The Threshold: Dario Amodei and Demis Hassabis on the Edge of AGI

In early 2026, two men sat on stage to discuss whether humanity would survive its own ingenuity. This wasn’t hyperbole or clickbait. Dario Amodei, CEO of Anthropic, had just spent his vacation writing an essay about existential risk framed around a scene from Contact: “How did you do it? How did you manage to get through this technological adolescence without destroying yourselves?”

Demis Hassabis, co-founder of Google DeepMind, had spent two decades building toward artificial general intelligence — AI that can do everything humans can do — and was now racing to ensure it didn’t arrive before humanity was ready. His timeline: 50% chance by the end of the decade. Amodei’s timeline: possibly 12 to 24 months.

The gap between those predictions matters less than what both men agreed on: we are approaching a threshold unlike any in human history. On one side lies the potential to cure all disease, understand the universe, and enter what Amodei calls a “post-scarcity world.” On the other side lies the risk of catastrophic misuse, loss of human agency, and civilizational disruption we lack the institutions to manage.

The conversation that unfolded — moderated by a journalist who noted this was like “chairing a conversation between the Beatles and the Rolling Stones” — revealed not just technical timelines but a deeper tension: between the inexorable logic of exponential progress and the desperate human need for more time. Between competition that drives innovation and cooperation that might save us. Between what we can build and what we should build.

A year earlier, these two had sat on a comically small loveseat in Paris discussing similar themes. Now, in early 2026, with AI systems already writing most of the code at major companies, the timeline had compressed. The question was no longer whether AGI would arrive, but whether we’d have our shit together when it did.

The moderator opened with the timeline question, directing it at Amodei. Last year in Paris, he’d predicted models capable of performing at Nobel laureate level across many fields by 2026–2027. It was now 2026. Did he stand by it?

“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,” Amodei replied. The mechanism he’d envisioned was already unfolding: models good at coding and AI research would produce the next generation of models, creating a self-reinforcing loop that would accelerate development.

“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.’” He paused, calculating. “I think we might be six to 12 months away from when the model is doing most — maybe all — of what those engineers do end to end.”

The question then became how fast the loop closes. “Not every part of that loop is something that can be sped up by AI,” Amodei noted. Chip manufacturing, training time, physical constraints — these remain bottlenecks. “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.”

Then came the critical admission: “If I had to guess, I would guess that this goes faster than people imagine. That key element of code and increasingly research going faster than we imagine — that’s going to be the key driver.”

The moderator turned to Hassabis, who’d been more cautious last year. He’d given a 50% chance of AGI-level capabilities by decade’s end. Did he still stand by that?

“I think I’m still on the same kind of timeline,” Hassabis said carefully. “There has been remarkable progress, but I think some areas of engineering work, coding, mathematics — these are a little bit easier to see how they would be automated, partly because they’re verifiable. You know what the output is.”

Natural sciences, he noted, are harder. “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.”

More fundamentally, there were missing capabilities. “Not just solving existing conjectures or existing problems, but actually coming up with the question in the first place. Coming up with the theory or the hypothesis. That’s much harder. That’s the highest level of scientific creativity, and it’s not clear we will have those systems.”

The disagreement wasn’t about whether AGI would arrive — both were certain it would. The question was whether it would take one year or five, and that difference contained everything: time to prepare institutions, time to solve safety problems, time to understand what we’d built before deploying it at scale.

The past year had scrambled the competitive landscape in ways both men were still processing. Twelve months earlier, DeepSeek’s release from China had sent shockwaves through the industry. Google DeepMind had seemed to be lagging OpenAI. Now the tables had turned dramatically.

“We were always very confident we would get back to the top of the leaderboards,” Hassabis said, “because I think we’ve always had the deepest and broadest research bench. It was about marshalling that all together and getting the intensity and focus and startup mentality back to the whole organization.”

The results spoke for themselves: Gemini had reclaimed leadership positions. The product was gaining market share. “We’re making great progress, but there’s a ton more work to do. We’re bringing to bear Google DeepMind as the engine room of Google, getting used to shipping our models more quickly into product surfaces.”

