在两大 AI 巨头 Anthropic与DeepMind 的达沃斯论坛对话中，我看到的2026年
📕链接：传送门🔗
会议视频链接：The Day After AGI | World Economic Forum Annual Meeting 2026
文章内容
2026年1月，DeepMind CEO Demis 和 Anthropic CEO Dario在 达沃斯论坛上进行了一场对谈。
趁着假期，博主仔细回顾了这场论坛，收获了不少insight，借这篇笔记，做一个简单的记录和分享。
整场讨论最触动我的，是大佬们对于 2026年的大风向的判断。
过去的2025年，行业的焦点是人类如何构建并落地各类 AI Agent；而到了2026年，行业焦点可能会转向「构建能够自己构建出下一代AI 的AI」，也就是把构建下一代 AI 的工作交还给 AI 本身，彻底激活它的 自我提升循环（Self-improving Loop）。
首先简单概括一下现场的对话：
会议一开始，主持人提到Anthropic的Dario曾在25年的会议上预言，一个能在多领域达到诺贝尔奖得主水平、完成人类所有工作的模型（a model that can do everything a human could do at the level of a Nobel laureate across many fields）可能会在2026或者2027年出现。现在已是2026年，他是否还坚持这个时间预测？
Dario解释道，他认为这 取决于AI完成自我提升（Self-improving）循环（Loop）所需要的闭环时间，即AI模型还需要多久才能自主地对自己进行升级。
在他看来，这个闭环需要AI具备两个关键能力：一个是足够强的编程能力（good at coding），另一个是足够强的做AI研究能力（good at AI research）。在这两个能力的加持下，AI能够发现自己的不足，提出创新想法，并用编程的方式实现自己的想法，从而完成自我进化。
目前，AI在编程方面已经有明显突破，但在做AI研究方面仍有较大距离。此外，整个闭环还涉及如芯片、算力、模型训练周期等现实因素。
总体来说，Dario认为 这个“闭环时间”不会太远，甚至可能比很多人预想得更快， 但他也承认，目前的预测 仍有很大的不确定性（a lot of uncertainty）。
而Deepmind的Demis则认为，当下的AI在编程和数学上之所以进展较快，是因为这两个学科的验证机制（verifying）相对简单明确。但是 在自然科学的研究上，验证往往需要长期实验，使得AI的推进难度显著提升。
总体来说，Demis对「AI自我提升的循环」能否在没有人类参与的情况下闭环，持一定的保留态度；同时他也指出，仅靠工程和数学也许并不能解决所有科学问题，这个理论上限在哪目前仍不清晰。所以即便AI攻克了工程和数学， 也不一定能攻克所有的科学问题。
但是他们都赞同，如果这种AI自我提升的闭环真的实现，AI的进化速度将显著加快，并给社会带来前所未有的安全挑战
Dario更是说到，也许未来几年我们的核心挑战在于：如何有效控制这些高度自主且智能超越人类的系统（How do we keep these systems under control that are highly autonomous and smarter than any human）
在会议的中间，主持人向两位嘉宾问了一系列AI可能给社会带来的问题，这里不再展开。
但在会议的最后，主持人问Dario和Demis未来这一年，AI行业可能会发生的巨变是什么？
Dario又再一次强调，他认为是「AI systems building AI systems」
Demis 基本认同Dario的这一判断，但表示他还会关注「世界模型」和「持续学习」等方向，因为他认为如果AI自我提升这种方式无法达到预期效果的话，这些技术可能会成为关键；此外，他认为机器人领域可能迎来突破时刻。
以上，是博主对这场对谈的简单梳理。
非常有意思的是，站在现在 2026 年 2 月这个时间点往回看这场1月20号的会议，我们会发现他们所讨论的观点，在 1 月底随着 OpenClaw 和 Moltbook 的爆火，竟已在一定程度上得到了初步验证。
当然，OpenClaw 目前还没有进化到能够直接修改自身基座模型的程度，它更多是对自身 Agent 能力的扩展。但它确实让“自我提升”这件事，从抽象概念，变成可以被具体讨论的技术。
以上是自己对达沃斯论坛这场讨论的简单整理。之后，我想针对 OpenClaw 等近期技术再做一些分享，敬请期待～
相关摘录
Part1
Moderator
The title of our conversation is the day after AGI, which I think is perhaps slightly getting ahead of ourselves because we should probably talk about how quickly and easily we will get there. And I want to do a bit of a sort of update on that and then talk about the consequences.
So firstly on the timeline, Dario, you last year in Paris said we’ll have a model that can do everything a human could do at the level of a Nobel laureate across many fields by 26 27. We’re in 26, do you still stand by that timeline?
Dario from Anthropic
It’s always hard to know exactly when something will happen, but I don’t think that’s going to turn out to be that far off.
