2026年人工智能展望：基础模型、IPO与机器人技术——与Sarah Guo和Elad Gil的对话

还有其他对2026年的预测吗？

Any other predictions for '26?

有的。

Yeah.

我觉得我原本不相信我们会看到这么多除了ChatGPT之外的独特消费体验。

I I I think that I did not believe that we were gonna see that many, like, unique consumer experiences besides, like, ChatGPT.

我认为我们会看到一系列消费硬件，它们大多会失败，但我对此持开放态度。

I think we are gonna see, like, a slate of consumer hardware that mostly fails, but I'm so open minded to it.

而且确实，这让我想到去看看这些是否能规模化，但我已经看到了一些非常神奇的消费型代理软件体验，我确实想要并会使用它们。

And then definitely, actually, like, it reminds me to see if any of these scales, but I am seeing magical experiences of, like, really different consumer agent software that I, like, I actually want and will use.

我认为人们正开始真正意识到，这些公司目前还处于保密状态。

And I I think people are really beginning to well, these companies are in stealth right now.

但我确实认为，明年会有更多产品经理和模型公司去尝试这个方向。

But I I do think that, like, there's gonna be a lot more product people that experiment with this and model companies that experiment with this next year.

所以我对此非常乐观。

And so I'm I'm pretty optimistic about that.

是的。

Yeah.

我完全同意。

I agree with that a 100%.

我认为，关键问题是哪家初创公司会脱颖而出，毫无疑问会有一些公司做到。

And I think, the big question is what will end up being a breakout startup, and it'll undoubtedly be some.

然后，哪些初创公司会快速增长，但随后被主要实验室或谷歌复制，并最终整合进核心产品中。

And then what will be a startup that'll grow really fast, and then it'll get copied by the main lab slash Google, and then it just gets incorporated into the core product.

有趣的是，除非一家公司真正实现起飞，建立起网络效应或其他真正可防御的壁垒，否则通常现有巨头在两三年后就能追上。

And the the interesting thing is unless a company truly hits escape velocity and build out a network effect or something else that's really defensible, usually incumbents can launch two, three years later and catch up.

所以，如果它们拥有渠道、核心产品，但正如你所说，我觉得这非常令人兴奋，我已经等待很久了。

And so if they have the distribution and they have the core product and they have but, you know, to your point, I think it's very exciting, and I've been waiting for this for a while.

我认为两年前、三年前，当时在我团队里的大卫·宋在斯坦福大学主导了一个为期两个学期的项目，我们从那里的工程专业招募了不同团队。

I think two years ago, three years ago, this guy David Song, who was on my team at the time, ran a two quarter thing at Stanford where we had different team supply, from the engineering programs there.

当时是人们组队使用人工智能构建消费类应用。

And it was, like, groups of people building consumer apps using AI.

因为我们说这一波AI太迷人了。

Because we said this wave of AI is so fascinating.

为什么没人做任何面向消费者的产品呢？

Why isn't anybody building anything consumer?

所以我们基本上就是免费提供GPU，让人去尝试各种东西。

So we basically just gave people free GPU to go and try stuff.

而且他们那边没有任何义务要做什么，比如让我们参与其中。

And there was no, like, obligation on their side to do anything with it, you know, in terms of us getting involved.

就是去玩点酷的东西吧，因为这是一个绝佳的试验场。

It was just go do cool stuff because this is such a good playground.

那些非常棒的体验当时都只是原型。

And those really neat experiences that were being prototypes.

然后我惊讶地发现，接下来几年里，根本没什么真正有趣的消费者产品出现。

And then I was just shocked that nothing happened for a couple years in terms of, you know, really interesting consumer products.

所以我和你意见一致。

So I agree with you.

这方面还有很大的空间。

There's so much room for that.

我总是想知道，是不是因为新一代的创业者不想做消费类产品，或者已经忘了怎么做？

And I always wonder, is it because there's a different generation of founders who don't wanna work on consumer or forgotten how?

因为那些大型消费类公司已经逐渐老化了。

Because, you know, the big consumer companies have kinda aged out.

是因为现有巨头太吓人了吗？

Is it the incumbents are just too scary?

到底为什么AI的消费端创新这么少呢？

Is it like, why is there so little innovation actually on the consumer side of AI?

我仍然不太明白问题出在哪里。

I still don't quite understand what the issue is.

我 okay。

I okay.

我们来列一列原因吧。

Let's let's, like, list the the reasons.

我认为现有的巨头确实很吓人。

I do think that the incumbents are pretty scary.

任何经历过上一代有趣消费者创意的人，都亲眼目睹了这些创意被整合进现有平台的过程，正如你所说。

And anybody who was around for the last generation of interesting consumer ideas saw actually the ingestion of those ideas into the existing platform as you put out.

是的。

Yeah.

所以这一点也是原因之一。

So there's that.

