20VC：DeepMind的德米斯·哈萨比斯谈为什么AGI比工业革命更重要 | 为什么大型语言模型不会商品化以及我们尚未触及扩展定律 | 人工智能的瓶颈与人工智能引发的能源危机 | 人工智能会加剧还是缓解不平等

尽管回报相比最初大规模扩张时有所下降，但依然非常可观。

The returns are still very substantial, although they're a bit less than they were, obviously, at the start of all of their scaling.

我认为，支撑现代人工智能产业的突破性成果，有大约90%是由谷歌大脑、谷歌研究或DeepMind完成的。

I would say about 90% of the breakthroughs that underpin the modern AI industry were done by either by Google Brain or Google Research or DeepMind.

那些有能力发明新算法理念的实验室，在未来几年将开始获得更大的优势。

Those labs that have capability to, you know, invent new algorithmic ideas are gonna start having bigger advantage over the next few years.

最后一波想法，差不多已经把它们的潜力榨干了。

The last set of ideas are sort of, you know, all the juices being rung out of them.

我有时会量化一下，比如通用人工智能（AGI）。

And I sometimes quantify, like, AGI.

通用人工智能的到来，相当于工业革命的十倍规模，但速度是它的十倍。

The coming of AGI is, like, 10 times the Industrial Revolution at 10 times the speed.

这里是公元前20年，我是哈里·斯蒂宾斯，非常期待今天的节目。

This is twenty BC with me, Harry Stebbings, and I'm so excited for the show's day.

我走路来参加这次采访，向我妈妈描述时是这么说的。

I walked to this interview, and I described it to my mother like this.

我们的节目嘉宾都非常出色，但老实说，很少有人能被视作与牛顿、图灵、爱因斯坦同属一个级别。

We have amazing guests on the show, but very few honestly will be considered in the same realm as Newton, Turing, Einstein.

今天的嘉宾是地球上最杰出的头脑之一，我为自己有机会与他坐下来讨论今天的内容感到无比幸运。

Our guest today is one of the greatest minds on the planet, and I consider myself incredibly lucky to have had the chance to sit down with him and discuss what we did today.

这确实是一期特别的节目，我会铭记很久。

This is a truly special one and one that I will remember for a very long time.

希望你们喜欢这期节目，我也非常感激能与这位非凡的人物共度这段时光。

Enjoy the episode, and I so appreciate the time we had with a very special human being.

我非常高兴欢迎来自DeepMind的德米斯·哈萨比斯。

I'm thrilled to welcome Demis Hassabis at DeepMind.

但在我们开始今天的节目之前，你知道商务旅行的平均预订时间是45分钟吗？

But before we dive into the show today, did you know the industry average for booking a business trip is forty five minutes?

这严重浪费了团队的时间。

That's a massive waste of your team's time.

而使用Novan，你的员工平均只需7分钟就能完成旅行预订。

Well, with Novan, your employees can book a trip in just seven on average.

Novan 是一款专为注重效率的企业设计的AI驱动差旅与费用管理平台。

Novan is the AI powered travel and expense platform designed for companies that value efficiency.

它通过高员工使用率和自动化政策控制，带来切实的业务影响。

It drives real business impact through high employee adoption and automated policy control.

现在，内置的AI会自动批准符合政策的预订，并自动阻止不符合的预订。

Now the built in AI approves in policy bookings and blocks the rest automatically.

这使得财务团队不再需要追要收据，也无需再经历月末的混乱，同时您还能获得实时可见性，帮助公司节省高达15%的差旅预算。

This allows finance teams to stop chasing receipts and skip the month end chaos, and you get this real time visibility that can save your company up to 15% on your travel budget.

这就是为什么Visa、Stripe、Figma，甚至Anthropic等企业如今都依赖它。

And that's why leaders like Visa, Stripe, Figma, and even Anthropic rely on these days.

立即访问 navan.com/20vc 亲自体验，您还有机会赢取两张美国大陆境内任意目的地的商务舱机票。

Go to navan.com/20vc today to see for yourself, and you'll get a chance to win two business class flights anywhere in Continental US.

无需购买即可参与。

No purchase necessary.

规则适用。

Rules apply.

立即前往 navan.com/20vc。

Head over to navan.com/20vc now.

Novan 简化差旅，Airwallet 简化其背后的支出。

Novan simplifies the travel, Airwallet simplifies the spend behind it.

创始人们，让我们现实地谈谈增长成本。

Founders, let's get real about the growth tax.

你已经获得了风险投资，并正在全球扩张，现在的问题不再是推出产品。

You've raised VC funding and you're scaling globally, and it's no longer about shipping product.

而是要在各大洲之间协调运营。

It's about orchestrating operations across continents.

但突然间，你的支付和财务系统正在拖累你的增长。

But suddenly, your payments and finance stack is choking your growth.

你不得不登录多个银行门户，等待数日才能完成转账，并跨实体进行报表汇总。

You're logging into lots of different banking portals, waiting days for transfers, and reporting across entities.

这是运营负担，而且在你这种规模下尤为严重。

It's operational drag and it's at your scale.

