learn from it. In fact, our very first Atari program, DQN, one of the ways it was able to master Atari games was by doing experience replay. We borrowed that from neuroscience and replayed successful trajectories many times. That's way back in 2013, the dark ages of AI."
"大脑做这事做得极好——在睡眠期间,尤其 REM 睡眠,会把那些重要的事件回放,让你能从中学习。我们最早那个玩 Atari 的程序 DQN,它能掌握 Atari 游戏靠的就是 experience replay——这是直接从神经科学借过来的:把成功的轨迹反复回放。这都 2013 年的事了,放现在就是 AI 的'黑暗时代'。"
"And I agree with you — we're kind of using duct tape right now: shove it all in the context window. This seems a bit unsatisfying. Even though we're working on machines, not biological brains, you could potentially have millions or tens of millions sized context windows. There's still a cost to looking it up and finding the right thing actually relevant for the specific decision you've gotta make right now."
"我同意你的判断,现在确实是胶带工程:全塞进 context window。这做法不太对劲。虽然我们造的不是生物脑,理论上 context window 可以做到几百万甚至几千万 token,可能还能记得很完美——但'从那一大堆里挑出此刻这个决定要用的那一段',这件事本身仍然是有成本的。"
"That cost is non-trivial, even if you can potentially store it all. I think there's actually a lot of room for innovation in areas like memory."
"就算你存得下所有东西,这个'查找成本'也不是 trivial 的。我觉得在 memory 这块还有大量创新空间。"
"The wild thing is that a million-token context one is actually plenty big, honestly."
"百万 token 的 context window 老实说已经很大了。"
"It's plenty big for most things it should be used for. If you think about the context window as sort of equivalent to working memory — humans have a few digits, maybe a dozen, average of seven. We've got million or 10-million context windows. The problem is we're trying to store everything in that, things that aren't important, things that are wrong. It's pretty brute force currently, and that doesn't seem right."
"对大多数用途来说够大了。如果把 context window 类比成 working memory——人类只有几个数字,顶多十几位,平均 7 位。我们有上百万、上千万的 context window。但问题是我们试图把所有东西都塞进去——不重要的、错的全塞。这就是 brute force,直觉上不对。"
"And then the problem is, if you're now trying to process live video and you're just gonna naively record all the tokens, then actually a million tokens isn't that much — it's only like 20 minutes. So you need more if you want something that's gonna understand what's going on in your life over maybe a month or two."
"另外,如果你要处理实时视频,naively 把每一个 token 都录下来——百万 token 其实没多少,大概只有 20 分钟。如果你想让系统理解你接下来一两个月生活里发生的事,这点容量远远不够。"
"DeepMind has historically leaned into reinforcement learning and search — AlphaGo, AlphaZero, MuZero. How much of that philosophy is actually embedded in how you're building Gemini today? Is RL still underrated?"
"DeepMind 历史上一直押注强化学习和 search——AlphaGo、AlphaZero、MuZero。这套思想现在多大程度上还嵌在你们做 Gemini 的方式里?RL 是不是仍然被低估?"
"I think potentially it is. It sort of goes in ebbs and waves. We've worked on agents since the beginning of DeepMind. In fact, that's what we said we were working on. All of the Atari work and AlphaGo and most specifically — they're agent systems."
"潜在意义上确实被低估。这种东西就是一波一波。Agents 我们从 DeepMind 第一天就在做——其实我们当年公司定位说的就是这个。Atari 那套、AlphaGo、最重要的是这些都是 agent 系统。"
"What we meant by that is systems able to accomplish goals on their own, make active decisions and make plans. We were doing it in the domain of games to make it tractable, then increasingly complex games, things like StarCraft after AlphaGo — AlphaStar. We basically did all the games that were out there."
"我们当时定义 agent 是:能自主达成目标、能主动决策、能制定计划的系统。为了让问题 tractable,我们先在游戏域做——AlphaGo 之后是 StarCraft,做 AlphaStar。基本上当时存在的游戏我们都做了一遍。"
"And then the question is — can you generalize those models to be world models, or models of language, not just models of simple games or even complex games? That's what the last few years has been about."
"接下来的问题是——能不能把这些模型 generalize 成 world model,或者语言模型,而不只是某个游戏的模型。过去几年我们就在做这件事。"
"You can think of a lot of the things we're doing today — all the leading models with thinking modes and chain-of-thought reasoning — as aspects of what was sort of pioneered with AlphaGo, coming back now. There's a lot of work we did back then that is relevant today."
"你今天看到的事情很多——所有顶级模型的 thinking mode、chain-of-thought reasoning——本质上是 AlphaGo 当年开创的东西在以新形式回归。当年我们做的不少工作放到今天依然 relevant。"
"We're sort of re-looking at some of those old ideas at scale today in a more general way, including things like Monte Carlo tree search and other ways of augmenting the RL on top of the reinforcement learning we already do today. A lot of those ideas, both from AlphaGo and AlphaZero, are really relevant to where we are with today's foundation models — and I think a lot of that is what we're gonna see of the advances the next few years."