我还觉得，我看到的公司和创始人在打造新型消费者体验时，第一反应通常是用这一代技术去打造上一代体验的更好版本。

I also think, like, the first instinct that that I've seen from companies and from founders working on, like, new consumer experiences is essentially building, like, better versions of, like, last generation experiences with this generation technology.

但结果往往并不那么有趣。

And it ends up, like, not being that interesting.

因此，我认为你必须要么非常接近前沿研究，要么极具创意野心，才能打造出真正不同的东西，才有可能成功。

And so I actually think you have to be, like, either quite close to research or pretty creatively ambitious to build, like, something very different that has any chance.

所以我觉得，真正拥有这种经验或这种创造力的人并不多，而我们现在即将看到它们出现。

And so I think, like, I think, like, there's just not that many people who have had that experience set or that creativity, and now we're gonna see it.

是的。

Yeah.

我觉得这非常令人兴奋。

I think it's pretty exciting.

另一件事是，我最近和一位非常知名的消费者领域创始人聊过，他经营着一家大型上市公司，他认为在整个世界上，可能只有几百位真正专注于消费者产品的好产品人才。

The other thing is, I was talking to a really well known consumer founder who's running, you know, a giant public company, and his view is that perhaps in the entire world, there's a few 100 great product people for consumer, at least in terms of who are actually working on it.

当然，人类的潜力是巨大的，很多人并没有从事消费者产品工作。

Obviously, there's enormous human potential, and people aren't working in consumer products could.

但在那些从事消费者产品的人中，真正出色、能够独立构思并推出有趣或优秀产品的，可能只有几百人。

And, you know, but of the people working in consumer products, you can sit most with a few 100 people who are exceptional who could actually come up with and launch their own product that would be interesting or good.

因此，你也可以认为，鉴于当前从事这一领域的人才数量有限，这类创新产品的数量本身可能就存在上限——我觉得这是一个很有意思的观点。

And so you could also just say say that maybe there's just a limitation on how many of these things can exist just given human potential within the set of people who are already doing it, which I think is kind of an interesting argument.

我不确定自己是否同意这个观点，但我觉得他提出的这个论点确实很有意思。

I don't if I agree with it, but I thought it was an interesting argument that he made.

如果这个人数也限定于那些真正了解当前可能性的人，那我也会认同这个数字。

I would limit myself to that number if it's also the set of people who, like, have the context of, like, what is possible now.

嗯。

Mhmm.

如果你有出色的消费产品直觉，但却在不断打磨现有产品的第五十次迭代。

If you've got great consumer product instinct, but you're, like, work you're, like, grinding away on the, like, fiftieth iteration of an existing product.

比如

Like

嗯。

Yeah.

嗯。

Yeah.

你是在处理Gmail里那个小小的子按钮之类的，而不是真正去从底层数据库做起。

You're working on the the the little sub button in Gmail or whatever instead of actually doing off the database.

完全正确。

100%.

嗯。

Yeah.

酷。

Cool.

还有其他我们该讨论的吗？或者对2026年有什么其他重大预测？

Anything else we should talk about or any other big predictions for '26?

我觉得今年发生的一个非常重要的新兴事件是Neolabs获得了令人惊讶的融资，大约三到八亿美元。

I feel like a very big, emergent thing that happened this year was the surprising funding of, like, Neolabs, like, three through eight.

你对此怎么看？

What do you think of that?

你对替代架构有什么看法？

What do you think about alternative architectures?

比如，你对围绕让强化学习变得更通用、更持续学习的那些研究方向有什么观点吗？

Like, do you have any point of view on, all of the effort around, like, getting reinforcement learning to be more general continual learning, some of the research directions?

你知道吗，我认为现在正在进行大量非常有趣的研究。

You know, I think there's enormous amounts of really interesting research being done.

所以，我认为这些模型在不同方面仍然有巨大的潜力可挖，我觉得这非常令人兴奋。

So I you know, there's a lot of juice to be squeezed out of these models still in different ways, and I think that's really exciting.

最终，这些因素会成为某些方法或模型的资本收益，因为我们知道规模真的很重要，这意味着最终只有少数玩家能聚集起资本，因为资本会流向最有效的领域。

Well, ultimately, these things become capital gains for certain types of approaches or models because we know scale really matters, which means that eventually you have to have to lapse into a handful of players because capital will aggregate the things that are working the most.

它们正在产生收入。

They're generating revenue.

那么问题来了，这些究竟是哪些东西？

And so then the question is, what are those things?

在什么情况下，由于某种原因，使用模式会被锁定？

At what point do things just get locked in from a usage perspective for whatever reason?

你可以想象，随着时间推移，会以各种方式针对这些模型逐步构建起这样的格局。

And there's all sorts of ways you can imagine this being built over time against some of the models.

我觉得这很有趣。

So I think it's interesting.

我觉得这令人兴奋。

I think it's exciting.

我们会看到事情如何发展。

I think we'll see how it plays out.

