DeepMind & Anthropic: When AGI Arrives | Vital Views

The evolution of AGI is pushing humanity toward its fourth societal transformation.

At the January 2026 World Economic Forum (WEF), The Economist editor-in-chief Zanny Minton Beddoes, Google DeepMind CEO Demis Hassabis, and Anthropic CEO Dario Amodei gathered for a conversation on "The Day After AGI" — examining predictions, responsibility, risks, and global implications of what happens once AGI becomes widespread.

Rather than dwelling on technical details, the three focused on institutional arrangements, global coordination, and divergent development paths. We translated and compiled their exchange for you. The full transcript runs roughly 10,000 words, about a 15-minute read.

Enjoy.

We first build models that are really strong at programming and AI research, then use those models to generate the next generation of models, thereby accelerating the whole process.

The real question then becomes: how fast does that loop close?

— Dario

Moderator: I had the privilege of moderating a conversation between Dario and Demis last year. I said at the time that it felt like hosting the Beatles and the Rolling Stones on the same stage. You two haven't appeared together on stage since. So today is essentially the sequel — the band is back together.

Our topic today is "The Day After AGI." The title may be slightly ahead of itself, since we probably should first discuss how quickly and easily we get there. I'd like to update us on that, and then we can move to the consequences.

Let's start with timelines. Dario, you said in Paris last year that by 2026 or 2027, we would have a model capable of Nobel Prize-level human performance across multiple domains — able to do almost anything. Well, it's now 2026. Do you still stand by that timeline?

Dario: It's always hard to pinpoint exactly when something will happen, but I don't think it's that far off. The path I envisioned was this: we first build models that are really strong at programming and AI research, then use those models to generate the next generation of models, thereby accelerating the whole process — creating a continuously accelerating loop that drives ever-faster model development.

Right now, we're already quite far along on coding. I have engineers at Anthropic telling me they barely write code themselves anymore — they just have the model write it, then they edit and work around it. I think maybe in another six to twelve months, models will be able to end-to-end complete most, if not all, of the work currently done by humans.

The real question then becomes: how fast does that loop close. Of course, not every part of that loop can be accelerated by AI — chips themselves, chip manufacturing, model training time, and so on. So there's a lot of uncertainty. It's easy to imagine this still taking a few years. But it's hard for me to imagine it taking much longer than that.

If I had to guess, I'd say it will move faster than most people expect. And the key driver here is coding ability, plus increasingly more research work, both progressing faster than we originally anticipated. Exactly how fast this exponential acceleration will be is genuinely hard to predict, but it's safe to say some very rapid changes are coming.

Moderator: Demis, you were relatively more cautious last year. You said there was perhaps a 50% chance of a system demonstrating the full range of human cognitive abilities before the end of this decade. As Dario mentioned, progress in coding has been remarkable. Where do you stand now? Do you still hold to that assessment? What has changed in the past year?

Demis: Yes, I'm basically in the same place timeline-wise. I think the past year's progress has indeed been remarkable. But I also think some domains — engineering work, coding, or arguably math — are relatively easier to automate, partly because the results in these areas are verifiable.

Certain areas of natural science are much harder. You don't necessarily know immediately whether the chemical compound you've synthesized is correct, or whether your prediction about some physical phenomenon holds. Often you need experimental validation, and that whole process takes much longer.

Additionally, I think some key capabilities are still missing — not just solving existing conjectures or problems, but being able to ask the right questions in the first place, or construct new theories and hypotheses. That, to me, is the highest tier of scientific creativity, and it's significantly harder. Not that we'll never get there, but perhaps we're still missing one or two critical ingredients.

There's also an open question: whether this self-improvement loop we're all trying to build can actually close completely without human involvement. Of course, such systems carry their own risks, which we'll certainly discuss. If they do work, they could meaningfully accelerate progress.

Moderator: We'll get to risks shortly. But I think another notable change this past year has been the shifting competitive landscape. Just a year ago, we had the DeepSeek moment — everyone was incredibly excited about what had happened — and there was a general sense that Google DeepMind was somehow falling behind OpenAI.

Now, that looks quite different. They've declared a code red, right? It's been a year of massive change. So I'd like to ask you specifically: what genuinely surprised you? How has your year been? And what's your view of the overall landscape?