So the mechanism where I imagined it would happen is that we would make models that were good at coding and good at AI research, and we would use that to produce the next generation of model and speed it up to create a loop that would increase the speed of model development.
We are now in terms of the models that write code, I have engineers within Anthropic who say, I don’t write any code anymore. I just let the model write the code, I edit it, I do the things around it. I think, I don’t know, we might be 6 to 12 months away from when the model is doing most, maybe all of what Sws do end to end.
And then it’s a question of how fast does that loop close? Not every part of that loop is something that can be sped up by AI, right? There’s like chips, there’s manufacture of chips, there’s training time for the model. So know, I think there’s a lot of uncertainty. It’s easy to see how this could take a few years. It’s very hard for me to see how it could take longer than that, But if I had to guess, I would guess that this goes faster than people imagined, that that key element of code and increasingly research going faster than we imagine, that’s going to be the key driver.
It’s really hard to predict, again, how much that exponential is going to speed us up. Something fast is gonna happen.
Demis from Google Deepmind
Yeah, look, I think I’m still on the same kind of timeline. I think there has been remarkable progress, but I think some areas of kind of engineering work, coding, or so you could say mathematics, are a little bit easier to see how they’ll be automated, partly because they’re verifiable, what the output is.
Some areas of natural science are much harder to do than that. You won’t necessarily know if the chemical compound you’ve built or this prediction about physics is correct. You may have to test it experimentally, and that will all take longer. So I also think there are some missing capabilities at the moment, in terms of like not just solving existing conjectures, or existing problems, but actually coming up with the question in the first place, or coming up with the theory or the hypothesis. I think that’s much, much harder. And I think that’s the highest level of scientific creativity. And it’s not clear, I think we will have those systems. So I don’t think it’s impossible, but I think there may be one or two missing ingredients.
It remains to be seen how, first of all, can this self-improvement loop that we’re all working on actually close without a human in the loop? I think there are also risks to that kind of system, by the way, which we should discuss, and I’m sure we will, could speed things up. If that kind of system does work.
Part2
Moderator
We’ll get to the risks in a minute. But one other change, I think, in the past year has been a kind of change in the pecking order of the race, if you will.
This time a year ago, we just had the deepseek moment and everyone was incredibly excited about what happened there. And there was still a sense that Google DeepMind was kind of lagging OpenAI.
I would say that now it’s looking quite different. I mean, they’re code red, right? It’s been quite a year. So talk me through what specifically you’ve been surprised by and how well you’ve done this year and whether you think. And then I’m going to ask you about the lineup.
Demis from Google Deepmind
Well, look, I think I was always very confident we would get back to sort of the top of the leaderboards and the sota type of models across the board.
I think we’ve always had like the deepest and broadest research bench. And it was about kind of marshalling that all together and getting the intensity and focus and the kind of startup mentality back to the whole organization.
And it’s been a lot of work, but I think, and we’re still a lot of work to do. But I think you can start seeing the kind of the progress that’s been made in both the models with Gemini 3, but also on the product side with Gemini app getting increasing market share. So I feel like we’re making great progress, but there’s a ton more work to do and, you know, we’re bringing to bear. Google DeepMind is kind of like the engine room of Google, where we’re getting used to shipping our models more and more quickly into the product.
Part3
Moderator
But let’s then go on to the predictions area now, and we are supposed to be talking about the day after AI, but let’s talk about closing the loop. The odds that you will get models that will close the loop and be able to power themselves, if you will, because that’s really the crux for the winner takes all threshold approach.
Do you still believe that we are likely to see that, or is this going to be much more of a normal technology where followers and catch up can compete?
Demis from Google Deepmind
The full closing of the loop, though, I think is an unknown. I mean, I think it’s possible to do. You may need AGI itself to be able to do that in some domains, again, where these domains you know where there’s more messiness around them, it’s not so easy to verify your answer very quickly. There’s kind of MP hard domains. So as soon as you start getting more, and I also include, by the way, for AGI, physics AI, robotics, working, all of these kind of things. And then you’ve got hardware in the loop that may limit how fast the self-improving systems can work. But I think in coding and mathematics and these kind of areas, I can definitely see that working. And then the question is more theoretical one is what is the limit of engineering and maths to solve the natural sciences?
Dario from Anthropic
And that’s the frame that I used, which is that we are knocking on the door of these incredible capabilities, the ability to build basically machines out of sand. I think it was inevitable the instant we started working with fire, but how we handle it is not inevitable.
And so I think the next few years, we’re going to be dealing with how do we keep these systems under control that are highly autonomous and smarter than any human?
What are the economic impacts? I’ve talked about labor displacement a lot, and what haven’t we thought of, which in many cases may be the hardest thing to deal with at all.