我认为，要阐明新的研究方向可能有哪些论点，就像伊利亚最近做了一次访谈，他将其描述为—— paraphrase 一下，他基本上说，是的，我相信扩展，当然，但计算资源有一个非无限的下限，我们可以在这一范围内测试各种想法。

I think to articulate what, like, the the arguments could be for, you know, new research directions is like, Ilya, you know, did this interview recently where he describes it as And the age of to to paraphrase, he, like, basically says that, yes, I believe in scaling, of course, but, you know, there's there's some floor of compute that is not infinite where we can test ideas at scale.

如果我们有一些关于如何实现更快速或更高效计算改进的‘秘密’想法，那么这实际上就不仅仅是一场单纯的资源竞赛了。

And then if we have, let's say, secret ideas around, like, how to get to more rapid or more compute efficient improvement, then it actually isn't just a straight resource battle

嗯。

Mhmm.

而如今的这种竞争，确实有点像一场军备竞赛。

Which, like, the rat race does feel a little bit like today.

我认为另一个可能的论点是多种架构，虽然人们已经对此进行了一些研究，但多种架构在广泛的实用性领域中其实非常重要。

I think the other argument you you could take is actually, like, multiple architectures, and people have done some research on this, but multiple architectures are really relevant at big domains of of, usefulness.

只是它们还没有被大规模应用。

They just haven't been scaled.

对吧？

Right?

而且，现在有足够的资本来测试它们，无论是扩散模型、SSM 还是其他什么，这些都将在明年发生。

And, like, there's enough capital out there to test them, be they like diffusion or SSMs or whatever, and that's gonna happen this next year.

然后我认为还有一个资源聚焦的论点。

And then I think there's, like, a, like, a resource focus argument.

对吧？

Right?

如果伊利亚描述的是，某些实验室拥有海量的计算资源，但今天必须将大量算力用于推理，那么你会把多少资源投入到你的特定研究方向上呢？比如自我改进、训练后优化、情感智能或大规模智能体研究？

If Ilya is describing that some set of labs, they have an enormous amount of compute, but they have to spend a lot of that compute on inference today, then how much do you spend on your particular research direction, be it self improvement or post training or emotional intelligence or very large scale out agent stuff?

是的。

Yeah.

这取决于你在做什么，因为推理最终会为你带来收入，从而支撑其他所有开销。

It depends on what you're doing because the inference is what ends up then raising you money to pay for everything else because you're generating revenue.

所以我认为，这本质上是一种以自身收益为杠杆，逐步实现更大规模的自我循环过程。

So I think, sure, that it's effectively your weighted bootstrapping to more and more scale.

所以我一直认为——也许这种想法是错的，我实际上现在觉得它可能是错的——但我一直觉得，最终你会得到一种进化系统，这才是构建AI的真正方式，因为也许我过度类比了生物学：你的大脑实际上是由一系列具有不同功能或任务的模块组成的。

So I always thought perhaps incorrectly I I actually probably think it's incorrect, but I always thought that eventually you end up with evolutionary systems is really how you build AI Because and maybe I'm over extrapolating up a biology where effectively your brain has a series of modules that have different functions or tasks.

对吧？

Right?

你有一个视觉系统，它在很大程度上是预设好的，能够非常有效地处理视觉信息。

You have a visual system that's highly sort of prewired to deal with vision really effectively.

你有不同的区域负责思考和学习。

You have, different areas of pirate thought and learning.

你有记忆。

You have memory.

你有与共情相关的镜像神经元。

You have, mirror neurons that are involved with empathy.

对吧？

Right?

你的大脑在某些方面实际上是非常专门化的。

Your brain is actually very, specialized in some ways.

尽管，显然，当人出生时，大脑的半个半球就相当于半个大脑，但大脑会重新布线并覆盖所有功能。

Although, obviously, as people were born, it's literally like half a brain hemisphere, and the brain rewires and sort of covers all the functionality.

但像这样的一些著名案例。

But, like, a few famous cases like that.

但你知道，从根本上说，你拥有大量演化成高度专业化任务的机制。

But, you know, fundamentally, you have a lot of stuff that evolves into very specialized tasks.

这几乎就像一个MOE系统之类的。

It's almost like a MOE or something.

你知道吗？

You know?

问题是，在你进一步发展人工智能时，你在多大程度上会重演这一过程。

And the question is the degree to which you recapitulate that as you're doing further development of AI.

你什么时候开始只是不断生成大量实例，并通过某种效用函数进行选择、重组，以及所有其他你通常用来让这些机制发挥作用的手段，而不是采用更多分析性、实验性或迭代性的方式呢？

When do you start just spawning off a bunch of instances of something and just have some utility function involving against that you then have some selection and recombining and all the other stuff that you'd kind of do to to try and make some of that work versus how much of it is a more analytical approach or a more experimental and iterative approach or you know?

所以它是以一种有方向性的方式进行的。

So it's or in in a directed way.

因此，我认为提出这个问题非常有趣。

And so I think it's really interesting to ask.