For Amodei, the competitive picture looked different. Anthropic remained independent — not part of a tech giant — and that independence raised questions about survival. Could an independent AI company last long enough to reach profitability before the capital requirements became crushing?

His answer came through numbers. Anthropic’s revenue had grown 10x year-over-year for three consecutive years: zero to $100 million in 2023, $100 million to $1 billion in 2024, $1 billion to $10 billion in 2025.

“Those revenue numbers — I don’t know if that curve will literally continue, it would be crazy if it did — but those numbers are starting to get not too far from the scale of the largest companies in the world.” He acknowledged the uncertainty. “We’re trying to bootstrap this from nothing. It’s a crazy thing. But I have confidence that if we’re able to produce the best models in the things we focus on, things will go well.”

Then he offered an observation that implicitly excluded others from the race: “I think it’s been a good year for both Google and Anthropic. The thing we actually have in common is we’re both companies — or the research part of the company — that are led by researchers who focus on the models, who focus on solving important problems in the world, who have these hard scientific problems as a north star. Those are the kind of companies that are going to succeed going forward.”

The moderator resisted asking what would happen to companies not led by researchers. But the implication hung in the air: as AI development became more complex, the companies with the deepest scientific culture would pull ahead. Hype and marketing would matter less than mechanistic interpretability and understanding what you’d actually built.

The conversation shifted to the central technical question: closing the loop. Could AI systems become good enough at AI research to improve themselves without human intervention? This was the crux of the winner-takes-all versus normal-technology debate.

“I definitely don’t think it’s going to be a normal technology,” Hassabis said. “There are aspects already where it’s helping with our coding and some aspects of research. The full closing of the loop, though, I think is an unknown.”

He outlined the challenge: in domains where you can quickly verify answers — coding, mathematics — self-improvement seemed feasible. But messier domains presented problems. “As soon as you start getting into NP-hard domains, physical AI, robotics — you’ve got hardware in the loop that may limit how fast self-improvement systems can work.”

This raised a deeper question: “What is the limit of engineering and math to solve the natural sciences?” Could pure computational power crack problems that required experimental verification? Or would there always be a physical bottleneck, a need to test hypotheses in the real world?

Amodei had spent the previous year thinking about these questions while writing what would become a controversial pair of essays. The first, “Machines of Loving Grace,” had painted an optimistic picture of AI’s potential. He’d written it, he admitted, because “the positive essay was easier and more fun to write than the negative essay.”

But on vacation — because apparently even Amodei’s vacations involve existential risk assessment — he’d finally written the second essay. It was framed around that scene from Contact, the question about surviving technological adolescence.

“We are knocking on the door of these incredible capabilities,” he said. “The ability to build basically machines out of sand. I think it was inevitable that the instant we started working with fire…”

He trailed off, then continued: “But how we handle it is not inevitable. 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? How do we make sure individuals don’t misuse them? I have worries about things like bioterrorism. How do we make sure nation-states don’t misuse them? That’s why I’ve been so concerned about the CCP, other authoritarian governments.”

The list continued: economic impacts, labor displacement, and crucially, “what haven’t we thought of — which in many cases may be the hardest thing to deal with at all.”

For each risk, addressing it would require a mixture of actions: what they could do individually as company leaders, what they could do working together, and what would require “wider societal institutions like the government.”

Then came the statement that made the moderator’s head spin: “I just feel this urgency that every day there’s all kinds of crazy stuff going on in the outside world, outside AI. But my view is this is happening so fast and is such a crisis we should be devoting almost all of our effort to thinking about how to get through this.”

The conversation turned to jobs — the most immediate and politically volatile consequence of AI advancement. Amodei had been vocal about this, predicting that half of entry-level white collar jobs could vanish within one to five years. Yet so far, the labor market showed no discernible AI-driven displacement.

The moderator put the question to Hassabis: wouldn’t this be like every previous wave of automation, where new jobs emerged to replace disrupted ones?