Demis: I've always been confident that we could get back to the top of the leaderboard and lead comprehensively across state-of-the-art models. The reason is that we've always had the deepest and broadest research foundation in the industry. The key was how to truly integrate all of that, and to recapture the intensity, focus, and startup-like working style across the organization. That itself was a massive undertaking, and there's still much work to do.

But I think people are starting to see the results — both on the model side, with Gemini 3's progress, and on the product side, with the Gemini app's market share continuing to grow. So I feel we're making good progress, though there's still a huge amount of work ahead.

At the same time, we're gradually leveraging Google DeepMind's advantage as Google's "engine room," getting increasingly comfortable pushing models to products faster and actually landing them in various product scenarios.

Moderator: I want to ask you something on this front too, Dario — you're in the middle of a new funding round at a remarkable valuation. But unlike them, you're an independent model company, and there's growing concern about whether such standalone model companies can survive long enough to reach meaningful revenue.

This has been very publicly discussed with OpenAI. Can you talk about how you think about this? Then we can return to AGI itself.

Dario: I think the way we look at it is this: as models become more capable, there's an almost exponential relationship — not just between compute invested and model cognitive ability, but between model cognitive ability and the revenue it can generate.

Over the past three years, our revenue has grown 10x and 10x again: from zero to $100 million in 2023, $100 million to $1 billion in 2024, and $1 billion to $10 billion in 2025. Of course, I'm not sure this curve will continue — if it did, that would be pretty insane. But you can see the scale is already approaching that of some of the world's largest companies.

There's always uncertainty, of course. We're essentially bootstrapping this from scratch, which is itself a pretty crazy thing. But I'm confident that if we keep building the best models in our focused direction, things will trend in a positive direction overall.

Broadly speaking, I also think the past year has been a good one for both Google and Anthropic. What we truly have in common is that we're both research-driven companies — or rather, the research parts of our companies are led by people focused on models and on solving important real-world problems. We use these hard, important scientific problems as our North Star, and I believe it's companies like this that will succeed going forward. I think we're very aligned on that.

I also include AI in the physical world — robotics, embodied intelligence, the whole suite.

Once hardware is involved, that itself may constrain how fast self-improving systems can operate and accelerate.

— Demis

Moderator: Let's turn to the predictions themselves. Today's theme is the day after AGI, but first, let's talk about "closing the loop."

That is, whether models could truly close this loop — become self-driving, self-propelling in some sense — because that's at the heart of the winner-take-all threshold. Do you still think we're likely to see such systems? Or will this ultimately be more like a relatively ordinary technology, where latecomers and catch-up players still have room to compete?

Demis: I can say quite clearly that I don't think this will be an "ordinary" technology. As Dario mentioned, there are already many ways it's showing itself to be different — it's already helping us with programming and some research tasks.

But true "closed-loop" systems, I think, remain an open question. I think it's possible, but in certain domains, it may require AGI itself to achieve. Especially in more complex, messier domains where it's hard to quickly verify whether an answer is correct — these tend to be computationally complex, close to NP-hard problems.

I also include AI in the physical world — robotics, embodied intelligence, the whole suite. Once hardware is involved, that itself may constrain how fast self-improving systems can operate and accelerate.

That said, in domains like programming and mathematics, I can clearly see how the closed loop would work. The more theoretical question that follows is: to what extent can engineering and mathematics themselves be used to solve problems in the natural sciences?

Moderator: Dario published "Machines of Loving Grace"

(https://www.darioamodei.com/essay/machines-of-loving-grace) last year. It was a fairly optimistic piece, laying out the technological potential you see and the future that might unfold. You used metaphors like "AI as geniuses in a national data center."

I hear you're working on a new essay that hasn't been published yet — people will have to wait a bit. Perhaps you could give us a preview: a year later, what is your biggest updated conviction?

Dario: My overall conviction hasn't changed. I've always believed AI will become extraordinarily powerful. I think Demis and I agree on this; we just differ on the timeline. Precisely because it will be so powerful, it has the capacity to deliver the positive visions I laid out in Machines of Loving Grace: helping us cure cancer, potentially eliminating tropical diseases, helping us better understand the universe.