因为如果你再次将生物学作为潜在的先例——尽管可能是个很差的先例——来看蛋白质设计的话。

Because if you look again at biology as a as a potential precedent, although maybe a very bad one, you look at protein design.

长期以来，存在着这些高度分析性设计的蛋白质，但后来人们开发出各种系统，直接取代了它们。

And for a long time, there are these, like, super analytically designed proteins, and then they came up with all these systems and just abolish it.

比如噬菌体展示、诱变扫描之类的各种方法，其效果远胜于你只是坐下来苦思冥想。

You know, like phage display and, like, mutagenic scans and all sorts of things that give you dramatically better results than if you just sat and thought about it.

而现在，当然，我们通过AI基本解决了这个问题，出现了许多非常精准的三维结构预测技术。

And now, of course, we kind of solved it with AI where you have, all these three d structural predictions that are actually very good.

AlphaFold 和其他一些技术在这方面确实取得了突破性进展。

That was, AlphaFold and a few other things that really were breakthroughs there.

所以在AI的语境下，也许我们最终也会走向同样的方向，即依赖这些系统。

So it feels like in the context of AI, maybe eventually we end up there as well, right, where you just involve these systems.

而这可能是一种非常不同的方法和训练方式。

And then that may be a very different type of approach and training.

我认为，这正是事情出现有趣分岔的关键所在。

Know, That may be where I think things really have an interesting break.

这也是为什么人们如此关注代码的原因之一，因为代码可以说是加速AGI发展的启动器。

And that's one of the reasons arguably people are so focused on code because code is arguably a bootstrap into moving faster on development of AGI.

但我认为，代码加上自我进化才是真正有趣且能带来快速突破的途径。

But I think it's kind of code plus self evolution is really the the potential really interesting approach to it to to get some really fast lift off.

但也许并不是。

But maybe not.

对吧？

Right?

我们走着瞧。

We'll see.

对于2026年，你有什么与人工智能无关的预测吗？

What is, the one prediction you have for '26 that has nothing to do with AI?

萨拉，你还会想别的事情吗？

Do you think about anything else, Sarah?

我会想的。

I do.

我开玩笑的。

I'm joking.

真的吗？

Really?

我的意思是，顺便说一句，另一个与AI相关的预测是，我认为国防领域在初创公司和国防技术方面将会加速发展，无论是自主还是非自主系统，总体趋势都将转向无人机系统。

I mean, the other thing, by the way, one other prediction that does have to do with AI is I do think, defense will accelerate in terms of startups and defense tech and the shift to autonomous or not autonomous, but to drone based systems in general.

这彻底改变了我们对战争和国防的思考方式，我认为明年我们将会看到这一趋势加速得更快。

It's a massive reworking of how you think about war and defense, and I think that's gonna be a shoot shot that we'll see go even faster this coming year.

我认为这部分加速是因为特朗普政府对此的处理方式，以及沃伦部长和那里所有人对此的思考。

I think this is accelerating in part to, you know, how the Trump administration has been approaching it and the secretary of Warren and everybody there have been thinking about it.

我认为部分原因在于，现在已经有足够多的初创公司在做有趣的事情。

I think in part just you have enough density now of startups doing interesting things.

所以我认为这是另一个巨大的转变，目前它正处于一个炒作周期中，而我实际上认为，这一点被低估了，因为它将会变得如此巨大。

So I think that's the other thing that's like a huge shift that, you know, it's a hype cycle right now, and I actually think, again, it's a little bit underthought about because it's it's gonna be so big.

除了AI之外，我认为在太空领域，尤其是SpaceX和Starlink方面，正在发生一些非常有趣的事情，我关注的是通信和电话技术。

Outside of AI, I mean, I think there's obvious really interesting things happening in space with SpaceX and Starlink and I think about communications and telephony.

这是一个巨大的转变。

So that's a big shift.

在我看来，能源和采矿领域正在发生一些非常有趣的事情。

There's really interesting things, in my opinion, happening in energy and mining.

你知道，世界上正在发生很多事情。

And, you know, I I think there's a lot going on in the world.

我同意你对国防的看法，但有些担忧，比如我们必须等待预算真正从合同转向这些新公司的大规模投入。

I agree on defense with some, like, concern that, you know, we have to wait for budget to actually shift from contracts to primes to some of these new companies at scale.

但需求是显而易见的——在一个日益走向自主化的世界里，我们必须保持竞争力。

But the demands, like, the need to be competitive in a world that's increasingly autonomy driven is, like, so obvious.

对吧？

Right?

我认为，炒作周期和繁荣期是有好处的，因为它们吸引了大量的人才、资本、创始人和想进入这个行业的人。

And I think, you know, hype cycles and booms are good in that they bring a lot of people to the table, you know, capital, founders, people who wanna work in the industry.

所以，即使很多公司会倒闭，我们仍然能在短时间内取得巨大进展。

And so you can make a lot of progress in a quick amount of time even if a lot of companies die.

嗯。

Mhmm.

是的。

Yeah.