这正在耗费数百万美元。

It's costing millions.

这就是我为什么如此兴奋地与Air Wallets合作。

That's why I'm so excited to partner with Air Wallets.

Air Wallets不仅仅是HSBC或花旗银行的银行替代品。

Air Wallets are more than just a banking alternative to HSBC or Citi.

Airwallis为你提供了一个智能的财务操作系统，赋能全球企业的运营与增长，让你能够管理和自动化银行、资金管理、支付和支出。

Airwallis brings you an intelligent financial operating system that powers how global businesses operate and grow, allowing you to manage and automate banking, treasury, payments, and spend.

对我来说，最令人兴奋的是，他们正在大力投资代理金融。

The most exciting part for me, they're heavily investing in agentic finance.

如果你正在全球扩展，你需要一个无国界、实时且智能的银行与金融平台。

If you're scaling globally, you need a banking and finance platform that's borderless, real time, and intelligent.

今天就访问Airwallix，看看他们如何帮助Canva、迈凯伦等数千家企业实现规模化增长，网址是airwallix.com/forward/20vc。

Check out Airwallix today and see how they're helping thousands of businesses like Canva, McLaren, and deal scale at airwallix.com forward /20vc.

条款和条件适用。

Terms and conditions apply.

您的资金受到保护，但不受FSCS保障。

Your money is safeguarded, not FSCS protected.

了解更多详情，请访问airwallis.com。

See airwallis.com for more details.

虽然Airwallis帮助您的资金实现全球流转，Vanta则助您跟上安全步伐。

While Airwallis helps your money move globally, Vanta helps your security keep up.

安全与合规做不好，会带来巨大的麻烦。

Security and compliance done wrong is a giant headache.

但如果安全与合规做对了呢？

Security and compliance done right though?

那就是Vanta。

Well, that's Vanta.

Vanta帮助您赢得信任并加速增长。

Vanta helps you earn trust and speed up growth.

无需使用电子表格。

No spreadsheets required.

对于时间与资源有限的初创公司，Vanta 通过人工智能与自动化成为你的首个安全 hires，帮助你快速实现合规并促成重大交易。

For start ups low on time and resources, Vanta becomes your first security hire using AI and automation to get you compliant fast and unblock really big deals.

而对于大型企业，Vanta 是你的 AI 驱动型合规与风险中枢，整合企业各处的数据并自动化工作流程，让你随时都能证明自身的可信度。

And if you're big enterprises, Vanta is your AI powered hub for compliance and risk, bringing together data from across your business and automating workflows so you can prove trust at any moment.

Vanta 能伴随你在每个阶段共同成长。

Vanta scales with you at every stage.

这就是为什么从 Cursor 这样的初创公司到 Snowflake 这样的大型企业，顶尖公司都选择 Vanta。

That's why top companies from startups like Cursor to enterprises like Snowflake choose Vanta.

正确地做好安全与合规工作。

Do security and compliance right.

我的听众可以通过访问 vanta.com/20vc 获得 1000 美元的折扣。

My listeners can get $1,000 off Vanta by going to vanta.com/20vc.

前往 vanta.com/20vc，即可享受 Vanta 1000 美元的优惠。

That's vanta.com/20vc for $1,000 off Vanta.

您已到达目的地。

You have now arrived at your destination.

德米斯，我非常兴奋能参与这次对话。

Demis, I'm so excited to be doing this.

非常感谢你今天抽出时间来参加。

Thank you so much for joining me today.

很高兴能来到这里。

Great to be here.

我们可以从很多地方开始，但我其实看了你做的那部纪录片，非常棒。

Now, there are many places that we could start, but I was watching actually the documentary that you did, which was fantastic.

我其实想从AGI开始聊。

And I actually wanted to start on AGI.

关于AGI的定义有很多种。

Definitions are very varying.

你对它对你而言意味着什么，思考得非常深入。

You've been very thoughtful about what it means to you.

所以我想从这里开始。

And so I wanted to start.

你能向我解释一下，你今天是如何看待AGI的吗？这样我们可以以此作为基础。

Can you explain to me how you think about it today so we get that as a kind of ground center?

是的。

Yeah.

我们对AGI的定义一直很一致，基本上是指一个具备人类心智所有认知能力的系统。

We've been very consistent how we define AGI as basically a system that exhibits all the cognitive capabilities the human mind has.

这一点很重要，因为大脑是我们所知的、可能是宇宙中唯一证明通用智能是可能的实例。

And that's important because the brain is the only existence proof we have that we know of in maybe in the universe that general intelligence is possible.

因此，对我来说，这就是AGI应该达到的标准。

So that for me is the bar for what AGI should be.

这是个最糟糕的问题。

It's the worst question.

我们离目标有多近？

How close are we?

每个人的说法都不一样。

Every everyone says different things.

当你看到一些非常知名的人物说AGI可能早在2026年或2027年就会出现时，这确实非常困难。

And it's very difficult when you have, you know, very prominent figures saying it could be as early as 2026, 2027.

是的。

Yeah.