"我们现在是把那些老 idea 重新拿出来,放到 scale 上、放到更通用的语境下来用——包括 Monte Carlo tree search,以及其他给 RL 做增强的手段(在我们今天已经用的 RL 之上再叠一层)。这里很多 idea 来自 AlphaGo / AlphaZero,放在今天 foundation model 的语境下依然非常 relevant。我认为未来几年的大进展会从这里来。"
"Today you need bigger and bigger models to be smarter and smarter. But then we're also seeing distillation working — smaller models can be quite a bit faster. You guys have incredible Flash models. You're finding they're 95% as good as the frontier and at like one-tenth the price. Is that right?"
"今天你要更聪明就要更大模型。但同时蒸馏也 work 起来了——小模型快很多。你们的 Flash 模型很猛,你们发现它能做到前沿的 95%、价格只有十分之一。是这样吗?"
"I think that's one of our core strengths. You have to build the biggest models to have the frontier capabilities, but one of our biggest strengths has been distilling and packing that power into smaller and smaller models very quickly."
"这是我们的核心优势之一。你必须先造最大的模型,才会有前沿能力;但我们的一大强项就是把这种能力快速地'蒸馏'并打包进越来越小的模型里。"
"We invented the kind of distillation process — Jeff and Oriol and others — and we're still world experts in that. We also have a huge need to do it because we've got to serve the biggest probably AI surfaces there are: Search with AI Overviews and AI Mode, the Gemini app, and increasingly every product at Google — Maps, YouTube — has some aspect of Gemini in it. That's billions of users, more than a dozen billion-user products, and they have to be served extremely fast, efficiently, cheaply, with low latency."
"distillation 这套方法基本是我们发明的——Jeff、Oriol 和其他人——我们至今还是这块的全球专家。我们也'被迫'必须做好它:我们要服务的大概是世界上最大的 AI 部署面——Search 的 AI Overviews 和 AI Mode、Gemini app、还有越来越多 Google 的产品(Maps、YouTube 等)都内嵌了 Gemini。这意味着十几亿用户、十几个十亿用户级产品,它们必须极快、极高效、极便宜、低延迟。"
"That gives us a really important incentive to make these Flash and even smaller models — Flash-Lite models — extremely efficient. And hopefully that ends up being really useful for many of the workloads all of you use these for."
"这给了我们一个非常实在的 incentive,把 Flash 甚至更小的 Flash-Lite 模型做得极致高效。这些东西希望对在座各位的 workload 也有用。"
"Are there limits to the distillation process? Could a 50B or 400B model be as smart as a Mythos for today?"
"蒸馏有极限吗?50B 或 400B 模型能不能达到今天 Mythos 那种水平?"
"I don't think we've gotten to any kind of informational limit — or at least none of us know yet if we have. Maybe at some point there's an information density we can't get beyond. But for now, the assumption we make is that a year after one of our leading Pro or Frontier models goes out, half a year later, a year later, you'll have them in the really tiny almost edge models."
"我不认为我们已经触到任何'信息密度的物理上限'——至少没人知道。也许哪一天会撞到那个上限。但目前我们的工作假设是:某个 Pro / Frontier 模型出来之后大概半年到一年,你就能在端侧小模型里看到它的能力。"
"You see some of that goodness in our Gemma models, our Gemma 4 models — really amazing power for their sizes. So I don't see any theoretical limit yet. We're still pretty far off."
"Gemma、Gemma 4 你都能看到这种'同尺寸下能力突破'的迹象。我看不到理论上限,我们还离它很远。"
"That is really good — because one of the weirder things we're seeing right now is engineers can do like 500 to 1000 times the amount of work they were doing six months ago. There are people in this room doing a thousand X the work a Google engineer from the 2000s was doing. Steve Yegge talks about this."
"这很关键——因为我们现在看到的怪事是,工程师能干出半年前 500-1000 倍的活。这个房间里有人产出已经是 2000 年代某个 Google 工程师的一千倍。Steve Yegge 写过这个。"
"I think it's very exciting. Small models have many uses — cost obviously, but speed too. If you're collaborating with the system, in coding for example, you can iterate a lot faster. There's a lot of need for fast systems that maybe aren't quite frontier — 95%, 90% — but that's plenty good enough, and you gain back more than the 10% on iteration speed."
"非常令人兴奋。小模型用途很多——成本是一个,速度更是。如果你是和系统协作,比如写代码,小模型让你能极快迭代。市场上对'非顶配但够快'的系统需求巨大——做到顶配的 95%、90% 已经绰绰有余,你从迭代速度上拿回来的远超那 10% 的能力损失。"
"And then the other big thing is running these things on the edge — for efficiency, but also for privacy and security reasons. Different devices that process very personal information. You can also think about robotics — robots in your house. You're gonna want very efficient, very powerful local models, maybe orchestrated with bigger frontier models in the cloud, but you only delegate to that in certain circumstances."
"另一个大方向是端侧——出于效率,也出于隐私和安全。处理高度个人信息的设备,以及未来的机器人——比如你家里的机器人——你会希望它跑非常高效、强大的本地模型,只在特定时刻 delegate 给云端的 frontier 模型。"
"Perhaps you process all of the audiovisual feed locally, and that stays local. I could imagine that would be a very good end state."
