Demis Hassabis has a test for artificial general intelligence, and current systems fail it cleanly.
Give a frontier model a knowledge cutoff of 1901. Could it produce what Einstein produced in 1905 — special relativity among them? Not retrieve it, not recombine adjacent published work, but arrive at the insight (5:32). His answer is that today's systems plainly cannot, and that the gap is not one of scale but of a capability that has not appeared.
Four months earlier at Davos he had put his timeline at the end of the decade. Here he sharpens it: around 2030, plus or minus a year, arriving gradually rather than as an identifiable moment (4:01). The Einstein test is what he is waiting for.
The advantage he claims, and the number behind it
The competitive section of the interview is the least interesting part, except for one argument that is structural rather than promotional.
Hassabis claims Google is the only organisation holding the full stack — chips, data centres, cloud, a frontier research lab, and products reaching billions of users (9:07). The claim is checkable in principle and the feedback loop he describes is the point: the lab's output reaches product surfaces immediately, and the resulting usage informs the lab.
His evidence is 900 million monthly users on the Gemini app (8:18), which he attributes less to model quality than to an engineering decision — rewriting the product stack to be AI-first and then agent-first, so that improvements ship to every surface rather than being integrated one at a time.
He is also careful about the framing that the press prefers. Asked whether Google will win, he declines the question's shape: he expects a handful of significant players and says the more important obligation is shared across all of them.
What acceleration should look like
The most instructive passage is not about strategy at all. It is a story about a decision made in a meeting.
AlphaFold could predict a protein structure in seconds, and roughly two hundred million proteins were known to science. The conventional plan was a submission service — researchers upload a sequence, wait, receive a structure. Hassabis did a back-of-envelope calculation and asked why they did not simply fold all of them and release the results (16:52).
He offers this as his working definition of acceleration worth having, and the contrast he draws is deliberate. The service model would have been useful, revenue-adjacent, and defensible. Folding everything and publishing it was none of those, and it made a category of research possible that the service model would have rationed.
Read alongside his complaint that the industry does not demonstrate unequivocal benefits often enough, the story functions as an argument about what such demonstrations require: not better communication, but occasionally choosing the version that captures less.
Simulation as the missing input
The technical thread most likely to be underrated concerns what simulation is for.
He describes using an interactive world model to generate driving scenarios that essentially never occur — a forest fire surrounding the road, an aircraft making an emergency landing on a motorway, an elephant appearing in traffic (22:36). These are used to test autonomous systems against situations that will never be in the training data, and to generate more data where the system performs badly.
The framing worth extracting is that simulation is not a cheaper substitute for reality here. It is the only access to the tail of the distribution, and the tail is where the failures that matter live.
Where he says the gap actually is
His account of what remains missing is more specific than the usual list.
Self-improvement is being pursued everywhere, and he expects it to compound in coding and mathematics for the reason he gave at Davos: those outputs are verifiable and synthetic data is cheap to generate (28:04). The harder problems are elsewhere.
Continual learning — systems that keep learning after release rather than being frozen at training — he names as a genuine gap (30:24). Alongside it he puts consistency, and his description of the current failure is precise: models that are brilliant on hard problems yet trip on a strictly simpler version posed slightly outside the expected form. He calls this jagged intelligence, and observes that human intelligence does not behave this way. Someone competent at a task can handle an easier task in the same domain. A system that cannot has not generalised; it has memorised a region.
The advice, and why it lands
He closes on a warning aimed at the place he was standing.
He spends about half his time in the Bay Area and values the energy. He also detects a frenetic quality that he does not think is conducive to thinking deeply about hard problems for a long time — which is what the next advances will require, since he expects them in deep technical work applied to hard science rather than in faster iteration on what exists.
His formulation is blunt enough to survive quotation: running a hundred miles an hour in the wrong direction is worse than standing still long enough to get the bearing right (40:37).
Coming from someone who has just given a date for artificial general intelligence, it is an unexpected note to end on. It is also consistent with the rest of the hour. His timeline is not a prediction that speed produces the outcome. It is a prediction about when a specific missing capability arrives — and nothing about working faster makes that capability appear sooner.