我的意思是，我对时间点有一个概率分布，但我认为未来五年内实现的可能性非常大。

I mean, I think, look, I've got a probability distribution around the timings, but I I would say there's a very good chance of it being within the next five years.

所以这根本不算久。

So that's not long at all.

这比你之前想的更近了吗？

Is that closer than you thought?

这个看法随着时间改变了吗？

Has that changed over time?

没有太大的变化。

Not really.

我的意思是，实际上，当你去想的时候，挺有意思的。

I mean, actually, you when you, it's funny.

我的联合创始人谢恩·莱格，这里是首席科学家，我们在2010年创立DeepMind时，他经常写博客文章预测AGI何时会出现。

My cofounder, Shane Legg, who's chief scientist here, when we started out DeepMind back in 2010, he used to write blog posts sort of predicting about when AGI would happen.

但请记住，2010年我们起步时，几乎没人从事人工智能，大家都觉得人工智能——尽管非常尊重地讲——

And bearing in mind, in 2010, when we started, almost nobody was working in AI, and everyone thought AI And with the greatest

没人会读这些博客文章。

of respect, no one was reading the blog posts.

是的。

No.

但这些文章现在仍然在网上，供人们查阅。

And but they're still there on the Internet for people to check.

我们当时经常对算力和算法进展进行外推。

And we used to do this extrapolation of compute and algorithmic progress.

基本上，我们预测从我们起步开始需要大约二十年时间。

And basically, we predicted around twenty years it would take from when we started out.

我认为我们目前的进度基本符合预期。

And I think we're pretty much on track.

今天在纪录片中，你认为最大的瓶颈是什么？

What are the biggest bottlenecks when you look today in the documentary?

你说你永远都不够算力。

You said you just never have enough compute.

嗯。

Mhmm.

当你审视我们今天所处的位置时，最大的瓶颈是什么？

What are the biggest bottlenecks when you look at where we are today?

我认为算力是最大的瓶颈，不仅因为显而易见的原因——根据所谓的扩展定律，不断扩展你的想法和系统，构建更大、参数更多的架构。

I think compute is the big one, not just for the obvious reason of scaling up your ideas and your systems as the scaling laws, as they're called, keeping on building bigger and bigger architectures with more and more parameters.

当你这样做时，你会得到更智能的系统。

And as you do that, you get more intelligent systems.

但你需要大量算力的另一个原因是做实验。

But the other thing you need a lot of compute for is for doing experiments.

计算机和云平台基本上就是我们的工作台。

The computers, the cloud is our workbench, basically.

所以如果你有一个新想法、一个新的算法思路，你想测试它，你就得在合理的规模上进行测试。

So if you have a new idea, a new algorithmic idea, but you want to test it, you kind of got to test it at a reasonable scale.

否则，当你真正把它应用到主系统中时，它就站不住脚了。

Otherwise, it won't hold when you actually put it into the main system.

因此，如果你有很多研究人员带着大量新想法，你就需要相当多的计算资源。

So you need quite a lot of compute if you have a lot of researchers with lots of new ideas.

你提到了‘扩展定律’这个词。

You mentioned the word scaling laws.

嗯。

Mhmm.

很多人认为我们正在触及扩展定律的极限，已经开始看到这种平台期效应。

A lot of people suggest that we're hitting scaling laws, and we're starting to see that plateauing effect.

是的。

Yeah.

你认为这是真的吗？

Do you think that's true?

我不这么认为。

I don't think so.

我觉得这比那要复杂一些。

I think it's a bit more nuanced than that.

所以，当领先的公司都开始构建这些大型语言模型时，每一代新系统都会带来巨大的性能提升。

So, of course, when the leading companies all started building these large language models, you're getting enormous jumps with each generation of new system.

你知道，它们的性能几乎翻倍了。

You know, maybe they're almost, like, doubling in performance.

在某个时候，这种增长必须放缓。

At some point, that had to slow down.

所以它并不是持续呈指数增长，但这并不意味着进一步扩大现有系统的规模就没有巨大的回报了。

So it's not kind of continuing to be exponential, but that doesn't mean there isn't great returns still for scaling the existing, you know, systems up further.

因此，我们和其他前沿实验室在计算能力扩展方面获得了丰厚的回报。

So we and the other frontier labs are getting a lot of great returns on on that kind of compute expansion.

所以我会说，回报仍然非常可观，尽管显然比最初大规模扩展时要低一些。

So I would say the returns are kind of still very substantial, although they're a bit less than they were, obviously, at the start of all of their scaling.

我们在哪些方面落后于你的预期？

Where are we behind where you thought we'd be?

我认为，实际上在大多数领域，我们已经超过了我原先的预期。

I think, actually, in most areas, we are ahead of where I thought we would be.

如果你想想视频模型，或者我们现在最新的系统如Genie，它们是交互式世界模型，如果你退一步来看，我觉得这简直不可思议。

If you think about things like the video models or even now with our newest systems like Genie, they're interactive world models, which I think is kind of incredible if you sort of step back and think about it.

如果五年前或十年前你给我看这些，我会非常惊讶。

I think if you'd shown me that five, ten years ago, I would have been pretty amazed.