But at the same time, there are genuinely large and serious risks. Not that they're unmanageable — I'm not a doomer — but we have to think seriously about them and confront them head-on. The reason I wrote Machines of Loving Grace first wasn't some deeply calculated strategy; honestly, it was just easier and more fun to write a positive piece than a risk-focused one.

I finally took a vacation, which gave me time to sit down and write something specifically about risks. Even when writing about risks, I approached it from an optimistic angle: not emphasizing the catastrophes themselves, but thinking about how we overcome them, how we develop a "battle plan" for addressing them.

One narrative framework I used comes from a scene in the film Contact. In the movie, humanity has discovered alien life, and an international committee is interviewing candidates to select a representative to make contact. One candidate is asked: if you could ask the aliens only one question, what would it be? The character responds: I'd ask, how did you do it? How did you survive your technological adolescence without destroying yourselves? How did you make it through?

I first watched this film about twenty years ago, and the question has stayed with me ever since. I use it as a metaphor: we now stand before a door to tremendous capability — the ability to essentially "make machines from sand." This path has been almost inevitable since humans first began using fire. But how we use this capability is not predetermined.

So in the coming years, we'll have to confront a series of questions: How do we control these highly autonomous systems that are already smarter than any human in many respects? How do we prevent their misuse by individuals? I have deep concerns about bioterrorism, for example. How do we prevent state-level misuse? What are the economic implications? I've spoken many times about labor displacement. And what are the problems we haven't even realized yet — which are often the hardest to address.

So I've been thinking systematically about how these risks can be addressed. This includes what we as company leaders need to do within our own organizations, what we can accomplish through collaboration with each other, and inevitably, the broader societal institutions — governments — that need to be involved.

But I feel a constant, intense sense of urgency. Every day, crazy things are happening in the world beyond AI. But to me, this is moving too fast and at too large a scale — we should almost be devoting our full attention to one question: how does humanity get through this?

Moderator: So now I'm not even sure what's more surprising to me: first, that you actually took a vacation; second, that you spent it thinking about AI risk; or third, that your entire essay centers on one question: can we get through the "technological adolescence" of this technology without destroying ourselves. But I really am looking forward to reading your next piece.

You've already mentioned several directions we could explore further. Let's start with employment, since you've been quite outspoken on this. I recall you saying that within one to five years, perhaps half of entry-level white-collar jobs could disappear.

But I want to pose this first to Demis — so far, we haven't really seen clear disruption in the labor market. Yes, unemployment in the United States has risen somewhat, but much of the economic research I've seen, and that we've published ourselves, suggests this is more a post-pandemic correction from over-hiring rather than AI-driven. One could even argue that many companies are hiring more to build out AI capabilities.

Do you think things will play out as economists have long argued — that there is no "lump of labor" fallacy, that new technology will ultimately create new jobs? At least the current evidence seems to point in that direction.

Demis: Yes, I think in the near term, that's indeed how it will play out. This largely follows the normal evolution path when a breakthrough technology emerges: some jobs will be disrupted, but new ones will be created — potentially more valuable, even more meaningful work.

I think we'll start seeing some effects this year, concentrated in junior roles, entry-level positions, intern-type roles. We're seeing some signs ourselves — hiring at these levels has slowed somewhat. But I think this can be fully offset by something else: the emergence of incredibly powerful creative tools now available to almost everyone, essentially for free.

If I were speaking to a group of undergraduates right now, I would strongly urge them to become as fluent as possible with these tools. In a way, even those of us building these systems are so busy constructing the systems themselves that we struggle to find time to fully explore them — even the current models and products have vast untapped capability, let alone future versions.

In my view, truly mastering these tools deeply could be more valuable than a traditional internship, allowing you to "leapfrog" more quickly into genuine usefulness within a specialized domain. That's roughly my prediction for what happens in the next five years.

Of course, we may still have some subtle differences on the timeline. But I think once true AGI arrives, that's an entirely different question: because then we enter genuinely uncharted territory.

Moderator: Do you feel now that this might take longer than you expected last year? You said at the time that perhaps half of white-collar jobs could disappear.