"也许整个音视频流就在本地处理、永远不离开本地。我能想象那是一个很理想的终态。"
"Going back to context and memory — models are currently stateless. What would the developer experience even be like for someone using a continual learning model? How would you steer it?"
"回到 context 和记忆——现在模型是 stateless 的。如果有 continual learning 的模型,开发者体验会是什么样?你怎么 steer 它?"
"It's really interesting. Not having continual learning currently is one of the things holding back agents from doing full tasks. They're really useful for aspects of tasks right now, you can patch them together and do some really cool things, but they don't adapt well with the context that you're in."
"这是个很有意思的话题。目前 agent 没法干完整任务,根本卡点之一就是没有 continual learning。它们在某些任务环节里很有用,你可以把它们拼起来做出挺酷的东西,但它们没法适应你所在的具体 context。"
"That's the missing piece for them being really kind of fire-and-forget and they'll figure it out themselves. They need to be able to learn about the specific context you're gonna put them in. We have to crack that to get full general intelligence."
"这就是它们离'扔过去就不管了、自己搞定'的关键差距。它们必须能学到你给它们部署的具体 context。这件事不解决,我们拿不到 full general intelligence。"
"Where are we on reasoning? Models can do really impressive chain-of-thought now, but they still fail on things a smart undergrad wouldn't. What specifically needs to change?"
"reasoning 进展到哪了?模型能做 chain-of-thought 已经很厉害了,但还会在一个聪明本科生不会犯的地方掉链子。具体缺什么?"
"There's a lot of innovation left in the thinking paradigms. We're doing fairly simplistic things, fairly brute force. There's a lot of scope for monitoring the chain of thoughts, maybe interjecting midway through a thought process. I often get the impression with our systems and competitor systems that they're almost overthinking — almost getting into sort of loops."
"thinking paradigm 这块还有大量创新空间。我们现在做的比较朴素,比较 brute force。比如监控 chain-of-thought、在思考过程中半路介入,这都有空间。我对我们自己以及对手系统的一个观感是——它们几乎在 overthinking,会陷入某种循环。"
"One thing I sometimes like to do is play chess against Gemini. All the leading foundation models are pretty poor at games, which is interesting. It's very cool to look at the thinking traces, because chess is well-understood — I can tell quickly if it's going off on a tangent."
"我有时候会跟 Gemini 下国际象棋。挺有意思的现象是:所有顶级 foundation model 下棋都挺差。看它的思考轨迹特别有意思——国象是被充分研究过的,我一眼就能看出它有没有跑偏。"
"Sometimes it will consider a move, it will realize it's a blunder, but it can't find anything better, so it kind of goes back to that move and does it anyway. You shouldn't be seeing that happening in a precise reasoning system."
"有时候它会想一步,它意识到这是一招 blunder(送子),但又找不到更好的——然后它就回去把那一步走了。这种事在一个精密推理系统里本不该发生。"
"There are huge gaps still, but it may only be one or two tweaks required to fix those. That's why you get this kind of jagged intelligence. On the one hand, it can solve gold medal problems in IMO — super hard — but on the other hand, as we've all seen, it can still make basic elementary maths errors if you pose the question in a certain way."
"这种 gap 现在还很大,但可能只需要一两处调整就能补上。这种现象就是所谓的'jagged intelligence(锯齿状智能)'——同一个模型既能拿 IMO 金牌,又会在你换个问法之后犯小学算术错误。"
"There's something almost about an introspection about its own thought process that I feel like there's something missing there."
"我感觉它缺的是某种'对自己思考过程的内省(introspection)'。这一块还差点东西。"
"Agents are really big. Some would say they're hyped. I personally think they're just getting started, and it's totally insane. What does DeepMind's internal research tell you about where agent capabilities actually are right now versus the hype?"
"Agents 现在很热,有人说被炒过头了。我个人认为它才刚刚开始,而且 totally insane。DeepMind 内部研究告诉你,agent 能力的真实水平和外界 hype 之间是什么关系?"
"I agree with you. We're just at the beginning. You have to have an active system that can actively solve problems for you to get to AGI. That was always clear to us. So agents are that path, and I think we're just getting going."
"我同意你的判断,我们才刚开始。要走到 AGI,你必须有一个能主动替你解决问题的 active system——这件事我们从一开始就清楚。所以 agent 是必经之路,而我们才刚出发。"
"All of us are getting used to how to best work. How do you incorporate it into your workflow in a way that isn't just a nice-to-have but actually starting to do fundamental things? My impression is at the moment we're all experimenting on lots of things, but only in the last couple of months are we starting to find the really valuable places."
"现在每个人都在摸索:怎么把 agent 整合进 workflow,让它不只是'锦上添花',而是真正承担基础工作?我的判断是大家这几个月才开始找到真正有价值的落点。"
"The technology is probably only getting good enough for that to be the case — where it's not a kind of toy, nice demonstration, but actually really adding value to your time and efficiency. I see a lot of people setting off dozens of agents for like 40 hours, but I'm not sure I've seen the output yet quite justify that level of input. But I think it will come."