所以我认为，在大多数领域，我们已经领先于整个领域原先的预期。

So I think in most domains, we're we are ahead of where the field thought.

不过，仍然有一些重要的东西缺失，比如持续学习。

There's still some big things missing, though, like continual learning.

这些系统在训练完成后，一旦部署到现实世界中，就不再学习了。

These systems don't learn after you finish training them, after you put them out into the into the world.

你知道，它们不太擅长学习新的东西。

You know, they're not very good at learning further things.

我认为还有一些关键能力尚未实现。

And I think some critical capabilities are about missing.

不好意思，我问的问题可能有点直接和基础。

I'm sorry to ask blunt and basic questions.

为什么我们还没有实现持续学习？

Why do we not have continuous learning?

嗯，人们还没有完全搞清楚，所有领先的实验室都在研究如何将新的学习能力整合到那些你花了数月时间训练的现有系统中。

Well, people haven't quite figured out yet, and all the leading labs are working on this, like, how to integrate new learning into the existing systems that, you know, you spent months training.

当然，大脑在这方面做得非常优雅。

Of course, the brain does this very elegantly.

对吧？

Right?

而且很可能是通过睡眠和强化学习等方式实现的。

And probably through things like sleep reinforcement learning.

所以，你知道，大脑中会有一种叫做‘巩固’的过程，即你白天的记忆会被重新激活，然后其中一部分信息会被优雅地融入你现有的知识体系中。

So, you know, you just kind of get consolidation, it's called, in the brain where, you know, your memories during the day are replayed, and then some of that information is elegantly incorporated into your existing knowledge base.

也许我思考了一段时间，我们需要类似的东西，将新信息与现有知识库结合起来。

And perhaps we I've I thought for a while, maybe we need something like that to incorporate new information along with the existing information base.

你提到了视频模型。

You mentioned video models.

你提到了媒体和图像。

You mentioned kind of media and image.

似乎DeepMind进展迅速，已经赶超甚至超越了其他厂商。

It seems that DeepMind has progressed very quickly and caught up slash overtaken other providers.

我想我发过推文，你好像点赞了。

I think I've tweeted I think you liked it.

但我基本上是分享了我使用过的东西，以及它如何随时间变化。

But I basically tweeted what I used and how it's changed over time.

现在，DeepMind是我研究新内容的首选。

And and DeepMind now is my number one for research for new shows.

是的。

Yeah.

以前并不是这样的。

It wasn't that way before.

是什么导致了DeepMind的加速发展，而这种发展在两三年前可能并不存在？

What has led to the acceleration and progression of DeepMind in a way that it wasn't maybe there two to three years ago?

是的。

Yeah.

嗯，我们进行了一些组织调整。

Well, we made some organizational changes.

我认为，我们一直拥有谷歌和DeepMind最深入、最广泛的科研团队。

So I think we've always had the deepest and broadest research bench at Google and at DeepMind.

我的意思是，如果你回顾过去十年甚至更久，大约十五年，我会说现代人工智能产业的突破性成果中，有90%都是由谷歌大脑、谷歌研究或DeepMind完成的。

I mean, if you look at the last decade or plus, you know, fifteen years, I would say about 90% of the breakthroughs that underpin the modern AI industry were done by either by Google Brain or Google Research or DeepMind.

所以，我们的一个团队，比如AlphaGo、强化学习，当然还有Transformer，这些都是关键性的突破。

So one of our groups, if you think of, like, AlphaGo and reinforcement learning and, of course, transformers, you know, these are all the key breakthroughs.

因此，如果未来还有任何缺失的突破，我相信我们有能力实现它们。

So I would back us to sort of make those breakthroughs in the future if there are any missing ones.

我认为我们基本上把公司内部所有人才汇聚在一起，朝着同一个方向努力。

And I think we've basically helped put together all the talent from around the company sort of pushing in one direction.

我们之前也谈到了计算资源的问题。

And then we talked earlier just about compute resources.

这还涉及到将我们所有的资源整合在一起，以便构建最大的模型，而不是让公司内部存在两三个不同的版本。

It was also about combining all of our resources together so we could build the biggest models rather than having two or three versions around the the company.

所以我认为，很大程度上是把我们已有的所有要素整合起来，然后以 relentless 的专注和节奏推进，像初创公司一样行动，重新回到前沿，并在许多领域保持领先。

So I think a lot of it was assembling together all the ingredients we already had and then kind of pushing with relentless sort of focus and and and pace, acting almost like a start up, really, to get back to the the frontier and and be ahead in in many areas.

你说如果有人能实现突破，那应该是

You say if anyone's gonna do the breakthrough, it could and should be

我们。

us.

是的。

Yes.

当你想到这一点时，持续学习是你最期待的下一个突破吗？

When you think about that, is continuous learning the next breakthrough that you're most excited by?

我觉得还有很多方面是缺失的。

I think there's quite a few things that are missing.

持续学习就是一个方面。

There's there's continual learning.

我认为在不同内存系统上还有很多潜力可挖。

I think there's a lot of mileage in looking at different memory systems.