Dario: My view hasn't really changed. I actually agree with you and Demis — when I said that, the labor market wasn't being affected, and I wasn't claiming it was already happening.

But now, I think we may be starting to see very early signs, mainly in software and programming. Even within Anthropic, I can foresee a situation where at more junior levels, and some mid-level positions, we may need fewer people in the future rather than more. We're thinking about how to navigate this change in a thoughtful, measured way.

If you asked me six months ago, I would have stood by the "one to five years" timeframe. Connecting this to what I said earlier reveals an apparent inconsistency: we may have AI that surpasses humans in almost every dimension within one to two years, or slightly longer, yet labor market changes appear slower.

The reason is that there's a lag period, and the substitution process itself takes time. I certainly know the labor market has adaptive capacity. In the past, 80% of people worked in agriculture; then agriculture was automated, people became factory workers, then knowledge workers. So the market does have some adaptability, and we need to view this through a more sophisticated economic lens.

But what I genuinely worry about is that as this exponential change continues compounding — and I don't think this will take long, perhaps in that one-to-five-year window — it will eventually outpace our collective societal capacity to adapt.

Demis: I think we're actually saying the same thing, just isolating the difference in timeline. And ultimately, this disagreement still comes down to how quickly you can get that closed loop to actually run.

Moderator: How confident are you that governments truly understand the scale of this and have begun seriously thinking about what policy responses are needed?

Demis: I don't think the amount of work being done on this is adequate. Even at events like today, when I speak with economists, I'm often surprised that they haven't thought more, in a professional economist way, about what's coming next — beyond just the path to AGI itself.

Even if we handle all the technical issues Dario mentioned perfectly, job displacement is still just one problem. What concerns us is the structural economic transformation that follows. There may be ways to distribute the resulting productivity gains and new wealth more equitably, but I'm not sure we yet have the institutions in place to do so. In theory, it should approach something like a post-scarcity society.

But there are other problems that genuinely keep me up at night. These are bigger than economic issues — they concern meaning, purpose, and all the things we get from work that aren't purely economic. That's one problem. But strangely, I actually think this one might be easier to solve than "how the overall human condition will change."

I'm still optimistic that humanity will find new answers. Today, we're already doing many things, from extreme sports to artistic creation, that don't have direct economic returns as their goal. I believe we'll continue finding meaning in these directions, and even develop more mature, more sophisticated forms. Plus, I think we'll begin exploring interstellar space, and all of this will become part of our "sense of purpose."

Because of this, I think even by my relatively conservative timeline — say, five to ten years — that's actually not a very long time. We should start thinking seriously about these questions now.

Through AlphaFold, our scientific research, and companies we've incubated like Isomorphic, we're working to solve diseases, cure diseases, find new energy sources. From society's perspective, these are obviously things people want to see happen.

But the industry's current investment proportion in these directions is still not high enough.

— Demis

Moderator: How significant do you think the risk of public backlash against AI is? Could this backlash push governments into making decisions that you would consider irrational? I ask because I think back to the globalization period of the 1990s — there was indeed job displacement, but governments didn't respond adequately, which eventually triggered strong public backlash and led us step by step to where we are today.

Do you think there will be an increasingly strong political resentment directed at what you're doing, and at your companies themselves?

Demis: I do think that risk exists, and the concern itself is understandable. People will feel fear and anxiety about their jobs and livelihoods. I think the next few years will become very complex, both at the geopolitical level and with various real-world factors intertwining.

But at the same time, I think there's another side. For example, through AlphaFold, our scientific research, and companies we've incubated like Isomorphic, we're working to solve diseases, cure diseases, find new energy sources. From society's perspective, these are obviously things people want to see happen. But I also feel the industry's current investment proportion in these directions is still not high enough. We need more cases like AlphaFold — and I know Dario agrees with this — actually doing things that have clear, unambiguous positive value for the world.

And I think this isn't just about "talking about it" — it's that the industry, and those of us in core positions, have a responsibility to actually demonstrate these results, to prove them with facts rather than staying at the level of rhetoric. Of course, all of this will inevitably be accompanied by shocks and disruptions in other areas.

