At the World Economic Forum Annual Meeting 2026 in Davos, Anthropic CEO Dario Amodei and Google DeepMind CEO Demis Hassabis reunited on stage for a wide-ranging conversation titled “The Day After AGI.” Moderated by Zanny Minton Beddoes, editor-in-chief at The Economist, the discussion picked up where the two left off last year — but with noticeably higher stakes.

What followed was not a debate about whether artificial general intelligence (AGI) is coming, but how fast, what will drive it, and whether society can adapt in time.

The conversation opened with a return to a question both leaders have been asked repeatedly over the past year: how close is artificial general intelligence, and how confident should we be in any timeline?

Dario Amodei argued that progress is already moving faster than many people expect, particularly in areas like coding and AI research itself. While he acknowledged uncertainty, he said it is becoming increasingly difficult to imagine AGI taking much longer than a few years.

“It’s always hard to know exactly when something will happen,” Amodei said, “but I don’t think that’s going to turn out to be that far off.”

He described a mechanism he believes could accelerate timelines further: AI systems that are good enough at coding and research to meaningfully help build the next generation of models. According to Amodei, that dynamic is no longer theoretical. He suggested that models could be handling most software development tasks end-to-end within the next 6–12 months — a shift that would further accelerate development cycles. He pointed to concrete changes already underway inside Anthropic.

“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,’” he said.

If those capabilities continue to improve, Amodei suggested, AI systems could increasingly reduce human bottlenecks in development. That compression, he argued, could push AGI timelines closer than many institutional planning models assume. While factors like hardware, manufacturing, and training time still impose limits, he argued that the direction of travel is clear and likely faster than most people imagine.

Demis Hassabis, by contrast, reaffirmed a more cautious stance. While he acknowledged dramatic progress over the past year — especially in coding and mathematics — he argued that those gains do not automatically translate to the kinds of breakthroughs required for AGI across all domains. Hassabis reiterated his view that there is roughly a 50 percent chance of systems exhibiting human-level cognitive capabilities by the end of the decade, but emphasized that significant uncertainties remain.

“Some areas of engineering work, coding or mathematics are a little bit easier to see how they would be automated partly because they’re verifiable what the output is,” Hassabis said, referring to tasks where success can be quickly measured.

In particular, Hassabis stressed that scientific discovery is constrained by physical experimentation and slow validation cycles, not just reasoning speed. In fields like biology, chemistry, and physics, progress depends on testing ideas in the real world, observing results, and revising theories based on incomplete or ambiguous data. Those feedback loops can take months or years and cannot be easily compressed by compute alone.

He also drew a distinction between solving well-defined problems and generating the underlying questions that drive major scientific advances. While today’s AI systems excel at optimization and pattern recognition, Hassabis argued that identifying which questions are worth asking — and proposing new theoretical frameworks — remains far harder to automate.

“Actually coming up with the question in the first place, or coming up with the theory or the hypothesis,” Hassabis said, “I think that’s much, much harder.”

The disagreement, then, was not about whether AGI is possible, but about which bottlenecks matter most. Amodei focused on the accelerating impact of AI-assisted research and coding, while Hassabis pointed to unresolved challenges in theory formation, experimental design, and real-world validation. Whether current systems can cross those remaining scientific thresholds as quickly as they have surpassed software benchmarks remains an open question — and a central uncertainty shaping the timeline debate.

Despite their differences on timing, both leaders repeatedly returned to what emerged as the most consequential question of the discussion: can AI systems substantially accelerate their own improvement by helping to build the next generation of AI?

Amodei described this process as a self-reinforcing development cycle. As models become better at tasks like writing code and assisting with AI research, they can shorten the time it takes to design, train, and refine new models. Those improved models can then contribute even more to development, creating a feedback effect that compresses timelines.

In Amodei’s view, this is no longer a hypothetical mechanism. He argued that parts of this cycle are already visible, particularly in software development, where AI systems increasingly handle large portions of the work that once required teams of human engineers. If that trend continues, he suggested, AI could remove many of the human bottlenecks that traditionally slow progress.

Hassabis largely agreed — with an important caveat. He distinguished between partial loop-closing and full loop-closing. In areas like coding and mathematics, he said, AI systems can plausibly improve themselves because success is quickly verifiable: code either works or it doesn’t, proofs either hold or they don’t. That makes it easier for AI to contribute meaningfully to its own advancement.