这些演讲者在其他会议上的发言： Hassabis and Amodei on the Day After AGI (Davos 2026)
关键数据
演讲章节
关键要点
- 01
Hassabis put AGI at around 2030, plus or minus a year, and described the arrival as gradual rather than a single identifiable moment. 4:01
- 02
His threshold for AGI is an Einstein test: a model with a 1901 knowledge cutoff producing the 1905 insights, including special relativity — a leap of insight today's systems clearly cannot make. 5:32
- 03
He put the Gemini app at 900 million monthly users, attributing the jump to rewriting Google's product stack to be AI-first and then agent-first. 8:18
- 04
He claimed Google is the only organisation holding the full stack from chips through data centres and cloud to a frontier lab and billion-user products. 9:07
- 05
He sized the transition at ten times the impact of the Industrial Revolution at ten times the speed — roughly a hundredfold — and called that an underestimate. 11:30
- 06
The AlphaFold decision to fold every known protein at once, rather than operate a submission service, is his working definition of acceleration worth having. 16:52
- 07
DeepMind uses Genie, its interactive 3D world model, to generate one-in-a-billion driving scenarios — a forest fire, an emergency aircraft landing — for testing Waymo systems. 22:36
- 08
He named continual learning and consistency as the missing pieces: an AGI should not show jagged intelligence, failing a strictly simpler problem posed slightly out of distribution. 30:24
- 09
His advice to the Bay Area was that direction beats velocity — running at a hundred miles an hour in the wrong direction is worse than standing still to get the bearing right. 40:37
提及的实体
相关演讲
A year after their first joint appearance, the heads of Anthropic and Google DeepMind returned to a shared stage and disagreed mainly about speed. Amodei held to a horizon of one to two years for systems that outperform humans across most cognitive work, resting the claim on a self-improvement loop that runs through code; Hassabis kept to the end of the decade, arguing that verifiable domains like coding and mathematics automate far earlier than natural science, and that the capacity to pose a new question rather than answer an existing one is still missing. The exchange is most useful where they converge: both accept the loop is the variable that decides everything, both are sceptical of doomerism without dismissing the risk, and both want more time than the competitive dynamic allows. Amodei's chip-export argument and Hassabis's call for minimum international safety standards are the two concrete policy asks.
Four of Google's model, product and search leads on what changes once agents run for hours rather than seconds, and the most quotable argument comes from Dean: the constraint is moving out of the model and into the tools around it. By Amdahl's law, an agent spending half its time in tools built for human-speed interaction cannot gain more than a doubling however fast the model becomes — which reframes a great deal of current infrastructure work as latency debt. Their internal response is concrete: rewriting Python tooling into Go, framed as a fully specified translation task rather than an open prompt, produced order-of-magnitude speedups overnight. Reid supplies the counterweight from Search, where acceptable latency turns out to scale with how much work is being taken off the user rather than being a fixed budget. Woodward's detail is the quietest and perhaps the most telling: teams that have stopped writing product documents for humans and now write context files for models to act on directly.
The anecdote that opens the panel does the work: a robot asked to pick up the extinct animal selected a dinosaur toy, with nothing in its training data connecting the phrase to the object. That transfer from language models into machines with hands is the premise of the current wave. What the practitioners then describe is where it stops. Physical intelligence is about exerting force and using a body to do it, which is knowledge about consequences — the one thing a corpus of internet images contains almost nothing about. The humanoid question gets an honest treatment: not that human shape is optimal, but that the world is already built for it, plus a development-loop argument about collecting data and deploying on the same hardware. The most useful passage is scepticism about the field's favourite shortcut: generated video looks realistic and does not hold up for dexterous manipulation, because looking right and being physically consistent are different properties.
The most honest moment here is an aside about how the presenters have tracked their own projects: plans in documents, plans in spreadsheets, plans in bug comments, and once a plan written on a receipt. That describes the actual category being addressed — not software nobody has built, but the small internal tool every team improvises badly because building it properly was never worth the effort. The demo turns on a single question: the generated app is strictly single player, so what happens when you want to share it with the team? That boundary is where improvised tools historically died, because it is where accounts, shared storage and access rules begin. Here it is crossed in one step, with the access rules generated and deployed automatically — which is convenient, and is also the moment the application acquires obligations nobody reviewed.
The clearest statement at I/O of how an agent API differs from a chat API, and the reasoning behind each departure is stated rather than assumed. Three changes matter. Conversation state moves to the server: a call returns an identifier, and passing it back continues the thread, retiring the client-side history array. The data model abandons alternating user and model turns for discrete steps, on the argument that a trace containing reasoning, tool calls, environment responses and compaction was never really a conversation and modelling it as one distorted it. And agents receive their own persistent remote environment rather than acting on the caller's machine — addressable by identifier, and shareable, so a research agent's output files become an application builder's input without passing through the context window. Schmid is explicit that scaffolded environment files are deliberately not model input, which is what keeps large artefacts out of the context budget. Schaeff's first half covers the real-time voice path, where the notable property is speech-to-speech across ninety languages with transcription of both directions.
The observation that organises this session is not about capability but about attention. Engineers increasingly file a ticket rather than opening an editor, and the code comes back — which changes what the surrounding tools are for. If the agent works while you do something else, the conversation between you is no longer a workspace; it is the mechanism by which the agent surfaces a question it cannot resolve alone. Interfaces built for continuous conversation optimise for flow, and interfaces built for interruption should optimise for the opposite. A runtime constraint follows immediately: an agent that starts a long-running job cannot block until it finishes, which turns out to be a workflow-engine problem rather than a model one. The panel's closing formulation — that deciding what to build is the hard skill and always was — reads as reassurance and functions as a warning, since that judgement is downstream of exactly the work now being delegated.