目前，我们使用的是这些长上下文窗口，有点像是蛮力做法。

At the moment, we have these long context windows, which are kind of a bit brute force.

你只是把所有东西都塞进去。

You just put everything in them.

我认为那里还有很多有趣、很可能全新的架构等待被发明。

I think there's there's there's a lot of interesting, probably, architectures to be invented there.

还有像长期规划、分层规划这样的内容。

And then there's stuff like long term planning, hierarchical planning.

这些系统在长期规划上并不擅长，比如预测多年后的未来，而我们人类的大脑却可以做到。

These systems are not very good at planning at long time horizons, many years into the future, which we, as with our minds, we can do.

我认为还有很多问题有待解决。

There's quite a lot of problems, I think, that are still left to overcome.

也许其中最大的一个是一致性。

Maybe one of the biggest is consistency.

所以我有时把这些系统称为参差不齐的智能，因为当你以某种方式提问时，它们在某些方面确实非常出色。

So, you know, the I sometimes call these systems jagged intelligences because they're really amazing at certain things when you pose the question in a certain way.

但如果你以稍微不同的方式提问，它们仍然可能在一些非常基础的事情上失败。

But in if you pose a question in a slightly different way, they can actually still fail at quite elementary things.

所以，通用智能不应该是这种参差不齐的。

So a general intelligence shouldn't be that sort of jagged.

当你重新组织文件并设置代理以特定方式运行时

When you reposition files and you set up agents to perform in certain ways

是的。

Yeah.

然后文件无法再正确配置，整个系统就会彻底崩溃。

And then the files no longer configure, it completely falls over.

没错。

Exactly.

百分之百。

A 100%.

这是一场灾难。

That's a disaster.

是的。

Yeah.

我的意思是，通用智能，如果你想想我们的思维是如何运作的，它不应该存在这样的漏洞。

Well, I mean, the general intelligence, you know, if you think about how our minds work, it shouldn't have those kinds of holes in it.

我们之前提到过规模扩展的停滞。

We said about a plateauing of scaling.

每个人都谈论模型在能力上的商品化。

Everyone talks about a commoditization of models in terms of capabilities.

你觉得我们会看到这种现象，还是认为有一两个会持续领先于其他模型？

Do you think we see that, or do you think we see one to two continuously accelerate ahead of the others?

是的。

Yeah.

我觉得，也许现在那三四个领先的实验室——我们就是其中之一——差距开始拉大了。

I feel like maybe, you know, the three or four leading labs now, which we're one, I think the gap is sort of starting to pull away.

因为很多这些工具当然也能帮助你构建下一代模型。

Because a lot of these tools also, of course, help you build the next generation.

比如编码工具、数学工具。

So things like coding tools, math tools.

我认为，仅靠重复使用相同的想法来获得同等进展，正变得越来越难。

And it's getting harder and harder, I would say, to kind of eke out the same gains from just the same ideas.

所以我认为，那些有能力发明新算法思路的实验室，在未来几年将获得更大优势，因为上一代的想法已经差不多被榨干了。

So I think those labs that have capability to, you know, invent new algorithmic ideas are gonna start having bigger advantage over the next few years as the the last set of ideas are sort of, you know, all the juices being rung out of them.

多年来，你对很多研究都持开放态度，我们也看到了许多高质量的开源模型。

You were very open with a lot of your research for years, and we see many very good quality open models.

你如何看待开源的未来？

How do you think about the future of open?

我有许多投资组合公司使用前沿模型作为基准，然后用开源模型尽可能接近其性能，同时更具成本效益。

I have many portfolio companies that kind of use frontier models, and then they use that to set a benchmark, and then they use open models to kinda get as close as possible but with more cost effectiveness.

这种未来会是什么样子？

What does that future look like?

是的。

Yeah.

我认为这可能和我们今天看到的情况类似。

I think it's probably similar to what we're seeing today.

我的意思是，我们非常支持开放科学和开源模型，我们已经做了很多工作，从最初的Transformer到AlphaFold都是如此。

I mean, we're we're big supporters of of Open Science and and Open Models, and we've done many, many things, obviously, from from the original transformers to AlphaFold.

你知道，这些我们都已经向外界开放，以帮助研究社区。

You know, these are all things we've sort of given out into the world and to help the research community.

我们计划继续这样做，特别是在应用领域，比如将人工智能应用于科学——这显然是我的热情所在。

And we plan to continue to do that, especially in applied domains, you know, scientific domains, applying AI to science, which is obviously my passion.

但越来越多的情况是，开源模型可能会比最前沿的模型落后一步。

But increasingly, what you're going to see is the open source models are probably one step back from the absolute frontier.

开源社区通常需要大约六个月的时间来重新实现并理解这些想法。

It usually takes about six months for the open source community to sort of reimplement and figure out what those ideas are.

但我们也在大力推动一套名为Gemma的开源模型，我们决心让它们在各自规模下成为最佳选择。

But we are also pushing hard on a kind of suite of open source models called Gemma, which are we're determined to kind of make best in class for their sizes.