而且在短时间内，热情更加高涨。

And there's there's more enthusiasm over a short period of time.

所以我同意这一点，而且我认为这也不一定是坏事。

So I agree with that, and I also don't think that's necessarily bad.

对吧？

Right?

你的非AI预测是什么？

I What's your non AI prediction?

我认为，像我这样的人不止一个，但我觉得GLP这一件事，尽管有如此多的热情，仍然被低估了，因为它带来的影响巨大。

I think that, like I'm not the only one, but I I think the the, like, GLP one thing is just despite all of the enthusiasm, like, still underrated for how much impact it is having.

对吧？

Right?

所以我认为，这些技术的持续采用是不可避免的。

And so I think that the continual adoption of these is like inexorable.

我实际上认为，它为其他肽类和激素疗法开辟了一条有趣的路径。

I actually think it creates a path that is interesting for like other peptide and hormone therapies.

我认为它如此有效这一事实带来了许多次级影响，一方面人们直接减轻了大量体重，另一方面也促使人们更愿意关注其他工程化肽类；我认为现在每个人都明白了，递送方式至关重要。

I think the fact that it has been so effective has like lots of second order effects, both from people way like, just being a lot less overweight, like, directly and the willingness to look at other engineered peptides or like, I think it like, everybody understands now that, like, delivery matters.

这些药物真的非常出色。

There are these really incredible medicines.

我认为这将推动更多投资进入任何类似机会的领域。

And I think that the impact of that is going to, like, fuel much more investment in, anything that looks like that type of opportunity.

所以我认为是的。

And so I think that's Yeah.

我实际上觉得你提到的一点非常有趣：如果你观察生物黑客群体，现在有很多人使用各种肽类，这些肽类能实现不同的效果，比如有人患有慢性腕管综合征，就会飞到迪拜接受肽类注射之类的。

I actually think, one thing that you mentioned is really interesting where if you look at the sort of biohacking community, there's a lot of peptide use now, different, you know, different peptides that will do different things in terms of, you know, somebody will have some chronic carpal tunnel thing, and they'll fly to Dubai to get, you know, peptides injected or whatever.

通常，这些是社会层面更大规模采用的早期信号。

And, usually, those are sort of early indicators of potential larger scale adoption societally.

因此，我认为目前整个肽类及其用途的世界是一个非常有趣的趋势。

And so I think that's a really interesting trend right now in general, like, this whole, like, world of peptides and their uses.

那么，肽类物质有没有类似HIMSS的盛会呢？

And is there a HIMSS of peptides?

那接下来会有什么发展呢？

Like, what's the what's coming there?

所以我觉得这非常有趣。

So I think that's super interesting.

是的。

Yeah.

我还觉得，正如你所说，生物黑客群体中，有一群人很早就开始超适应症使用GLP-1，他们对长寿、超声波神经调节、干细胞注射等很感兴趣。

I also think, like, the biohacking community, as you said, it like, the set of people who were really, really early off label GLP one adopters interested in longevity, neuromodulation with ultrasound, stem cell injection, for example.

这曾经是一个边缘的小众群体。

That has been a fringe small community.

但我认为它将不再那么边缘化。

And I think it's going to get less fringe.

而很多这类东西在十年前，通常都源自健美圈。

And a lot of these things traditionally, ten years ago, came out of the bodybuilding community.

对吧？

Right?

健美界以前流行的是肌酸和这些现在广泛使用的东西，还有助眠的镁和其他各种物质。

The bodybuilding community was like creatine and all these things that are more broadly used now, but also other other things for sleep aids or other, you know, magnesium and all those stuff.

为了总结这一年的最后一期节目，我们邀请了一些朋友对2026年做出预测。

And to round out this year end episode, we've asked some of our friends for their predictions for 2026.

我非常好奇。

I'm so curious.

我对明年的预测是，推理系统将直接转化为更加灵活、更加稳健的AI。

My prediction for next year is that the reasoning systems are going to translate directly to AIs that are much, much more versatile, much, much more robust.

推理不仅将改变语言模型，还将影响从生物学到自动驾驶汽车再到机器人技术的每一个行业。

And reasoning is going to impact is gonna revolutionize not just not just language models, but reasoning is going to impact every single industry from biology to self driving cars to robotics.

因此，我认为推理是即将改变众多应用和行业的重大突破。

And so reasoning, I think, is is the big huge breakthrough that that is going to transform a lot of different applications and industries.

到2026年，AI将不再是一个需要我们主动提示的被动工具。

In 2026, AI will stop being a reactive tool that base for us to prompt it.

相反，它将变得非常主动，并深度融入我们的工作生活。

Instead, it will become very proactive and get deeply integrated in our work life.

它会伴随我们到任何地方，听到我们听到的内容，了解我们需要完成的任务，事实上，大多数时候甚至在我们开口之前就会帮我们完成这些任务。

It will go where we go, hear what we hear, know what tasks we need to work on, and in fact, most of the times complete those for us before we even ask it to do so.

