Hassabis and Amodei on the Day After AGI (Davos 2026)

Demis Hassabis and Dario Amodei disagree about roughly twelve months. Everything else on stage at Davos — the risks, the policy asks, the scepticism about doom, the discomfort with the pace — they hold in common. But the twelve months is not a rounding error, because both men locate the decisive variable in the same place, and their estimates of when it turns are what separates an orderly transition from an emergency.

The variable is whether AI systems can improve AI systems without a human in the loop.

What they actually agree on

It is worth being precise about the shared ground, because it is larger than the framing of the session suggests.

Both expect systems that outperform humans across most cognitive work, and neither treats that as speculative. Both reject the position that the outcome is fixed and catastrophic — Amodei calls himself not a doomer, Hassabis a cautious optimist — while insisting the risks are real enough to organise around. Both want more time than the current competitive dynamic allows. And both name the same mechanism as the accelerant: models good enough at code and at research to compound their own development.

Amodei's version is concrete. Engineers inside his own company, he says, have stopped writing code and now edit what the model produces; he puts the moment when models handle most or all of what a software engineer does end to end at six to twelve months away (1:57). His revenue figures are offered as supporting evidence that capability converts to commercial reality rather than staying in demos — roughly zero to a hundred million dollars in 2023, a hundred million to a billion in 2024, a billion to ten billion in 2025 (6:40).

The disagreement is about which domains resist verification

Hassabis holds to the end of the decade, and his reasoning is not caution for its own sake. It is a claim about where automation stops working.

Coding and mathematics automate early because their outputs are checkable. A program either compiles and passes its tests or it does not; a proof either holds or it does not. Large parts of natural science are not like this. A predicted compound has to be synthesised, a physical prediction has to be measured, and the feedback loop runs at the speed of laboratories rather than of inference.

His sharper objection is about a capability that has not appeared at all. Current systems answer questions; they do not pose them. Forming the hypothesis, choosing the conjecture worth attacking — what he calls the highest level of scientific creativity — is missing, and he is not confident that scaling produces it (3:48). If he is right, the loop closes in the verifiable domains and stalls at the boundary of the rest, which is precisely the shape that yields his slower timeline.

Amodei does not really contest this. He concedes the loop has components AI cannot accelerate — chip fabrication, training runs, the physical supply chain — and frames his own estimate as a guess that things move faster than people expect. The two positions are closer to a shared model with different parameters than to a genuine dispute about mechanism.

Amodei's policy ask is an attempt to escape a trap he describes

The most revealing exchange is not about capability at all.

Amodei says plainly that he would prefer Hassabis's timeline — that five to ten years would be better than one to two. Asked why he does not simply slow down, he gives the honest answer: he cannot, because geopolitical competitors are building the same technology at a similar pace and no enforceable agreement exists to make mutual restraint stable.

His proposal follows from that, and it is worth reading as an attempt to change the structure of the game rather than as trade policy. Halt advanced chip exports, he argues, and the race stops being between countries — at which point it becomes competition between a handful of labs, which he says with evident confidence he and Hassabis could work out between them (23:30). He reaches for a deliberately uncomfortable analogy to reject the counterargument that export access binds a rival into your supply chain (24:21).

Hassabis's ask is different in kind: minimum international safety standards for deployment, on the reasoning that a technology this diffuse will affect everyone regardless of where it is built. Neither man suggests the current environment makes either outcome likely.

On jobs, both of them are careful

The exchange most likely to be quoted out of context is the one where they are most disciplined.

Amodei restates his estimate that half of entry-level white-collar roles could disappear within one to five years, and immediately concedes the labour-market evidence for it has barely begun to appear (16:30). His worry is not that the market cannot adapt — he notes it has absorbed larger transitions — but that a compounding exponential outruns the speed at which adaptation happens.

Hassabis expects the ordinary pattern in the near term: disruption alongside new and possibly better work. His advice to students is unusually specific and does not depend on either timeline being right — become unusually proficient with the tools that already exist, because the capability overhang in current systems is large enough that mastering it may beat a conventional early-career apprenticeship (15:08).

Both then step past the economics. Hassabis's stated concern is not distribution but meaning: whether the institutions that would redistribute abundance exist, and separately whether the sources of purpose that people currently draw from work survive its restructuring. He thinks the second problem is harder than the first.

What to watch

Asked what will have changed by the time they next appear together, both name the same thing without hesitation: whether AI systems building AI systems actually closes the loop (29:47).

That is the useful takeaway from an hour of forecasting. Not the dates, which are estimates offered as estimates, but the agreement about which single observation would settle the question. If the loop closes in the coming year, Amodei's timeline was right and the policy conversation is already late. If it stalls where Hassabis expects — at the boundary of domains where you cannot check the answer quickly — then the additional years exist, and the question becomes whether anyone uses them.