“In the near term, that is what will happen,” Hassabis agreed. “The normal evolution when a breakthrough technology arrives. Some jobs get disrupted, but new, even more valuable, perhaps more meaningful jobs get created.”

He saw early signs of impact at the junior level: internships, entry-level positions. “I think there is some evidence — I can feel it ourselves — of maybe a slowdown in hiring. 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.”

His advice to undergraduates: “Get really unbelievably proficient with these tools. I think that can be maybe better than a traditional internship would have been, in terms of leapfrogging yourself to be useful in a profession.”

But he drew a sharp line: “That’s what I see happening probably in the next five years. What happens after AGI arrives — that’s a different question. We would be in uncharted territory at that point.”

Amodei’s view hadn’t changed from six months earlier when he’d made his prediction. “At the time I made the comment, there was no impact on the labor market. I wasn’t saying there was an impact at that moment.” But now, in early 2026, he could see the beginnings.

“Even within Anthropic, I can look forward to a time where on the more junior end and then on the more intermediate end, we actually need less and not more people. We’re thinking about how to deal with that within Anthropic in a sensible way.”

He stood by his one-to-five-year timeline. The lag existed because replacing human workers takes time even when the technology is ready. “The labor market is adaptable. Eighty percent of people used to do farming. Farming got automated and they became factory workers and then knowledge workers.”

But here was the concern: “My worry is as this exponential keeps compounding — and I don’t think it’s going to take that long, somewhere between a year and five years — it will overwhelm our ability to adapt.”

The difference between the two men’s views on labor displacement, like their difference on AGI timelines, came down to one variable: how fast the loop closes on code generation. If AI systems rapidly become capable of doing their own AI research, Amodei’s timeline holds. If key capabilities remain elusive, Hassabis’s more measured pace prevails.

“I don’t think anywhere near enough work is going on about this,” Hassabis said when asked about government preparedness. “I’m constantly surprised even when I meet economists at places like this that there aren’t more professional economists thinking about what happens.”

The problem wasn’t just economic. “Even if we get all the technical things right that Dario is talking about, and the job displacement is one question we’re worried about — the economics of that — maybe there are ways to distribute this new productivity, this new wealth more fairly. I don’t know if we have the right institutions to do that.”

Then he went deeper: “There are even bigger questions than that to do with meaning and purpose. A lot of the things we get from our jobs, not just economically — that’s one question, but that may be easier to solve strangely — than what happens to the human condition and humanity as a whole.”

He remained optimistic that solutions existed. “We do a lot of things today from extreme sports to art that aren’t necessarily directly to do with economic gain. I think we will find meaning, and maybe there’ll be even more sophisticated versions of those activities. Plus I think we’ll be exploring the stars. There’ll be all of that to factor in as well for purpose.”

But the timeline worried him: “Even on my timelines of five to 10 years away, that isn’t a lot of time before this comes.”

The moderator raised the specter of political backlash. In the 1990s, globalization had displaced workers, governments hadn’t responded adequately, and the resulting resentment had reshaped politics globally. Could AI trigger something similar, leading to “stupid” policy responses that would hamper development?

“I think there’s definitely a risk,” Hassabis acknowledged. “That’s kind of reasonable. There’s fear and worries about these things like jobs and livelihoods.”

His prescription focused on demonstrating value: “We want to, and we’re trying to do this with AlphaFold and our science work and Isomorphic — our spinout company — solve all disease, cure diseases, come up with new energy sources. As a society, it’s clear we’d want that. Maybe the balance of what the industry is doing is not enough towards those types of activities. I think we should have a lot more examples of AlphaFold-like things that are unequivocal good in the world.”

He emphasized: “It’s incumbent on the industry and all of us leading players to show that more, demonstrate that, not just talk about it.”

But good intentions collided with geopolitical reality. “The other issue is the geopolitical competition between the companies but also US and China primarily. Unless there’s international cooperation or understanding around this — which I think would be good actually in terms of things like minimum safety standards for deployment — I think it’s vitally needed. This technology is going to be cross-border. It’s going to affect everyone. It’s going to affect all of humanity.”