However, Hassabis argued that extending this loop across all domains remains uncertain. In fields involving physical systems — such as robotics, hardware, or experimental science — progress depends on real-world testing, slow feedback, and constraints that AI cannot yet bypass. Manufacturing chips, training large models, running experiments, and integrating systems into the physical world all impose limits that software alone cannot eliminate.

As the session drew to a close, both leaders were asked what signal to watch to understand whether AGI arrives sooner or later.

Amodei summarized it succinctly: “AI systems building AI systems.”
Hassabis agreed: “I agree on that.”

Whether this self-improvement cycle remains partial or becomes broadly self-sustaining may determine whether AGI arrives in a few years or stretches further into the decade — and how much time society has to adapt.

When the conversation turned to labor, the moderator noted that economy-wide indicators have not yet shown a sustained rise in unemployment that economists can clearly isolate as being driven primarily by AI. That does not mean workers have not already been displaced.

Rather, early AI-related disruption is showing up unevenly through layoffs concentrated in specific sectors and companies, slower hiring, role consolidation, and workers moving between jobs. Historically, these kinds of shifts tend to appear before broader labor-market effects show up in aggregate unemployment data. Hassabis largely agreed with that narrow assessment, with an important caveat.

“In the near term, that is what will happen — the kind of normal evolution when a breakthrough technology arrives,” Hassabis said.

However, he argued that this does not mean disruption is absent. Instead, he suggested that pressure is likely to surface first in junior and entry-level positions — the very roles that traditionally serve as on-ramps into careers.

“I think we’re going to see this year the beginnings of maybe impacting junior-level, entry-level jobs — internships, this type of thing,” he said.

Hassabis did not frame this as a reason for young workers to disengage. Instead, he suggested that early-career learning itself may shift. With access to advanced AI tools, individuals could acquire skills, experiment, and receive feedback faster than in many traditional entry-level roles — effectively using AI systems as a kind of informal, self-directed apprenticeship.

The tradeoff, however, is that this learning increasingly happens outside paid employment. Hassabis did not point to any existing systems built by companies or institutions to replace entry-level roles. Rather, the burden shifts to individuals to invest time in learning AI tools on their own, often without income. The hope is that this self-directed learning can help them qualify for higher-level work. That opportunity depends heavily on access, education, and personal resources — and it does not replace the financial stability, mentorship, or structure that entry-level roles have traditionally provided.

Amodei agreed that labor markets have historically adapted to technological change, but he warned that the pace of AI progress could strain that adaptive capacity. His concern was not immediate mass unemployment, but timing — specifically, whether new roles and training pathways can emerge quickly enough to replace the ones being disrupted.

“There’s this lag and there’s this replacement thing,” he said.

In past transitions, new types of work typically appeared as older roles faded, giving workers and institutions time to adjust. Amodei warned that if AI capabilities continue to compound rapidly, that adjustment window could shrink — leaving gaps in employment, fewer entry points for new workers, and insufficient time for education systems and employers to respond.

On safety, both leaders rejected fatalistic “doomer” narratives — but neither minimized the seriousness of the risks created by increasingly autonomous and capable systems.

Amodei emphasized that concerns about control, misuse, and unintended behavior have shaped Anthropic’s work from its earliest days. His focus was not on speculative scenarios, but on practical failure modes that could emerge if development outpaces safeguards.

“I’m not a doomer,” he said, “but if we go so fast that there’s no guardrails, then I think there is risk of something going wrong.”

He outlined a range of concrete concerns, from individual misuse to large-scale threats involving nation states. Among them were risks such as bioterrorism, the misuse of AI by authoritarian governments, and the challenge of maintaining control over systems that may operate with a high degree of autonomy.

“How do we keep these systems under control that are highly autonomous and smarter than any human?” Amodei asked.

Hassabis framed the issue similarly, describing advanced AI as a dual-use technology: the same systems capable of accelerating scientific discovery can also be repurposed for harm. He argued that many technical safety challenges — such as managing misuse, limiting harmful behavior, and improving system reliability — are likely solvable in principle.

The constraint, he said, is not whether those problems can be solved, but whether the field has the time and coordination needed to solve them before deployment accelerates. Addressing safety effectively requires sustained collaboration across organizations, shared standards, and the ability to slow down when necessary.