"技术可能也就是最近才到那个 threshold——不再是 toy demo,而是真的在为你节省时间、提高产能。我看到不少人一次性派 dozens 个 agent 跑 40 小时,但我还没看到产出能配得上这种 input 量。我相信它会来。"
"We haven't seen a triple-A game that tops the app store charts that was sort of vibe-coded yet. I've programmed and seen many nice demos. I can do a prototype of a theme park in half an hour now, which took me six months back when I was 17. It's mind-blowing — I get this feeling if I spent the whole summer working on it, I could make something really incredible."
"我们还没看到一个 vibe-coded 出来的 triple-A 游戏冲到 app store 榜首。我自己也写过、做过不少漂亮的 demo——我现在 30 分钟能做出一个 Theme Park 的 prototype,我 17 岁那年做这个用了 6 个月。这种对比令人震撼——我有种感觉,要是真给我一整个夏天死磕,能搞出非常牛的东西。"
"But it still needs craft and human soul into it and taste. That's something you have to make sure you still bring to whatever you're building."
"但它仍然需要 craft、human soul 和 taste。这些东西不管你做什么,都还得你自己注入进去。"
"It's not quite there yet — why haven't we seen a kid making a hit game that sells 10 million copies? That should be possible given the effort that's gone in. So something's still missing. Maybe it's to do with the process, or maybe it's to do with the tools."
"还没到位——为什么我们还没看到一个小孩做出一个卖 1000 万份的爆款游戏?按现在投入的 effort 量,这件事本该已经发生。所以某个东西还缺。可能是流程问题,也可能是工具问题。"
"How much of it will be autonomous? I don't think we'd see autonomous first — we'd actually see people in this room operating 1000X."
"哪一部分会变完全自动?我不觉得我们会先看到完全自动——我们会先看到这屋里的人产能拉到 1000X。"
"That's what you should see first. And then many of you — there'll be like games companies or other types of companies that have built some kind of best-selling app, best-selling game using these tools. That's what you should see first. And then more of that will get automated."
"先发生的是这件事。然后是在座的人或者新兴公司——做游戏的、做 app 的——用这些工具做出某种 best-selling 的产品。这是应该最先看到的。再之后才会有更多被自动化。"
"Part of it might be — we wanna discuss creativity. What I often say about that is, if we look at the things we've done, like AlphaGo. Very famously, the Move 37 in game two — and for me, I was waiting for a moment like that to start the science projects like AlphaFold. We started AlphaFold the day we got back from Seoul, which is 10 years ago now."
"另外一部分可能是创造力问题。我经常这么讲:看看我们做过的事——比如 AlphaGo,大家都知道第二局的 Move 37。对我个人,我一直在等那种时刻才好开启 AlphaFold 这种科学项目。我们从首尔回来那天就启动了 AlphaFold,距今刚好 10 年。"
"But it's not enough to come up with Move 37. That's pretty cool, very useful, but can it invent Go? That's what I want — a system that can invent Go if you give it a high-level description."
"但仅仅想出 Move 37 是不够的。Move 37 很厉害、很有用——但它能不能发明 Go?这才是我想要的——你给它一段高层描述,它能给我反向构造出 Go 这个游戏本身。"
"You give it: a game you can learn the rules of in five minutes but takes many lifetimes to master. It's beautiful aesthetically, but you can play it in an afternoon. That would be the high-level description I'd give. And the thing I get back is Go, right?"
"我给的描述是:一个 5 分钟能学会规则、一辈子学不完、视觉上美、一个下午能玩完一局的游戏。这就是我给的高层描述。然后我希望它返回给我的——是 Go 本身。"
"Clearly today's systems can't do that. So the question is why — there's something still missing there."
"今天的系统显然做不到。所以问题是为什么——这里还差一样东西。"
"Well, someone in this room might make it. Then the answer would be there's nothing missing — it just was the way we were using the systems. That might actually be the answer. It might be that today's systems are capable of it with a brilliant enough creative person providing that impetus, the soul of the project, and being au fait enough with the tools to almost be at one with them."
"也许这屋里就有人会先做出来。如果这事真发生了,那答案就是'其实什么都不缺,只是之前没人这么用'——这也可能就是答案。今天的系统也许已经具备这个能力,只是要等到一个足够创造性的人——他能注入项目的'灵魂',又对工具熟到几乎'人机合一'——来把它召唤出来。"
"I could imagine that happening if you experimented with the tools all day and all night, like probably many of you are doing, and you combine that with proper deep creativity — something more incredible could be done."
"如果你日夜都在折腾这些工具——很多在座的人就是这么干的——再叠加真正深的创造力,那种'更不可思议的东西'是可能被造出来的。"
"Switching gears to open weights — the recent release of Gemma. You're making highly capable open and accessible models that can actually run locally. What does that mean? Will AI be in the hands of users instead of primarily in the cloud? Does that change who gets to build with these models?"
"换个话题聊 open weights——你们最近放了 Gemma 出来,做出了能在本地跑的、开放且高能力的模型。这意味着什么?AI 是不是会从云端为主转向用户手里?这是不是会改变'谁能用这些模型 build 东西'?"
"We're huge proponents of open source and open science in general. AlphaFold — we put it all out there for free. All of our science work, even still today, we publish in the big journals. We wanted to create world-leading models for their sizes — that's hopefully what we've done with Gemma. We're very committed to that path."