Then came a striking admission: “Maybe it would be good to have a slightly slower pace than we’re currently predicting, even on my timelines, so that we can get this right as a society. But that would require some coordination.”

The moderator noted the irony: “I prefer your timelines.”

“Yes, I will concede,” Hassabis replied with a slight smile.

Since the Paris conversation, the geopolitical landscape had, if anything, deteriorated. A new US administration had adopted a no-holds-barred approach toward AI competition with China while simultaneously selling chips to China. Europe’s relationship with the United States had become strained. Against this backdrop, Hassabis’s vision of a CERN-like international collaboration seemed impossibly distant from reality.

“We’re just trying to do the best we can,” Amodei said when asked about operating in this environment. “We’re just one company, trying to operate in the environment that exists, no matter how crazy it is.”

But his policy recommendation hadn’t changed: “Not selling chips is one of the biggest things we can do to make sure we have the time to handle this.”

He returned to his preferred timeline argument: “I said before, I prefer Demis’s timeline. I wish we had five to 10 years. It’s possible he’s just right and I’m just wrong. But assume I’m right and it can be done in one to two years. Why can’t we slow down to Demis’s timeline?”

The answer: geopolitical competition. “The reason we can’t do that is because we have geopolitical adversaries building the same technology at a similar pace. It’s very hard to have an enforceable agreement where they slow down and we slow down.”

Here was the logic: “If we can just not sell the chips, then this isn’t a question of competition between the US and China. This is a question of competition between me and Demis, which I’m very confident we can work out.”

The moderator pressed on the administration’s apparent logic: binding China into US supply chains by selling them chips. Amodei’s response revealed how seriously he took the stakes.

“I think it’s a question not just of timescale but of the significance of the technology. If this was telecom or something, then all this stuff about proliferating the US stack and wanting to build chips around the world to make sure random countries build data centers that have Nvidia chips instead of Huawei chips…”

He paused, then offered an analogy so stark it silenced the room: “I think of this more as: are we going to sell nuclear weapons to North Korea because that produces some profit for Boeing? Where we can say, ‘Yeah, these cases were made by Boeing, the US is winning, this is great.’”

The comparison landed. “That analogy should just make clear how I see this trade-off. I just don’t think it makes sense. We’ve done a lot more aggressive stuff toward China and other players that I think is much less effective than this one measure.”

The final area of risk the moderator raised was the one that animates doomer scenarios: malign AI, systems capable of deception that slip human control. Had the past year changed how either man thought about this?

“Since the beginning of Anthropic, we’ve thought about this risk,” Amodei said. “Our research at the beginning was very theoretical. We pioneered this idea of mechanistic interpretability — looking inside the model, trying to understand why it does what it does, like human neuroscientists try to understand the brain.”

Both he and Hassabis had backgrounds in neuroscience, and that perspective shaped their approach. “As time has gone on, we’ve increasingly documented the bad behaviors of the models when they emerge and are now working on trying to address them with mechanistic interpretability.”

He drew a distinction: “I’ve always been concerned about these risks. I think Demis has also been concerned. But I have been — and I would guess Demis as well — skeptical of doomerism, which is ‘we’re doomed, there’s nothing we can do,’ or this is the most likely outcome.”

Instead: “I think this is a risk that if we all work together we can address. We can learn through science to properly control and direct these creations that we’re building. But if we build them poorly, if we go so fast that there’s no guardrails, then I think there is risk of something going wrong.”

The moderator turned to Hassabis with a broader question: over the past year, had he grown more confident about the upside potential or more worried about the risks?

“Look, I’ve been working on this for 20-plus years,” Hassabis replied. “We already knew — 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 been obsessed with that since I was a kid. Building AI should be the ultimate tool for that if we do it in the right way.”

But: “The risks also we’ve been thinking about since the start — at least the start of DeepMind 15 years ago. We foresaw that if you got the upsides, it’s a dual-purpose technology, so it could be repurposed by bad actors for harmful ends. We’ve needed to think about that all the way through.”

His confidence came from a specific belief: “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.”