因此，对于小型开发者、学者或初创公司的初期阶段来说，我认为它们非常合适，同时也适用于边缘计算。

So specifically for small developers or academics or the, you know, the beginnings of a startup, I think they're perfect for that and also edge computing too.

因此，我们对开源模型在某些类型的应用中非常感兴趣。

So we're very interested in open source models for certain types of applications.

你如何看待大语言模型之后的世界？

How do you think about a world post LLMs?

不同的人有不同的看法。

You have different people with different views.

你的看法非常不同。

You have with very different views.

对我来说，我不认为像Jan所说的那样，我在一些方面持不同意见，我认为可能还有一些我们尚未发现的关键要素，需要突破才能取得进展。

For me, I don't think it's you know, I kinda disagree with Jan on a few things in terms of I think there might be there's there's a fifty fifty chance there's some things maybe missing that we still need to make breakthroughs in.

也许它们是世界模型，这类方法。

Perhaps they're world models, these kinds of approaches.

但我坚信，我们已经看到了这些基础模型有多么成功。

But my betting is pretty strongly is we've seen how successful these foundation models have been.

它们能够完成令人惊叹的事情。

They can do incredibly impressive things.

我认为这种趋势不会消失。

I don't think that's going to go away.

我们仍在从规模定律中获得收益。

We're still seeing gains from the from returns from the scaling laws.

我认为，当思考未来的通用人工智能系统时，唯一的问题是：LLM基础模型是唯一的關鍵组件，还是整个系统？

I think the only question really is when you think about a future AGI system is is an LLM foundation model gonna be the key component only, or is it the total system?

对吧？

Right?

所以我只是觉得，关键在于是否还需要其他东西？

So I just think it's it's a question of, you know, is there anything else needed?

不是，我不认为它会被取代。

Not is it not I don't think it's gonna get replaced.

我认为这些基础模型会被进一步构建，就像我们对待世界模型那样。

I think it's gonna get built on top of these foundation models, just like the way we do with our world models.

当我们展望未来五年，正如你所说，可能具备通用人工智能的世界会是什么样子？

When we think about that future, five years out, as you said, potentially with AGI, what does that world look like?

许多人有不同的担忧。

Many people have different concerns.

如果我们从一般角度出发，那个世界会是什么样子

If we just start generally, what does that world look

在你看来？

like to you?

我认为积极的一面是，显然，我整个职业生涯和人生都在为通用人工智能而努力，我认为它将成为科学和医学的终极工具。

I think on the positive side and the things, obviously, I've I spent my whole career and life building towards AGI is I think it will be the ultimate tool for science and medicine.

在推动科学发现、寻找疾病治愈方法方面，我认为我们需要这种技术。

So in terms of advancing scientific discovery, finding cures to diseases, I think we need that kind of technology.

因此，我希望在五年或更长时间后，我们将进入一个科学发现的新黄金时代。

And so I'm hoping, in five years plus time, we'll be sort of entering a new golden era, golden age of scientific discovery.

我母亲患有多种硬化症，所以这是我最期待的领域。

So my mother's got multiple sclerosis, so it's, like, something it's the thing that I'm always most excited about.

是的。

Yeah.

我担心的是药物研发的过程，以及通过所有临床试验所需的时间。

The thing I worry about is actually kind of drug discovery, the process of getting it through all the trials

嗯。

Mhmm.

而我知道，要等到我母亲真正受益，可能需要十年时间。

And knowing that it takes a decade before my mother will actually get any benefits from it.

我们该如何解决这个问题？

How do we solve that?

对。

Yeah.

所以我认为我们很快就能达到那个阶段。

So I think we'll get to that point soon.

首先，我们做了AlphaFold项目来预测蛋白质折叠，之后孵化了一家名为Isomorphic Labs的公司，它发展得非常好。

First of all, what we're doing is you know, after we did the AlphaFold project to do protein folding, then we spun out a company called Isomorphic Labs, which is doing extremely well.

那家公司的目标是专注于解决药物研发过程中的其他环节，这涉及大量化学工作，比如设计化合物、检查其毒性，以及确保药物具备安全所需的各种特性。

And that is supposed to you know, the idea there is we're focusing on solving the rest of the drug discovery process, which is a lot of chemistry, designing the compounds, checking it's not toxic, and all the different properties you need for drugs to be safe.

我认为，未来五到十年内，我们就能完成整个药物设计引擎的构建。

I think we'll have that whole drug design engine ready in, you know, the next five plus five to ten years.

你说得对。

Then you're right.

下一个问题是临床试验仍然需要很多年。

The next problem is the clinical trials still take many, many years.

对吧？

Right?

但我认为人工智能可以帮助解决这个问题，比如模拟人体部分代谢过程，以及对患者进行分层，确保特定患者获得与其基因组特征完全匹配的药物。

But I think AI can help there in terms of simulating parts of the human metabolism, also stratifying patients to make sure that certain patients get exactly the right type of drug that's suitable for their genomic makeup.

所以我认为AI在这方面也能提供帮助。

And so I think AI can help there too.