"我们是 open source 和 open science 的大力支持者。AlphaFold 我们直接全开出去。我们所有 science 工作至今仍在大期刊发表。我们想做的是'同尺寸下世界领先'的模型——Gemma 希望就是这样。我们会持续走这条路。"
"It's been about 40 million downloads now in just two and a half weeks. We're really excited. I also think it's important for there to be Western stacks on open source. A lot of the Chinese models are excellent and currently lead in open source. Gemma is very competitive for its sizes."
"两周半的时间下载了大概 4000 万次。我们很激动。另外我觉得很重要的一点是:开源生态里需要有'西方栈'。很多中国模型非常出色,目前在开源里也确实领先。Gemma 在它的尺寸级里有竞争力。"
"For us — there's a question of resources, talent, and compute. Nobody has enough spare compute to make two frontier models at maximum size with different attributes. But for now, we've decided that our edge models — the things we want to use for Android and glasses and robotics — it's best they're open models, because they're vulnerable anyway once you put them out on the surfaces. So they might as well be actually fully open."
"对我们来说——资源、人才、算力都有约束。没人有那么多余裕的算力去做两个不同侧重的最大 frontier 模型。但现在我们的决定是:edge 模型——我们要在 Android、眼镜、机器人上用的那批——干脆就开放。反正部署到设备表面之后它本就 vulnerable,索性彻底开放。"
"So we've made a decision to unify that at the nano-size level. That works for us strategically too. We hope as many people as possible build on it — and we'll be building on that too."
"于是我们在 nano 这个尺寸级把这两件事统一起来——既是技术决定也是战略决定。希望尽可能多的人在它上面 build。我们自己也会 build 在它上面。"
"Earlier, I got to show you a demo of my version of Samantha from Her. It worked, which is amazing. Gemini was built multimodal. The depth of the context and tool use with speech directly to the model — there's nothing like it. Bar none, the best one actually."
"在之前我给你 demo 了我自己版本的 Samantha(《Her》里那个)。它居然 work 了。Gemini 一开始就是多模态。speech 直接进模型 + tool use + 上下文深度——没有别的能比。是同类里最强的。"
"That's the slightly underappreciated aspect of the Gemini series — we started multimodal from the start. That made it a bit more difficult initially, compared to just focusing on text, but we believe we'll gain from that in the long run."
"这是 Gemini 一直被略微低估的地方——我们一开始就是 multimodal。这在早期反而更难,比起只盯 text 来说;但我们相信长期会赢回来。"
"We're seeing that now for things like world model building, stuff like Genie that we build on top of Gemini. And it's gonna be really important for robotics — Gemini Robotics will be built on multimodal foundation models. We have a competitive advantage with Gemini being so strong at multimodal."
"现在我们已经在 world model(比如基于 Gemini 之上的 Genie)看到这个回报。机器人也会大用——Gemini Robotics 是搭在多模态 foundation model 之上的。Gemini 的多模态强项是我们的核心竞争力之一。"
"We're using it increasingly in things like Waymo. If you imagine devices and digital assistants that come with you into the real world — on your phone or glasses or some other device — it needs to understand the physical world around you, intuitive physics, the physical context you're in. Our systems are extremely good at that."
"在 Waymo 里我们也在越来越多地用它。再想想未来那种'跟你一起进入物理世界'的设备和数字助手——手机、眼镜、其他形态——它必须理解你周围的物理世界、直觉物理、你所处的物理上下文。我们的系统在这一点上极强。"
"The cost of inference is dropping fast. What becomes possible when inference is essentially free, and how does that change what your team is actually optimizing for?"
"推理成本在快速下降。当推理本质上免费时,什么会变得可能?这会改变你团队的优化目标吗?"
"I'm not sure inference will ever be essentially free. There's sort of Jevons paradox and other things. I think we'll just end up using whatever we can get our hands on."
"我不太相信推理会便宜到本质免费。Jevons 悖论之类的东西在这里同样适用。最后会发生的事是——我们会把所有能拿到的算力全用掉。"
"You could imagine millions of agents — swarms of agents — working together on things. That's one way to use the inference. Or single agents, or smaller groups of agents thinking in multiple directions and then ensembling that. We're experimenting with all these things, probably many of you are. All of that will use up any inference that's available."
"你可以想象数百万 agent——成群的 agent——协同工作。这是一种用法。或者单 agent、小群 agent 在多个方向上同时思考然后 ensemble 起来。这些我们都在做,在座很多人也是。所有这些用法都会把'可用算力'耗光。"
"One day maybe it can be almost cost-zero — certainly the energy if we solve fusion or superconductors or optimal batteries or some set of those things, which I think we will do with material science. Energy costs will be essentially zero. But there'll still be the physical creation of the chips. There'll be some bottleneck for at least the next few decades."
"也许某一天接近零成本——能源至少有可能,如果我们靠材料科学解决了聚变、超导体、最优电池里的某一项,我相信我们会做到。能源就会本质免费。但芯片的物理制造仍然在。这一关至少未来几十年都会是 bottleneck。"
"So they'll still be rationing on the inference side. You'll still have to use it efficiently."