“It may be we don’t have that,” Hassabis warned.

In a highly competitive environment, he explained, companies and countries may feel pressure to move faster rather than more carefully. That kind of race makes it harder to test systems thoroughly, share safety research, or align on common standards — increasing risk not because safety is impossible, but because competition, speed, and fragmentation work against it.

The conversation sharpened when it turned to geopolitics — not because the leaders disagreed, but because both acknowledged how little room there may be to slow AI development even when doing so would be preferable.

Amodei was explicit that unilateral restraint becomes unrealistic in a world where geopolitical rivals are pursuing similar capabilities. In that context, he argued that export controls on advanced chips are not about economic competition, but about buying time.

He rejected the idea that selling advanced AI infrastructure to adversaries is justified by market logic or supply-chain influence, framing the decision instead as a matter of existential risk.

“Are we going to sell nuclear weapons to North Korea,” Amodei asked, “because that produces some profit?”

The analogy underscored how he views the stakes: not as a normal technology trade-off, but as a decision that directly affects how much time the world has to prepare for systems that could rapidly surpass human control.

Hassabis echoed the concern, emphasizing that AI is inherently cross-border and will affect all of humanity. In principle, he argued, this creates a strong case for international coordination around safety standards and deployment. In practice, however, geopolitical competition makes that coordination extremely difficult.

Both leaders acknowledged the tension directly. While they expressed a preference for slower timelines and more deliberate deployment, neither suggested that current global dynamics make that outcome likely.

The result is an uncomfortable reality: even the people most aware of the risks feel constrained by competition between nations. The challenge is not simply a lack of goodwill or awareness, but a system that rewards speed over caution — even when caution would better serve the long-term interests of humanity.

Beyond economics and safety, Hassabis raised what he described as a deeper and potentially more difficult challenge: how societies preserve meaning and purpose if work no longer plays the central role it has for generations.

He acknowledged that much of the AI debate focuses on productivity gains and economic redistribution — questions he suggested may be solvable in principle through policy, new institutions, or post-scarcity models. But Hassabis argued that material security alone does not address what people derive from work beyond income.

Jobs, he noted, provide identity, daily structure, and a sense of contribution — elements that are far harder to replace than wages. In that sense, the human challenge may ultimately prove more difficult than the economic one.

When discussing what could fill that gap, Hassabis pointed to activities that already exist outside traditional labor markets, such as art, exploration, physical pursuits, and creative expression. The implication was not that everyone must become an artist, but that societies may need to broaden how they define contribution and value beyond paid employment.

“I’m a big believer in human ingenuity,” he said.

What concerned him most was timing. On the timelines being discussed, he warned that society may have only five to ten years to begin grappling seriously with these questions — making this not a distant philosophical issue, but a near-term societal one.

Hassabis did not propose concrete solutions, nor did he suggest that technologists alone should define the answers. Instead, he framed the question of meaning in an AI-shaped world as a collective responsibility — one that will require input from educators, policymakers, employers, and citizens before AI-driven change becomes unavoidable.

In previewing an upcoming essay on AI risk, Amodei framed the current moment as a test of whether humanity can safely navigate a period of rapid technological acceleration without causing irreversible harm.

He described today’s AI transition as a kind of “technological adolescence” — a stage in which societies acquire extremely powerful tools before they have fully developed the institutions, norms, and safeguards needed to manage them responsibly. The danger, he suggested, is not the technology itself, but the mismatch between capability and maturity.

To illustrate the stakes, Amodei referenced the film Contact, in which future civilizations look back and ask whether earlier societies were able to survive their own breakthroughs without self-destructing.

“How did you manage to get through this technological adolescence without destroying yourselves?” he asked, posing it as the question humanity may one day have to answer.

For Amodei, this was not an abstract thought experiment. He argued that the pace of AI development is so rapid that, despite other global crises, societies may have far less time than they assume to put durable safeguards, coordination mechanisms, and norms in place to navigate this transition safely.

Taken together, the conversation suggested that the defining question is no longer whether AI will reshape the world, but whether humanity can manage and shape what comes next — before the window to do so closes.