但我认为真正的革命将发生在少数——可能十几种AI药物成功通过整个流程之后。

But I think the real revolution will come when a few, maybe a dozen or so, AI drugs get through the whole process.

然后政府和监管机构看到这些成果，拥有足够的数据来验证这些模型的预测。

And then the government and the regulatory body see that, and they have enough data to sort of back test the predictions of those models.

在未来再过十年左右，我们或许能够完全信任模型的预测，从而跳过某些步骤，比如不再需要动物实验。

And then maybe what we can do will be in the future where maybe ten further years where we can really just trust the predictions that the models are making and actually then maybe skip out some steps, perhaps like the animal testing is not needed anymore.

由于我们可以依赖这些模型，或许可以更快地提高剂量水平。

Maybe we can go up the dosage ladder quicker because you can rely on these models.

所以我认为我们需要分两步走。

So I think we gotta do it in two steps.

首先解决药物设计问题，然后再关注监管所需的时间。

Solve the drug design problem first, and then look at the regulatory length of time it takes.

说到监管，AI安全是一个重要且令人关注的话题。

Speaking of regulatory, AI safety is a big topic and a big concern.

我认为确实是这样，而且昨晚我吃饭时看了这部纪录片，非常精彩。

I think it was and, again, I I watched it last night over dinner, which was a great watch, which is obviously the documentary.

我认为是斯蒂芬·霍金说过，我们必须做好，因为我们可能没有第二次机会。

I think it was Stephen Hawking who said we must get it right because we might not get another chance.

你觉得他说得对吗？

Do you think that's right?

是的。

Yeah.

我也觉得他说得对。

I do think that's right.

我认为这正是我们需要面对的严峻现实。

I think that is the stakes that, you know, we have to deal with.

而且，我担心两件事。

And, you know, there's two things I worry about.

一是这些系统被不良行为者滥用，它们可能被重新利用。

One is the misuse of these systems by bad actors, and they can be repurposed.

这些是双重用途的技术。

These are dual purpose technologies.

它们可以用于科学和医疗领域的巨大福祉，正如我们刚才讨论的，但也可能被不良行为者重新用于有害目的。

They can be used for incredible good in science and health, as we just discussed, but they can also be repurposed for harmful ends by a bad actor.

所以这是一个问题。

So that's one issue.

第二个问题是技术性的：确保这些系统在变得更强大时——不是今天的系统，而是一两年后，当它们变得更具能动性、更自主，接近通用人工智能时——仍能被控制在我们希望的框架之内。

Second issue is a technical one, making sure these systems as they get more powerful, not today's systems, but maybe in a year or two's time when they become more agentic, more autonomous as we get towards AGI, can they be kept on the guardrails that we want?

我认为，适当的监管可以帮助确保所有领先提供商至少达到某种最低标准。

And I think regulation the right kind of regulation could help here in terms of making sure there's at least sort of minimum standards from all of the leading providers.

但理想情况下，这需要一种国际标准。

But it needs to ideally be a kind of international standards.

什么样的监管才是合适的？

What is the right kind of regulation?

而且，我再次引用你刚才的话。

And, again, I'm kind of quoting yourself back from this.

你提到你觉得我们需要更多的全球协调，这让我担心，因为我们在这方面正变得越来越差

You mentioned you were like, I think we need more global coordination, which worries me because we're getting worse at it

是的

Yes.

我认为这无疑是一个真实的状况。

Which I think would be an unwavering truth.

对

Yes.

当然

For sure.

我的意思是，我们现在的时机真不巧，正值人类历史上可能最具影响力的技术出现之际，而国际体系却愈发碎片化。

I mean, that's not it's sort of crazy, the timing that we're in, right, with this most consequential maybe technology the world's ever seen at the same time as a very fragmented sort of international system.

这并不理想，但我认为我们只能尽力而为，至少制定一套最低标准或一些检测不良特性的基准，比如欺骗行为。

And it's not ideal, but I think we're gonna have to try and do the best we can to at least come up with a sort of set of maybe minimum standards, some benchmarks that test for undesirable properties, for example, deception.

没人应该开发具备欺骗能力的系统，因为那样它们就可能绕过其他安全防护措施。

You don't you know, nobody wants should be building systems that are capable of deception because then they could be getting around other safeguards.

然后我想象一下，如果一切顺利，会有一种认证流程，几乎就像一个质量标志，表明这个模型具备某些安全措施和保障。

And then I imagine, you know, if things go well, some kind of certification process that basically it's almost like a kite mark of, you know, quality that this model has certain safeguards and certain guarantees.

因此，消费者和公司可以安全地在此基础上进行构建。

And so therefore, consumers and companies can safely sort of build on top of it.

我认为这理想情况下就应该是这样发展的。

And I think that is how it should go ideally.

但这必须是国际性的，因为这些系统是跨国界、跨地区的。

But it does have to be international because, of course, these systems are cross border, and they're cross territory.

最终的验证机构是谁呢？

Who is that ultimate verification system?

你显然一开始提到了主题公园。

You obviously started with Theme Park.