它可能是一位帮助我们提升技能的教练。

It could be a coach that helps us improve our skills.

它可能是一位帮助我们优先处理工作、管理时间的经理。

It could be a manager who helps us prioritize our work and manage our time.

简而言之，它将成为我们所能期望的最棒的工作伙伴。

In short, it's going to be the best work companion we could wish for.

我认为明年我对AI的主要预测是，上下文将成为每个产品的最重要部分。

I think the main AI prediction that I have for next year is I think context is just gonna be the most important part of every single product.

老实说，我到目前为止体验最好的就是ChatGPT的记忆功能。

And, honestly, one of the best experiences I've had with it so far is just memory and ChatGPT.

我认为将出现更多功能，其目标都是提取用户意图，减少用户必须提供大量上下文的负担，让模型、系统或产品能获得越来越多的上下文信息。

I think that there are gonna be a lot more features that basically their goal is to extract the user intent and make the onus less on the user to basically give all of the models or the system or the product more and more context.

换句话说，你如何让产品承担起从用户那里提取这些信息的责任，而不是让用户事先做所有这些工作。

So in other words, how do you put the onus on the product to actually extract that from the user instead of the user having to do all of the work to do this upfront.

我对2026年的预测是，将出现一系列基于更快推理运行的全新产品体验。

My prediction for 2026 is there will be a whole new suite of product experiences that run on much faster inference.

我对2026年的预测是我们终于会停止把内容复制粘贴到聊天框里。

My prediction for 2026 is that we'll finally stop copy pasting stuff into chat boxes.

相反，我认为我们将拥有更好地利用屏幕共享和关键来源上下文管理的应用程序。

Instead, I think we're going have applications that have better use of screen sharing and context management across the sources that matter the most.

我对2026年的一个预测。

One prediction for 2026.

现在关于智能代理的讨论非常多，而且已经持续一段时间了，但还没有人真正创造出大规模的消费级智能代理AI。

There's so much talk of agents right now, and there has been for a while, but no one has truly created a mass scale consumer agentic AI.

我认为今天的模型已经具备实现这一目标的能力。

I think the models are there today for this to be possible.

到2026年，我们将看到一个团队成功设计出正确的界面、系统和产品，带来像ChatGPT刚推出时那样巨大的体验跃升。

And in 2026, we will see the group that figures out the right interface and system and product that creates as big a step function and overall experience as Chad did when it first came out.

我认为这个领域并没有像人们想象的那样被实验室广泛占据。

And I think this area is not nearly as seeded to the labs as people assume.

这真的是谁都有机会。

It really is anyone's ballgame.

你好。

Hello.

我是亚伦。

Aaron here.

首先，我拍自拍视频时总是觉得很别扭。

First of all, I get quite awkward around doing selfie videos.

这是我拍的第九遍视频了，希望效果还行。

This is my ninth take of this video, so I hope it goes okay.

但我的2026年预测是，这将是AI代理持续发展的第二年，特别是AI代理在企业中，尤其是在深度垂直或特定领域中的应用。

But 2026 prediction would be that this is going to be certainly the continued year number two of AI agents, but in particular, AI agents in the enterprise in either deep vertical or domain specific areas.

我认为这将成为我们将当前AI模型取得的所有进展真正落地到企业中的主要方式。

I think this is going to be the main way that we actually take all of the progress that we're seeing in AI models and actually deliver them into the enterprise.

你必须能够与组织的工作流程相结合。

You have to be able to tie to the workflow of the organization.

你必须能够访问他们拥有的数据。

You have to get access to the data that they have.

你必须有正确的上下文工程，才能让智能体真正发挥作用。

You have to have the right context engineering to make the agents actually work.

然后你还需要进行变革管理，以确保智能体有效运行。

And then you have to do the change management that makes the agents effective.

因此，今年我们将开始看到这种模式越来越明显，这也意味着我们需要在智能体框架上投入更多工作。

So, this is going to be a year where we start to see this pattern emerge more and more, which equally means that we need to ensure that we have a lot more happening on agent harnesses.

所以，感谢Akhorma、Suhail和Dex给出的回答。

So, shout out to Akhorma, Suhail, and Dex for that answer.

但今年绝对是智能体框架之年，关键在于如何通过为模型搭建完善的支撑结构，实现模型能力数量级的提升。

But it's definitely going be the year of agent harness and seeing how do you start to get, you know, an order of magnitude improvement on the model's capabilities by having all the right scaffolding around the model.

最后，这将是经济实用型评估之年。

And then finally, it will be the year of economically useful evals.

因此，我们开始真正弄清楚这些模型如何在经济中承担越来越多的知识型工作角色。

So really starting to figure out how these models end up doing a lot more knowledge worker cast in the economy.

到2026年，我们将看到更多这样的趋势。

And that's gonna we're gonna see a lot more of that in 2026.

今年我们已经看到一些预兆，比如Apex、GDP Val以及少数其他案例。

We saw some previews of that this year with Apex and GDP Val and a handful of others.