Another issue that can't be ignored is geopolitical competition. Not just competition between companies, but national-level competition primarily between China and the United States. If there isn't some form of international cooperation or consensus on this — which I think is essential, such as establishing minimum safety standards for deployment — things become much more complicated. I believe Dario would agree. This technology is transnational; it affects everyone, all of humanity.

By the way, Contact is also one of my favorite movies — funny coincidence, I didn't know you loved it too. But I do think these issues need to be seriously worked through and resolved step by step. If possible, even if AI development could proceed slightly slower than currently predicted, even slower than my own timeline estimate, just to give society time to work through these issues, I think that would be a good thing. But this clearly requires a considerable degree of coordination.

Dario: Yes, in this situation I would also lean more toward your timeline assessment —

Moderator: But Dario, we do need to expand on this topic. Because since our last conversation in Paris, if the geopolitical environment has changed in any direction, it's become more complicated, more insane — pick whatever word you like.

Secondly, the United States' attitude toward China has also shifted noticeably, more toward an unlimited, full-speed-ahead posture, yet simultaneously selling chips to China. This approach itself seems quite contradictory. The overall American stance is already very different; and at the geopolitical level, the relationship between the United States and Europe is currently in a rather delicate, even somewhat strange state.

Against this backdrop, when I hear you talk about international organizations along the lines of CERN, to be honest, that sounds very far from our present reality. Returning to practical matters, do you think geopolitical risk is indeed rising? If so, what do you think should be done? And from what we can see so far, government actions seem to be precisely the opposite of what you were advocating just now.

Dario: Yes, you know, we can only do our best. We're just one company after all, and can only operate within the environment we're given, however crazy it may be.

But at least at the policy level, my core assessment hasn't changed. And the most important point among them is: don't sell chips. This is probably one of the most critical measures we can take to buy time, to ensure we have space to properly handle all of this.

I said earlier too — I actually hope Demis's timeline assessment is right. I hope we really do have a 5 to 10 year buffer. And it's entirely possible he's right and I'm wrong. But let's assume I'm right, that this really could happen within one to two years — then why can't we actively slow down, to align with Demis's timeline?

The problem is, we can't. The reason is simple: because our geopolitical rivals are pushing ahead with the same technology at a comparable pace. In this situation, it's very difficult to reach a genuinely enforceable agreement for everyone to slow down together.

But if we can at least do one thing — not sell chips — then this is no longer a US-China competition problem. It's just a competition between me and Demis. And that, I'm very confident we could negotiate.

Moderator: Then how do you view the government's current logic? As I understand it, their thinking is: we sell them chips because we need to bind them into America's supply chain system.

Dario: I think this isn't just about the timescale — it's more fundamentally about the nature of this technology itself. If this were telecommunications or something similar, then the whole logic about promoting the American technology stack, building data centers using American chips around the world, ensuring these countries use NVIDIA rather than Huawei chips — that whole framework might make sense.

But that's not how I view this. To me, this is more like making a choice: do we sell nuclear weapons to North Korea just because it would bring Boeing some profit? And then we can say, these weapons were made by Boeing, America wins, what a wonderful thing — the analogy itself is sufficient to show how I view this tradeoff. I really don't think this makes logical sense.

And by comparison, we've taken many more aggressive measures against China and other actors in the past, but in my view, none of those have been as direct and effective as this single measure.

Moderator: Let me ask one final question, and hopefully leave some time for one or two questions from the audience. Another risk that doomsayers often worry about is AI with absolute power and malicious intent. I know both of you have remained relatively skeptical of this "doomsday" narrative.

But in the past year, we have seen some changes — these models have begun exhibiting capabilities for deception, concealment, duplicity. Has your view on this risk changed compared to a year ago? In the models' evolutionary trajectory, have there emerged signs that deserve more of our attention and vigilance?

Dario: Yes, actually from the very beginning of Anthropic, we've been thinking seriously about this category of risk. Initially our research was more theoretical. For example, we pioneered the direction of mechanistic interpretability, trying to "open up the inside of models" to understand why they make certain behaviors, in a sense like human neuroscientists — and both of us have neuroscience backgrounds — trying to understand how the brain works.