"所以推理这一侧还是会被'rationing'(配额化)。你必须高效地用它。"
"We've got a lot of bio and biotech founders in the audience. AlphaFold 3 took us beyond proteins to a broad spectrum of biomolecules. How close are we to modeling full cellular systems, or is that still fundamentally harder?"
"听众里有不少 bio / biotech 创业者。AlphaFold 3 把范围从蛋白质扩到了更广的生物分子。我们离整细胞系统建模还有多远?这还是另一个量级的问题吗?"
"At Isomorphic Labs, which we spun out from DeepMind after AlphaFold 2 — it's going amazingly well. It's trying to build out, not just AlphaFold (which is one piece of drug discovery), but the adjacent biochemistry and chemistry to design the right compounds with the right properties. We'll have some big announcements very soon."
"我们 AlphaFold 2 之后从 DeepMind 拆出去的 Isomorphic Labs 进展非常好。它在建造的东西不只是 AlphaFold(那只是 drug discovery 的一块)——还包括相邻的 biochemistry / chemistry,目标是设计具备所需性质的合适化合物。我们最近会有大公布。"
"Eventually you want a whole virtual cell. I've talked about this — a full working simulation of a cell that you can perturb, and the outputs would be close enough to experimental that it's useful. You could skip a lot of the search steps and generate lots of synthetic data to train other models that predict things about real cells. I think we're about 10 years away from a full virtual cell."
"最终你要的是一整个 virtual cell——一个能扰动、输出和实验结果足够接近、因此有用的完整细胞模拟。它能省掉一大堆 search 步,还能生成大量合成数据去训练其它能预测真实细胞行为的模型。我估计距离完整 virtual cell 大概还有 10 年。"
"On the DeepMind side, we're starting with a virtual nucleus first, because it's relatively self-contained. The trick is — can you pick a slice of the complexity? Eventually you wanna model a human body, but at what level of detail? What slice will be self-contained enough that you can approximate the inputs and outputs?"
"DeepMind 这边我们从'virtual nucleus(细胞核)'开始——它相对自包含。所有这种问题的关键都是:你能不能切出一片合适的复杂度?最终你想模拟人体,但到哪个粒度?你能切出哪一块自包含得足够好、能把它的输入输出近似掉?"
"The other issue is — there's not enough data yet. If we could image a live cell without killing the cell, that would be game-changing, because then you could convert it into a vision problem, which we know how to solve. At the moment there are no techniques that give you nanometer resolution without destroying the cell."
"另一个 bottleneck 是数据不够。如果能在不杀死活细胞的前提下做纳米级成像,那 game changer——可以直接转换成 vision 问题,我们知道怎么解。目前没有这种'活细胞 + 纳米分辨率 + 不杀死它'的成像技术。"
"You've been looking at all kinds of science — material science, drug discovery, climate modeling, mathematics. If you had to rank which scientific domain will transform most dramatically in the next five years, what's on your list?"
"你看过非常多 science 领域——材料、药物、气候、数学。未来 5 年哪个领域会被最剧烈地改造?你怎么排?"
"They're all so exciting — that's been my main passion, the reason I've worked on AI for 30+ years: to use AI as the ultimate tool for science, to advance scientific understanding, scientific discovery, medicine, and our understanding of the universe."
"全都让我激动——这是我做 AI 三十多年的主要动力。AI 是科学的终极工具,用来推动科学理解、科学发现、医学,以及对宇宙的理解。"
"Our original mission statement had two steps: step one was solve intelligence — i.e. build AGI — and step two was use it to solve everything else. We had to change that over time because people were like, 'do you really mean solve everything else?' And we did mean that. People are starting to understand what that means today."
"我们最初的使命有两步:第一步,解决智能(也就是造 AGI);第二步,用它去解决其他一切。后来不得不改一改措辞,因为别人会问'你真的是说一切吗?'我们就是这个意思。今天大家开始懂了。"
"Specifically I meant: solve other 'root node' problems in science — areas of science that would unlock whole new branches or avenues of discovery. AlphaFold is the prototypical example. Over three million researchers around the world use AlphaFold now — pretty much every biology researcher. Some pharma executives told me almost every drug discovered from now on will have used AlphaFold at some point in the discovery process."
"我所说的'其他'特指科学里的 root node 问题——能解锁全新分支或方向的领域。AlphaFold 就是 prototype——全球已有超过 300 万研究人员在用,几乎覆盖所有生物研究者。一些 pharma 高管告诉我:从现在开始几乎每一款新药在 discovery 流程的某个环节都会用到 AlphaFold。"
"That's the impact we hope to have. I don't see any area of science or engineering that AI won't be helpful with. The ones you mentioned — we're at almost an AlphaFold 1.0 moment for those. Very promising results, but it hasn't quite solved the grand challenge yet. The next couple of years we'll have a lot to talk about across materials all the way to mathematics."
"这就是我们希望达成的影响。我看不到任何一块 science / engineering 是 AI 帮不上忙的。你提到的这些——目前大概都在'AlphaFold 1.0'阶段:结果非常 promising,但还没真正攻克 grand challenge。接下来几年我们会在材料一直到数学的整个范围里有很多消息要讲。"
"A lot of people in this room are trying to build companies applying AI to science. What's the difference between a startup that actually advances the frontier versus one that's just wrapping an API around a foundation model and calling it AI for science?"