我们将看到远多于此的现象。

We're gonna see way more of that.

这些就是我们的预测，我们2026年再见。

So those are the predictions, and we'll see you in 2026.

我认为2026年将是美国开源模型非常有趣的一年。

I think 2026 is going to be a very interesting year for American open models.

在过去一年中，开源智能的前沿已从美国转向中国，始于2024年DeepSeek的发布。

Over the last year, the frontier of open intelligence shifted from America to China, starting with the release of DeepSeek at the 2024.

美国机构对开源智能领域美国领导力的这种衰退反应迟缓。

And American institutions were slow to notice this erosion of American leadership in open intelligence.

但我认为在过去半年里，他们已经以一种显著的方式意识到了这一点，无论是从政府层面、企业层面，还是开始涌现出一些以开放智能为使命的新兴实验室，这类机构不止一家，不仅仅是Reflection。

But I think they've noticed in a big way over the last half year, both from the government level, from the enterprise level, and there are some really interesting neo labs starting to come out with open intelligence as their directive and there are a few of these, not just reflection.

这些公司正开始推出一些非常有趣的中小型开源模型，明年我认为我们将看到美国在最大规模的开源权重前沿重新夺回领导地位。

And these companies are starting to produce some very interesting small open models and next year I think we'll see The US regaining leadership at the open weight frontier at the largest scale.

我非常期待看到这一点。

And I'm really excited to see that.

大家好。

Hey, folks.

我对2026年的预测是，我认为人工智能将变得更加政治化。

My prediction for 2026 is that I think we will see AI become much more politicized.

我认为它将成为2026年中期选举的重要讨论议题，有些人会强烈反对它。

I think we'll see it become a major point of discussion for the twenty twenty six midterm elections, and some people will come out strongly against it.

有些人则会强烈支持它。

Some people will come out strongly supportive of it.

我不确定哪一方最终会占上风。

And I'm not sure which size is gonna win out.

2025年是人工智能药物发现领域令人惊叹的一年。

2025 has marked an incredible year in AI drug discovery.

在过去的一年里，我们从能够在计算机上设计简单分子，发展到设计简单的抗体，如今最新进展是无需任何示例就能在计算机上设计出具有药物特性的全长抗体。

In the past year alone, we've gone from being able to design simple molecules on the computer to designing simple antibodies and now most recently full length antibodies with drug like properties zero shot on the computer.

如果2025年是人工智能药物发现的研究之年，那么2026年将是部署之年。

If 2025 has been the year of research in AI drug discovery, 2026 will be the year of deployment.

这些模型终于进入了一个真正对药物发现变得有用的阶段。

The models have finally entered an era where they're becoming really useful for drug discovery.

它们不仅加快了进程，还让我们能够攻克那些传统技术长期以来难以应对的极具挑战性的靶点。

Not only do they make things faster, but they're also allowing us to go after really challenging targets which have been traditionally really difficult to do with traditional techniques.

我非常期待接下来的发展，因为这些模型丝毫没有放缓的迹象。

I'm really excited to see what comes next because the model show no signs of slowing down.

好的。

Okay.

我对2026年的预测是：那将是YOLO走向终结的一年。

My prediction for 2026 is it will be the year that YOLO dies.

我们将开始把自己从‘人生只有一次’转变为‘不要死去’。

We will begin transforming ourselves from a you only live once to don't die.

我认为我们现在某种程度上是一种自毁的物种。

I think right now we're kind of a suicidal species.

我们做着非常原始的事情。

We do very primitive things.

我们通过饮食毒害自己。

We poison ourselves with what we eat.

我们设计的生活方式让我们慢慢自杀。

We design our lives so that we slowly kill ourselves.

企业通过让我们上瘾和痛苦来获利。

Companies make profits by making us addicted and miserable.

我们摧毁了唯一的家园，却 somehow 将这些行为视为美德。

We destroy the only home we have and somehow we celebrate these things as virtue.

我觉得一切都颠倒了。

I think it's all backwards.

我认为总有一天我们会回望过去，惊讶于自己曾经如此行事。

And I think one day we'll look back and we'll be pretty astonished that we behaved like this.

我认为即将到来的转变将简单而彻底：我们对生命说‘是’，对死亡说‘不’。

I think the shift coming is gonna be simple and radical that we say yes to life and no to death.

这很简单，但我认为这可能是对人工智能进步的回应。

It's simple, but I think it could be in response to AI's progress.

我们将以一种坚定的方式团结起来，做出这样的选择。

And we do this defiantly as a form of unification.

不过，我认为我们要承认自己存在的神圣性，这需要极大的勇气。

I think it does require a lot of courage for us, though, to say we recognize how sacred our existence is.

我们不想抛弃它，而是想用尽全部的勇气和力量去捍卫它，因为它如此珍贵。

We don't wanna throw it away, and we want to defend it with every bit of courage and strength we have, because it is so precious.