As time has progressed, we've increasingly documented and identified bad behaviors that models exhibit in certain situations, and begun trying to address and correct these issues using mechanistic interpretability approaches. So you could say I've always maintained a high level of attention to these risks, and have discussed this with Demis many times. I'm sure he takes it very seriously as well.

However, I have indeed remained skeptical of "doomsday" narratives — the view that we're destined to fail, that there's nothing we can do, or that this is the most likely outcome. I don't think so. I think this is a real risk, but if we work together, we can gradually understand, control, and guide the systems we're creating through scientific methods.

Of course, if we build these systems in a crude way, if we get caught in a reckless race, pushing too fast with no guardrails, then yes, I do think things could go wrong.

Moderator: Then let me also take this opportunity to raise the question slightly: in the past year, have you become more confident in the positive potential of this technology — whether in science, medicine, or the areas you've repeatedly mentioned? Or by comparison, have you become more concerned about the risks we just discussed?

Demis: I've been working in this field for over twenty years. We've actually always known — the reason I've devoted my entire career to AI is because of its enormous positive potential — it's essentially the ultimate tool for science and understanding the universe. I've been fascinated by this since childhood, and if built in the right way, AI should become the ultimate tool for achieving this goal.

Meanwhile, we've been thinking about these risks from the very beginning — at least for the fifteen years since DeepMind was founded. We foresaw early on that once we truly unlocked this positive potential, AI as a dual-use technology could also be repurposed by bad actors for harmful ends. So this has always been something we need to keep in mind throughout.

But I've always had tremendous faith in human ingenuity. The real question is: do we have enough time, enough focus, and can we get the best people working together to solve these problems? If we have those things, I believe the technical risks can be addressed.

Conversely, if we don't have that time and space, things become more dangerous — the process becomes fragmented, with different projects and teams racing against each other, and in that scenario, ensuring the systems we build are technically safe becomes much harder. But in my view, as long as we have enough time and room, this remains a highly solvable, controllable problem.

Moderator: Let's take one final question from the audience.

Philip: Thank you very much. I'm Philip, co-founder of StarCloud — we're building data centers in space. I want to ask a slightly philosophical question. For me, one of the strongest arguments supporting "doomerism" is the Fermi paradox — the fact that we don't see other intelligent life in our galaxy. I'm curious whether you two have any thoughts on this.

Demis: I've actually thought about this quite a lot. I don't think that's the reason. Because if it were, we should be seeing traces of AI by now. Let me quickly explain the idea: if the Fermi paradox exists because extraterrestrial civilizations are destroyed by their own technology, then we should see "PaperClip factories" flying toward us from some corner of the galaxy, right? But that's not what we see. We don't see any structures like that, nor Dyson spheres — whether AI-created, naturally formed, or left behind by biological civilizations.

So in my view, the Fermi paradox must have some other explanation. I have some theories of my own, but I certainly can't get through them in the next minute. My intuition is that we may have already crossed the so-called "Great Filter." If I had to guess, I'd say that filter likely occurred at the stage of multicellular life — the evolutionary leap to that point was itself extraordinarily difficult.

So where we are now, there is no predetermined outcome of "what happens next." I think what comes next is something that humans themselves will have to write. This could have been a fascinating discussion, but it's clearly beyond what we can cover in the next few minutes.

Moderator: For this last question, just 15 seconds each — when we meet again, hopefully next year, and hopefully with the three of us, what do you think will have changed by then?

Dario: I think the most important thing to watch is how the development of AI systems building AI systems unfolds. Which direction that path takes will determine whether we still have a few years before the finish line, or whether we'll suddenly be facing a situation that is both miraculous and extremely urgent, requiring immediate response.

In short: AI building AI systems.

Demis: I agree with that, so we're closely following developments in that area. At the same time, I think there are other equally important directions being actively researched, such as world models and continual learning. These are problems that must be cracked. If self-improvement alone doesn't bring breakthroughs, we'll need these other capabilities to truly deliver. Additionally, I think robotics may be heading toward its own breakthrough moment.

Moderator: Though based on everything you've just said, perhaps all of us should hope this happens a bit more slowly — including yourselves, and everyone else. Honestly, I would prefer that; I think it would be better for the world. Of course, you do have some ability to influence this to a certain degree.

Thank you both very much.

(Click "read more" to watch the full video)