"在场很多人想做 AI for science 的创业。什么样的创业是真正推动前沿,什么样的只是套个 API 然后管自己叫 'AI for science'?"
"What I would do today if I was sitting in your place: you have to intercept where the AI tech is going. That's one hard part. But I think there's huge scope for combining where AI's going with some other deep technology area — materials, medicine, other really hard areas of science."
"如果我今天坐在你这位置上,我会怎么做:你必须 intercept(抢在前面拦截)AI 技术的发展方向。这本身就是难的。然后我会把 AI 走向 + 另一个 deep tech 方向叠起来——材料、医药、那些真正硬的科学领域。"
"Those interdisciplinary teams, especially if it involves the world of atoms — there's not gonna be a shortcut to that, at least in the foreseeable future. Those areas are pretty safe from just getting swamped by whatever the next foundation model update is."
"跨学科的团队、尤其是涉及原子世界的——这种事情可见的未来里没有 shortcut。这些方向也相对安全,不会被下一次 foundation model 更新一波带走。"
"I've always loved deep tech. Nothing that's really long-lasting and worthwhile is easy. AI was like that back in 2010 when we started — it was thought to just not work. Investors and even academia considered it a niche subject we tried in the '90s and know doesn't work. But if you have belief and conviction in your idea, why it's different this time, what special combination from your background you have — ideally you're expert in both the machine learning side and the other area you're applying it to, or you create a founding team with that expertise — there's huge impact and value to be built."
"我一直偏爱 deep tech。任何真正长寿、值得做的事都不简单。2010 年我们开始时 AI 也是这个状态——大家认为它不 work,投资人这么觉得,学界也是,认为它是 90 年代搞过、已经知道不行的小众方向。但你只要对自己的 idea 有信念,知道这次为什么不一样、你的背景有什么特殊组合——理想情况是你既懂 ML,又懂你要应用到的那个领域;或者你能凑出这样的 founding team——这里的 impact 和价值是巨大的。"
"AlphaFold was a spike you pursued and it worked. What makes a scientific domain ripe for an AlphaFold-style breakthrough? Is there a pattern?"
"AlphaFold 是你押的一个 spike,结果做成了。什么样的科学领域适合下一个 AlphaFold-style 的 breakthrough?有规律吗?"
"The lesson I've learned from all the alpha projects we've done — specifically AlphaGo and AlphaFold — the techniques we have are great if the situation can be described as a massive combinatorial search space. The more massive, the better in some ways — so no brute force or special-case algorithm will solve it. That's true of Go moves and of different configurations of proteins (far more than atoms in the universe)."
"做完所有 alpha 系列后我总结出的规律——尤其是 AlphaGo 和 AlphaFold——是这样:如果你的问题能被描述成一个'巨大的组合搜索空间'——越大反而越好(因为这样 brute force 或专用算法都解不掉),那我们的方法就特别合适。Go 的合法走法、蛋白质的可能构象(数量远超宇宙原子数)都属于这种情况。"
"Then you have a clear objective function — minimizing free energy in proteins, or winning the game of Go. You need to specify your objective function clearly so you can hill climb."
"然后你要有清晰的目标函数——蛋白质里是最小化自由能,Go 里是赢棋。你得把目标函数明确写出来,这样系统才能 hill climb。"
"And then enough data and/or a simulator that can generate lots of in-distribution synthetic data. If those things are true, with today's methods you can go a long way into tackling and finding the kind of needle in the haystack that you need."
"再加上足够的数据,或者一个能生成大量 in-distribution 合成数据的模拟器。这三样齐了,今天的方法就能在大海捞针里走很远。"
"I think of drug discovery the same way. There is a compound out there that would solve this disease, if one could only find it — and that wouldn't have any side effects. As long as the laws of physics allow it, the only question is how to find it efficiently, in a tractable way. We showed for the first time with AlphaGo that these systems could find those kinds of needles in a haystack — in that case, the perfect Go move."
"药物发现也是同一个结构。'某种化合物能治这种病、还没副作用'——只要物理定律允许它存在,问题就只剩'怎么高效地、可计算地找到它'。AlphaGo 是第一次证明这种系统能找到大海里那根针——那次是完美的 Go 步。"
"How close are we to AI systems that can do genuine scientific reasoning, not just pattern matching on data?"
"AI 离能做真正科学推理(而不只是数据上的模式匹配)还有多远?"
"I think we're close. We have a system called Co-Scientist, and we have algorithms like AlphaEvolve that go a bit beyond what basic Gemini will do. All the frontier labs are experimenting in this way. I haven't seen yet anything that is a true, genuine, massive discovery — that's my personal opinion. I think it's coming."
"我认为我们已经很近。我们有个叫 Co-Scientist 的系统,还有 AlphaEvolve 这种算法,它们比基础 Gemini 多走一步。所有 frontier lab 都在这条线上做实验。但我个人觉得还没看到任何一个真正意义上的 massive discovery——这只是我个人观点。它会来。"
"It may be related to creativity — going beyond the bounds of what's known. That's just not pattern matching at that point, because there is no pattern to match to, and it's a bit more than extrapolation. It's some kind of analogical reasoning — and I don't think these systems have that yet, or at least we're not using them in the right way."