So I’m thinking through how to address those risks and know for each of these, it’s a mixture of things that we individually need to do as leaders of the companies and that we can do working together. And then there’s going to need to be some role for wider societal institutions like the government in addressing all of these.
I just feel this urgency that every day there’s all kinds of crazy stuff going on in the outside world outside AI, right? But my view is this is happening so fast and it is such a crisis, we should be devoting almost all of our effort to thinking about how to get through this.
Part4
Moderator
Let’s start with jobs, because you actually have been very outspoken about that. And I think you said that half of entry-level white collar jobs could be gone within the next one to five years.
But I’m going to turn to you Demis, because so far we haven’t actually seen any discernible impact on the labor market. Yes, unemployment has ticked up in the us, but all of the kind of economic studies I’ve looked at and that we’ve written about suggest that this is over hiring post pandemic, that it is really not AI driven. If anything, people are hiring to build out AI capability.
Do you think that this will be, as you know, economists have always argued that it’s not a lump of labor fallacy, that actually there will be new jobs created because so far the evidence seems to suggest that?
Demis from Google Deepmind
I think we’re going to see this year the beginnings of maybe impacting the junior level, entry level, child of jobs, internships, this type of thing. I think there is some evidence. I can feel that ourselves maybe like a slow down in hiring in that, but I think that can be more than compensated by the fact there are these amazing creative tools out there pretty much available for everyone, almost for free, that if,
you know, I was to talk to us, a class of undergrads right now, I would be telling them to get really unbelievably proficient with these tools.
I think to the extent that even those of us building it, we’re so busy building, it’s hard to have also time to really explore the almost the capability overhang, even today’s models and products have let alone tomorrows and I think that can be maybe better than a traditional internship would have been in terms of sort of leapfrogging yourself to be useful, a useful in a profession. So I think that’s what I see happening probably probably in the next five years, maybe again, slightly different on time scales than that what happens after AGI arrives that’s a different question because I think really we would be in uncharted territory at that point.
Dario from Anthropic
I have about the same view. I actually agree with you and with Demis is that at the time I made the comment, there was no impact on the labor market. I wasn’t saying there was an impact on the labor market at that moment, and now I think maybe we’re starting to see just the little beginnings of it in software encoding.
I see it within Anthropic where, look, I can kind of look forward to a time where on the more junior end and then on the more and the more intermediate, and we actually need less and not more people. And we’re thinking about how to deal with that within Anthropic, since in a sensible way. One to five years. As of six months ago, I would stick with that. If you kind of connect this to what I said before, which is we might have AI that’s better than humans at everything in maybe one to two years, maybe a little longer than that, those don’t seem to line up. The reason is that there’s this lag and there’s this replacement thing.
Part5
Moderator
So I’m going to give you a chance to answer that in the context of a slightly broader question, which is, over the past year, have you grown more confident of the upside potential of the technology, science, all of the areas that you have talked about a lot? Or are you more worried about the risks that we’ve been discussing .
Demis from Google Deepmind
as I’ve been working on this for 20 plus years. So we already knew, look, the reason I’ve spent my whole career on AI , is the upsides of solving basically the ultimate tool for science and understanding the universe around us. I’ve I’ve sort of been obsessed with that since a kid and building AI is the, you know, should be the ultimate tool for that If we do it in the right way.
The risks also we’ve been thinking about since the start, at least the start of demin 15 years ago and we kind of sort of foresaw that if you’ve got the upsides, it’s a dual purposes technology. So it could be repurposed by, say, bad actors for harmful lens. So we’ve needed to think about that all the way through.
But I’m a big believer in human ingenuity. But the question is having the time and the focus and all the best minds collaborating on it to solve these problems. I’m sure if we had that, we would solve the technical risk problem. It may be we don’t have that. And then that will introduce risk because we’ll be sort of, it’ll be fragmented, there’ll be different projects and people will be racing each other. Then it’s much harder to make sure, you know, these systems that we produce will be technically safe. But I feel like that’s a attractable, problem If you, if we have the time space.
Part6
Moderator
This could be a great discussion, but is out of scope for the next 36 seconds, but what isn’t 15 seconds each what when we meet again, I hope next year, the three of us, which I would love, what will have changed by then?
Dario from Anthropic
I, well, I think the biggest thing to watch is this issue of AI systems, building AI systems, how that goes, whether that goes one way or another, that will determine whether it’s a few more years until we get there or if we have, you know, if we have wonders and a great emergency in front of us that we have to face.
Demis from Google Deepmind
I agree on that. So we’re, we’re keeping close touch about that. But also I think, outside of that, I think there are other interesting ideas being researched like world models, continual learning, these are the things I think they will need to be cracked if self improvements and doesn’t sort of deliver the goods on its own, then we’ll need these other things to work. And then I think things like robotics may have it sort of breakout moment.