我认为，今年将是终结‘人生只有一次’、开启‘不要死去’的年份。

I think it's gonna be the year we end YOLO and the beginning of don't die.

明年最引人注目的事情是，其他类型的知识工作将经历软件工程师如今所感受到的变化：他们年初时还主要靠自己编写代码，到年底却几乎不用自己敲代码了。

The most striking thing about next year is that the other forms of knowledge work are gonna experience what software engineers are feeling right now, where they went from typing, you know, most of their lines of code at the beginning of the year to typing barely any of them at the end of the year.

我认为这是所有形式的知识工作都将经历的Claude编码体验。

I think of this as the Claude code experience for all forms of knowledge work.

我还觉得，持续学习可能会以一种令人满意的方式被推广，我们会看到家用机器人首次投入测试，而软件工程本身在明年也将变得极其狂野。

I also think that probably continual learning gets sold in a satisfying way, that we see the first test deployments of home robots, and the software engineering itself goes utterly wild next year.

我对2026年的预测是，那将是所有人认知彻底反转的一年。

My prediction for 2026 is that it's the year where everyone's perceptions are flipped.

目前，每个人都认为除了谷歌之外，只能使用NVIDIA，但到那时，这一点将明显是错误的。

Currently, everyone believes that you can only use NVIDIA outside of Google, and that will be obvious that that's not the case.

目前，大约三分之一的美国人讨厌人工智能，认为它非常糟糕。

Currently, about a third of Americans hate AI and think it's really bad.

这个数字将会增加。

That number will increase.

目前，大多数美国人认为人工智能没有用处。

Currently, most Americans think AI is not useful.

这一点也将发生反转。

That will flip as well.

因此，每个人的先验认知都将被颠覆。

And so everyone's priors will be flipped.

这是因为人工智能的变革性应用将变得无处不在。

That's because the transformative use of AI will be so prevalent.

它的明显实用性将如此之高，以至于任何人都无法再维持原有的认知，认知失调将被彻底消除。

The the obvious utility of it will be so high that there is no way for anyone's priors, you know, cognitive dissonance will be wiped away.

嘿。

Hey.

我是本杰明·斯佩克特。

I'm Benjamin Spector.

我是阿舍·斯佩克特。

I'm Asher Spector.

我们的预测是，2026年将是高效能人工智能之年。

And our prediction is that 2026 is the year of energy efficient AI.

数据中心的建设主要受限于能源、电力供应、电网连接和高压设备等因素，这就是为什么XAI的Colossus最初由现场燃气轮机供电的原因。

Data center buildings are primarily constrained by energy, power availability, grid interconnects, high voltage equipment, things like that, which is why XAI's Colossus was initially powered by on-site gas turbines.

关键是，这种动态仍在持续增长。

The thing is the dynamic is continuing to grow.

像我们这样的实验室、新实验室以及像Kirsten这样的初创公司，对训练和计算的需求都非常显著，而当前的需求已经超出了我们向电网输送电力的能力。

Labs, neolabs like us, and startups like Kirsten have a pretty remarkably excisional demand for both training and compute, and this demand is currently outstripping our ability to push lots onto the grid.

这意味着在2026年，从每一分钱中榨取尽可能多的算力将变得至关重要。

This means that in 2026, it will be really important to squeeze every available bit of tons out of every wallet.

不过，从长远来看，芯片可能比电力更重要，因为芯片的折旧速度远快于基础电力基础设施。

That said, in the long term, chips probably matter more than power because chips depreciate much more quickly than the underlying power infrastructure.

例如，在数据中心电力成本为每千瓦时10美分的情况下，芯片的成本在五年折旧周期内实际上超过了电力成本。

So for example, with data center power supplies of 10¢ per kilowatt hour, the chips cost actually in order to manage more than the power in the five year depreciation cycle.

因此，在2026年，我们认为每单位能耗所能实现的智能水平至关重要，必须从每单位能源中榨取尽可能多的智能。

So in 2026, we think intelligence per watch is really important to squeeze as much intelligence you can out of every unit of energy.

但从长远来看，我们认为芯片才是更重要的。

But in the long term, we think it's the chips that matter more.

节日快乐。

Happy holidays.

新年快乐。

Happy New Year.

谢谢这一年的陪伴。

Thanks for the year.

2026年快乐。

Happy 2026.

2026年快乐，各位听众。

Happy 2026, listeners.

谢谢。

Thank you.

在Twitter上关注我们：no priors pod。

Find us on Twitter no priors pod.

如果你想看到我们的脸，请订阅我们的YouTube频道。

Subscribe to our YouTube channel if you wanna see our faces.

在Apple Podcasts、Spotify或你收听的任何平台关注本节目。

Follow the show on Apple Podcasts, Spotify, or wherever you listen.

这样你每周都能收到一集新内容。

That way you get a new episode every week.

并前往 no-priors.com 订阅邮件或获取每集的文字稿。

And sign up for emails or find transcripts for every episode at no-priors.com.