The caveat: “It may be we don’t have that, and then that will introduce risk because it’ll be fragmented. There’ll be different projects and people racing each other. Then it’s much harder to make sure these systems we produce will be technically safe.”

But the technical safety problem? “I feel like that’s a very tractable problem if you have the time.”

In the final two minutes, a questioner from the audience — co-founder of a company building data centers in space — raised the Fermi Paradox. If advanced civilizations commonly destroy themselves with their own technology, shouldn’t we see evidence of that across the galaxy?

Hassabis fielded it quickly: “That can’t be the reason, because we should see all the AIs. We should be seeing paperclips coming towards us from some part of the galaxy. Apparently we don’t see any structures, Dyson spheres, nothing, whether they’re AI or biological.”

His theory: “We’re past the great filter. It probably was multicellular life, if I would have to guess, was incredibly hard for biology to evolve. So we’re on — there isn’t a comfort of knowing what’s going to happen next. I think it’s for us to write as humanity what’s going to happen next.”

The moderator, with seconds remaining, asked each for a prediction: when they met again next year, what would have changed?

“The biggest thing to watch is this issue of AI systems building AI systems,” Amodei said. “Whether that goes one way or another will determine whether it’s a few more years until we get there or if we have wonders and a great emergency in front of us that we have to face.”

“I agree on that,” Hassabis said. “We’re keeping in close touch about that. But outside of that, I think there are other interesting ideas being researched like world models, continual learning. These are things that will need to be cracked if self-improvement doesn’t deliver the goods on its own. And I think robotics may have its breakout moment.”

The moderator offered a final thought: “Maybe on the basis of what you’ve just said, we should all be hoping that it does take you a little bit longer.”

“I would prefer that,” Hassabis said immediately. “I think that would be better for the world.”

“But you guys could do something about that,” the moderator noted.

The implication hung in the air: these two men, leading the organizations building the most powerful AI systems in the world, had just spent 30 minutes explaining why they wished they had more time. Yet the logic of competition — between companies, between nations — meant neither could simply choose to slow down.

They could regulate chips. They could share safety research. They could try to coordinate international standards. But absent unprecedented cooperation, the exponential would continue compounding. The loop would close or it wouldn’t. AGI would arrive in one year or five.

And humanity would discover whether it had written its future carefully enough to survive it.

The conversation ended to applause, but the questions it raised echo beyond any single stage or moment. Two of the people best positioned to understand what’s coming had just outlined a scenario where:

• AI systems capable of doing all human cognitive work may arrive in 1–5 years
• This will trigger labor displacement at a scale that could overwhelm societal adaptation
• Geopolitical competition makes coordination difficult even when everyone agrees it’s needed
• The technical safety problems are solvable given time, but time is the one thing competition denies us
• The upside — curing disease, understanding the universe, post-scarcity abundance — is extraordinary
• The downside — misuse, loss of control, civilizational disruption — is existential
• And despite understanding all of this, the people building these systems feel compelled to continue

Perhaps the most revealing moment came when Hassabis admitted he’d prefer his own slower timeline and that it would be better for the world. Not “might be better” or “could be better.” Would be better.

Yet neither man suggested stopping. Neither proposed a moratorium. Neither argued for pause. They argued for chip export controls. For safety research. For international cooperation on minimum standards. For demonstrating beneficial applications. For transparency about risks.

But not for stopping.

The logic is simultaneously compelling and terrifying: if we don’t build it, someone else will. If we don’t build it fast, someone less careful will get there first. If we don’t compete, authoritarian governments will control the most powerful technology ever created.

So we race toward a threshold we’re not ready to cross, hoping we can solve the safety problems fast enough, hoping institutions can adapt quickly enough, hoping humanity proves clever enough to survive its own ingenuity.

Dario Amodei framed his essay around a question from Contact: “How did you do it? How did you manage to get through this technological adolescence without destroying yourselves?”

In early 2026, standing on the edge of AGI, the honest answer is: we don’t know yet. We’re still in the middle of finding out.

And the two people who understand the technology best just told us they wish we had more time to figure it out — but the exponential waits for no one.