"这可能和创造力是同一件事——突破已知边界。这就不是模式匹配了——没有可匹配的模式;也不只是外推。它需要一种类比推理(analogical reasoning),而我不认为目前的系统具备,或者我们还没找到正确的用法。"
"In science I often say it this way: can it come up with a hypothesis that's really interesting, not just solve one? When I say 'just' — I'm talking about solving the Riemann hypothesis or one of the Millennium Prize problems. That would be obviously amazing — maybe we're a couple of years out. But I'd like to solve P equals NP, that's my favorite one."
"我经常这么对科学说:它能不能提出一个真正有意思的假设,而不只是解决一个?这里我说的'只是'是指——解决 Riemann hypothesis 或某个 Millennium Prize 问题。这本身就已经很 amazing 了,可能距离也就两年。我个人最希望解决的是 P = NP,这是我最喜欢的一题。"
"But even harder than that would be to come up with a new set of Millennium Prize problems that top mathematicians would regard as deep and meaningful and worthy of a lifetime of study. That's another level harder. We don't have that yet. I don't think it's magical though — these systems will eventually be able to do that. Maybe we're missing one or two things."
"比这还难的是——它能不能提出一组新的 Millennium Prize 级问题,被顶级数学家公认为深刻、值得花一辈子去研究?这是再上一个量级。我们离这个还远,但我不认为这是某种魔法——系统终归能做到,只是还差一两样东西。"
"The way we'd test that — I sometimes call it my Einstein test. Can you train a system with the knowledge cutoff of 1901, and will it then come up with what Einstein did in 1905, including special relativity, his annus mirabilis? We could run that test, and once a system passes, we're on the verge of these systems being able to invent something truly novel."
"测试方式——我有时候叫它 Einstein 测试。把模型的知识截断在 1901 年,看它能不能搞出 1905 年 Einstein 的那一波,包括狭义相对论(他的 annus mirabilis,奇迹之年)。这个测试我们可以跑——一旦哪天通过了,就说明这些系统离'发明全新的东西'只差一步。"
"Last question. For the deeply technical people in this room who want to work on something even close to the scale of what you've created — what's the thing you know now about building at the frontier that you wish you'd known at 25?"
"最后一个问题。在场有大量深技术的人想做接近你这个量级的事——你现在知道的、关于'在前沿做事'的某件事,你 25 岁时希望就懂的是什么?"
"Going after hard, deep problems is no more difficult in some ways than going after a shallower, simpler, more superficial problem. They're just differently difficult. Life's very short. You only have so much time and energy. You might as well put your life force into something that will really make a difference if you hadn't been there to push it."
"做深、做硬的问题,从某种意义上说和做浅、做简单的问题难度差不多——只是难的方向不同。人生很短,精力有限。你不如把生命力投到一件'如果不是你来推就不会发生'的事上。"
"I love interdisciplinary work — combinations of fields. That's gonna be even more prevalent in the next few years, finding the connections between fields. It's gonna be even easier to do that with AI."
"我热爱跨学科——领域和领域的组合。未来几年这种工作会更普遍——在不同领域之间找连接。有了 AI,做这件事会变得更容易。"
"The other thing — depending on what your AGI timeline is, mine's like 2030 or something — if you start off on a deep tech journey today, usually you're talking about a 10-year journey for true deep tech, in my opinion. So now you have to just consider AGI appearing in the middle of that journey. So what does that mean? It's not bad necessarily, but you have to take that into account."
"另外——取决于你对 AGI 的时间表(我自己估算是 2030 年附近)——如果你今天开始一个 deep tech 项目,真正的 deep tech 通常是 10 年起步。那你就必须把'AGI 在中点出现'这件事当成必发生事件来考虑。这意味着什么?不见得是坏事,但你必须 factor in。"
"One thing I see happening: Gemini, Claude, or one of these general systems making use of AlphaFold-like specialized systems as tools. I don't think we're gonna have it just in one giant brain — that will have too much regression. We don't need Gemini to do protein folding. So much better is to have really good general-purpose tool-usage models that orchestrate specialized tools."
"我看得到的一种走向:Gemini、Claude 这种通用系统会把 AlphaFold 这类专用系统当作 tool 来调用。我不认为它会全部塞进一个 giant brain ——那会有大量信息密度上的回退。我们不需要让 Gemini 自己去做蛋白质折叠。更合理的形态是:非常强的通用 tool-using 模型 + 一堆专用工具被它编排。"
"So I think that's interesting to think through the implications of. What you might build today — also physical things, what kinds of factories would you build, finance systems and so on. You really need to take that seriously, imagine what that world would look like, and then build something that would be useful if AGI comes in halfway through."
"想清楚这些含义很关键。你今天要建什么——包括物理的东西,你会建什么样的工厂、什么样的金融系统等等。把这件事认真当回事,先想象那个世界长什么样,再去 build 一个'AGI 在你做到一半时进场后,反而能变得更有用'的东西。"