是的。

Yes.

很长

A long

很久以前。

time ago.

是的。

Yes.

太棒了。

Brilliant.

别把汉堡放得太靠近过山车。

Don't put the burgers down too close to the roller coaster.

但你知道，作为一家媒体公司，我会接触各种媒体平台，却分不清什么是真实的，什么是虚假的。

But, you know, obviously, as a media company, I go through any media platforms that I don't know what's real or fake.

我总是不得不追问什么是真实的，什么是虚假的。

I'm always having to ask what's real or fake.

谁是那个验证的仲裁者？

Who is that arbiter of verification?

对。

Yeah.

嗯，我认为最终还得靠政府，我觉得。

Well, I think there are I mean, ultimately, it's gotta be government, I think.

但你知道，真正能做技术工作的机构可能是像人工智能安全研究所这样的组织。

But, you know, the types of technical bodies that would be able to do the technical work would be like maybe the AI Safety Institutes.

英国就有一个非常出色的机构，是在首相任内成立的，我认为他们做得非常好。

You know, there's a very good one in The UK that, you know, was set up under prime minister, I think he's doing great work.

美国也有一家。

And then there's one in The US.

也许那些拥有顶尖研究能力的国家也应该设立类似的机构，配备高质量的研究人员，能够根据某些基准评估和审计这类系统。

And maybe some of the leading countries that have the best research should also have an equivalent body that is staffed with high quality researchers too that can actually evaluate and audit these kinds of systems against certain benchmarks.

我会独立核查它们是否达到了应有的标准。

And I kind of, like, independently check whether they are, meeting the right standards.

如果我能拥有一个仅适用于人工智能安全的魔法棒，你会用它来实施什么样的具体方案呢？

If I could give you, like, a magic wand that was only applicable to AI safety, sadly, what would be your implementation idea program that you would put in place with this magic wand?

是的。

Yeah.

我认为我们需要一个类似的国际机构，也许类似于原子能机构，AI安全研究所可以向其提供支持。

I think we need some kind of international body, maybe similar to the Atomic Agency, something like that, that perhaps the AI Safety Institute sort of feed into.

研究界也必须参与进来，共同确定哪些是合适的基准，用于评估哪些特质和能力。

And the research community has to also do this and be involved in, like, what are the right set of benchmarks to check what types of traits, what types of capabilities.

也许还有其他保障措施，比如，AI系统输出人类无法理解的标记是不可取的。

Maybe there are other safeguards too, like, you know, it's it wouldn't be desirable to have AI systems output tokens that are not human readable.

也就是说，以我们无法理解的某种机器语言输出，我认为这会引入新的漏洞。

So, you know, in some kind of machine language that we couldn't understand, I think that would introduce a new vulnerability.

因此，还有不少类似这样的问题，我认为大多数领先实验室都会同意，最好不要这样做。

So there's quite a few sort of things like that, which I think most of the leading labs would agree probably not best to do.

而这些机构会针对这些方面进行测试。

And then these bodies would you know, these institutions would test against those things.

我认为这会增强公众的信心。

And I think that would give the public confidence.

此外，学术界和民间社会也可以参与进来，确保这些即将变得极其强大的系统已经经过了独立的检查和审计。

And and, you know, academia could be involved as well, as well as civil society that these systems, which are gonna get incredibly powerful, have been independently checked and audited.

就这样了。

That's it.

你的魔法棒现在完成了。

Your magic wand's done now.

是的。

Yeah.

那就是那个。

Then that was the one

也许我以前做错了事。

Maybe maybe I used to done the wrong thing.

但是啊。

But Yeah.

时间会证明一切。

Time will tell.

是的。

Yes.

没错。

Exactly.

你刚才提到，科学是近年来最令人兴奋的领域之一，但我必须问一下，因为这也是最大的担忧之一：劳动力替代问题。

You you said there about science being one of most exciting areas in I years' have to ask it because it's one of the biggest concerns, is the labor displacement problem.

我刚才其实请了马克·安德森来做节目，他说我是个马克思主义者。

I just had Marc Andreessen on the show actually, and he said that I was a Marxist.

我知道。

I know.

我觉得这简直是个误会。

Which I thought was like, it's a bulge.

为了提出……嗯，马克思很棒。

For bringing Well, Marx wonderful.

所以我并不怪他，他说这完全是胡说八道。

So I'm not blaming him, he was like, it's completely rubbish.

是的。

Yeah.

我完全不同意这一点。

I don't agree with that at all.

我们一直都能克服它。

We've always overpassed overcome it.

当你看到这些系统真正强大的能力时，你怎么看待劳动力替代问题？

How do you think about the labor displacement problem when you look at how truly capable these systems are

是的。

Yeah.

这对劳动力市场会造成什么影响？

And what that does to labor markets?

当然，你知道，过去每一次革命性技术的出现，都会导致大量工作岗位的变动。

Well, certainly, you you know, in the past, with every new revolutionary technology, there's been a lot of jobs disruption.

这一点毫无疑问。

So that's for sure.

我认为这肯定会发生。

And I think that's definitely gonna happen.