Q: Do Amodei and Hassabis agree on when AGI will arrive?
A: Not exactly. Amodei believes current trends suggest AGI could arrive sooner than many expect, driven by rapid progress in coding and AI-assisted research. Hassabis is more cautious, arguing that key aspects of intelligence — such as scientific creativity and experimental validation — remain harder to automate and could extend timelines.

Q: Why is coding such a central focus in the AGI debate?
A: Both leaders see coding as an early accelerant because results are relatively easy to verify. Amodei cited internal use where engineers increasingly rely on AI-generated code, while Hassabis noted that verifiability makes some domains easier to automate than others.

Q: What does “AI systems building AI systems” actually mean?
A: It refers to AI models becoming capable enough to meaningfully assist in developing the next generation of AI — writing code, optimizing architectures, or contributing to research. Both leaders agreed this loop is the most important signal to watch, as it could dramatically compress development timelines.

Q: Are AI-driven job losses already happening?
A: Both said broad labor market impacts are not yet clearly visible. Hassabis suggested early effects may appear in entry-level roles, while Amodei emphasized a lag between capability and displacement, followed by the risk of faster disruption if progress compounds.

Q: How worried are they about AI risks?
A: Both leaders rejected the idea that catastrophic outcomes are inevitable. However, they agreed that rapid progress without sufficient safeguards increases risk, particularly in a competitive global environment where coordination is difficult.

This Davos conversation matters because it shows that the people building the most advanced AI systems are no longer debating distant futures — they are debating how little time society may have to adjust.

The most important signal is not the exact AGI timeline, but the compression of uncertainty. When technology moves faster than institutions, businesses, and individuals, the result is not just disruption — it’s misalignment. Laws lag behind reality. Schools train people for roles that are already changing. Companies make long-term plans based on assumptions that may no longer hold. The risk is not that AGI arrives suddenly, but that it arrives into systems that cannot respond quickly enough, creating gaps in oversight, preparedness, and trust.

This is why the discussion about AI improving itself matters — not as a technical milestone, but as a speed problem. The leaders were not talking about abstract “loops,” but about something simpler: if AI can meaningfully help design and improve the next generation of AI, progress no longer depends entirely on how fast humans can hire, organize teams, or run traditional research cycles. That means capabilities could advance faster than most organizations — public or private — are structured to handle. In practical terms, this is when AI stops feeling like a new tool that gets adopted gradually and starts behaving like a force that reshapes systems before they have time to adapt.

The job discussion underscored how this plays out in the real world. Both leaders acknowledged that early disruption is likely to show up first in entry-level and junior roles — the very positions people rely on to gain experience and move up. Notably, neither offered a concrete solution for how societies should replace those pathways. The implicit message was that governments, educators, and employers will need to respond — but the conversation made clear that those responses are not yet in place, even as the pressure begins to build. That makes this a systems problem, not just a labor-market one.

The same pattern appeared in the discussion of governance. Both leaders said they would prefer more time to deploy AI carefully and safely. At the same time, they openly acknowledged that geopolitical competition makes slowing down unrealistic. For readers, this means many of the safeguards people assume will be designed before AI becomes deeply embedded may instead be created while those systems are already in use. That doesn’t guarantee failure — but it does mean risk management will likely be reactive rather than preventative, shaped by events as they unfold rather than by careful advance planning.

If productivity, intelligence, and decision-making are increasingly automated, societies will need new ways to define value, contribution, and purpose. Interestingly, the conversation was revealing for what it did not resolve. The leaders raised concerns about meaning, identity, and value in a world where work may no longer anchor daily life — but they did not claim ownership of those answers. Instead, responsibility was implicitly distributed outward to society at large. That matters because it highlights a gap: the builders of the technology recognize the implications, but they are not positioning themselves as the ones who will define how humans should live with it. Those questions, they suggest, belong to educators, policymakers, employers, and citizens.

So what, ultimately, are they planning? This conversation suggests they are planning for rapid capability growth, for partial automation of their own development processes, and for a world in which governance and social adaptation struggle to keep pace. The value of this discussion is not that it offered solutions — it didn’t — but that it signaled how seriously AI’s builders are treating the next few years as a critical window.

For readers, that is the takeaway: this was not a theoretical exercise or a Davos thought experiment. It was an early warning that decisions about education, work, governance, and safety may need to be made before certainty arrives, not after — because by then, progress may already be moving too fast to pause.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.