The man with the Nobel Prize has a question for the men raising the billions. Google’s DeepMind CEO Demis Hassabis, who shared the 2024 Nobel Prize in Chemistry for AlphaFold’s protein-structure work, told the audience at Google I/O 2026 last month that no AI on earth could pass a test he built around Albert Einstein. The test is simple. Train a system on everything humanity knew in 1901, before Einstein published special relativity. Then ask it to derive general relativity on its own. It cannot, because in 1901 the answer does not exist. The only way to pass is to do what Einstein did.
That is the bar. The rivals selling AGI-imminent to the public have set the bar considerably lower. Hassabis’s challenge lands as OpenAI, Anthropic, and Elon Musk’s xAI race to ship products on the promise that human-level AI is months, not years, away.
🔍 THE BOTTOM LINE
The loudest voices in AI are telling you the finish line is in sight. The only person in the room with a Nobel Prize says the machines cannot do the one thing that would make them genuinely intelligent: have an idea that has never existed before. The contradiction is too stark to ignore.
The Einstein Test: What It Actually Asks
Hassabis’s test, reported by The Outpost’s I/O 2026 coverage and analysed in depth by Zaruko, cuts through the AGI definition problem in a way most benchmarks do not. Today’s systems have read every physics textbook ever written, including the theory of relativity. When they explain relativity back to you, they are repeating an answer that already exists in their training data. That is not intelligence. That is the best librarian in history.
Cut the data at 1901 and the trick stops working. The model has no relativity to retrieve, no Lorentz transformations to recombine, no Maxwell’s equations to extend. It has to generate the next step itself, the way Einstein actually did, working from a stack of late-19th-century papers and a stubborn belief that the equations were incomplete. Hassabis, in his IISc Bangalore talk at the India AI Impact Summit in February 2026, was blunt: “It’s clear today’s systems couldn’t do that.”
The point of the test is not the physics. The point is originality. Can the system produce knowledge that was not in its training set? If not, what we have is a compression of human output, not a generator of it. As Zaruko’s analysis frames it: LLMs interpolate, they find patterns and recombine them. AGI would need to extrapolate, to reach conclusions that do not exist anywhere in the data. This is the same gap Hassabis was pointing at in his foothills-of-the-singularity remarks at the same I/O keynote — a thousand-fold compute increase will not, on its own, close it.
Beyond AlphaGo: Even Move 37 Is Not Enough
It would be easy to assume Hassabis is being humble. He is not. His own team built AlphaGo, the system that in 2016 produced Move 37 against Lee Sedol, a play no human had made in roughly two thousand years of Go. Most people called it a sign of creative intelligence. Hassabis does not.
Move 37 was a brilliant move inside Go. It was not the invention of Go. The deeper test, Hassabis said, is whether an AI could invent a game as deep and as beautiful as Go in the first place. No model that exists today can do that. The space of possible games is enormous; Go is a single elegant point in it. A system that can only traverse known spaces, however cleverly, has not earned the word general.
He also called this pattern jagged intelligence, the right phrase for a system that can win gold at the International Mathematical Olympiad one minute and stumble on arithmetic the next, depending on how the question is framed. A true general intelligence should not be jagged. It should be roughly as good at a wide spread of problems as a capable human adult. Today’s systems are not. The shape of their failure, on tasks that should be easy, is the giveaway.
The Fundraising Tell
When AGI timelines align with capital raises, pay attention. Hassabis says 2030, plus or minus a year. Musk has said the end of 2026. Anthropic’s Dario Amodei has said 2027, with the memorable framing of “a country of geniuses in a datacentre.” Sam Altman has been saying “any day now” for several years. Each of these predictions comes from a company whose valuation rests on the AGI-imminent thesis. Hassabis has a Nobel. He has nothing left to sell you.
The original thread that crystallised this contrast, a long read from X writer @Ric_RTP that has clocked 850,000 views since Thursday, made the obvious point that may be worth saying out loud: the people telling you AGI has already arrived are the same people raising hundreds of billions of dollars on that exact promise. The valuations only work if the finish line is right in front of us. So the finish line keeps getting quietly dragged closer, and AGI keeps getting redefined down to “does useful work,” until the products they already sell happen to qualify.
OpenAI’s published definition, that AGI is “a system that can outperform humans at most economically valuable work,” is the tell. By that bar, a spreadsheet with a good plugin is on the path to AGI. Hassabis calls that bar “a joke,” which is about as close as a polite British Nobel laureate gets to calling his rivals out. It is the same dual-message pattern Sam Altman has been running for years, and it deserves a separate column.
🇳🇿 NZ Angle
New Zealand has no Nobel laureate in technology. We are also a small, open economy with a tech sector that lives or dies on global sentiment. When US venture capital decides AGI is six months away, that sentiment reaches us through Nvidia’s share price, the AI adoption mandates coming out of Wellington, and the steady drumbeat of corporate training programmes telling Kiwi workers to upskill for jobs that may not exist in 18 months.
Hassabis’s slower timeline is, paradoxically, the honest one. A 2030 horizon gives small economies a real window to adapt, to invest in literacy and policy on human terms, instead of scrambling to train for a labour market that turns out to be 4-7 years further away than the loudest pitch deck claimed. The Einstein Test, in a small way, is a useful piece of consumer protection. Anyone telling you AGI is imminent should be asked: would your system pass it? If they squirm, you have your answer.
⚖️ The Other Side
It is fair to note that Hassabis is not a pessimist. He believes real AGI is coming, and he is spending his life building it. He just refuses to pretend it is already sitting in your phone. He also has skin in the game: Google is racing the same frontier labs he is calling out, and his I/O 2026 stage was, after all, a Google marketing event. He is not neutral. He is just better positioned than most to know when the bar is being lowered.
The Move 37 critique is also worth one more pass. There is a serious argument that within a closed system, an AI that produces a Move 37 is doing something like invention. Hassabis’s reply, which is the right one, is that invention of a new move is not invention of a new game. The bar is what matters, and the bar he has drawn is the right bar.
❓ FAQ
Q: What is the Einstein Test, in one sentence?
A: Train an AI on all human knowledge up to 1901, then see if it can independently derive general relativity the way Einstein did from 1905. If it cannot, you have a sophisticated retriever, not a general intelligence.
Q: Why does Hassabis dismiss AlphaGo’s Move 37 as evidence of AGI?
A: A brilliant move inside Go is a brilliant move inside Go. The deeper test is whether an AI could invent a game as deep and elegant as Go. No current system can. That is the bar Hassabis is drawing.
Q: What is “jagged intelligence”?
A: Hassabis’s term for systems that ace the International Mathematical Olympiad one minute and fail basic arithmetic the next. A true general intelligence should not be jagged; it should be roughly as reliable across the range of human cognitive tasks as a capable human adult.
Q: How does Hassabis’s AGI timeline compare to his rivals’?
A: Hassabis says around 2030, plus or minus a year. Musk says end of 2026. Anthropic’s Dario Amodei says 2027. OpenAI’s Sam Altman has been saying “any day now” for years. The contradiction with the fundraising cycle is hard to miss.
Q: Should Kiwi workers and businesses trust the AGI-imminent narrative?
A: Treat it the way you would treat any unverified investment pitch. Hassabis’s slower 2030 timeline gives a real window to adapt, invest in literacy, and write policy on human terms, instead of scrambling for a labour-market shift that the actual evidence suggests is years, not months, away.
🔍 THE BOTTOM LINE (Synthesis)
Hassabis has nothing to sell you that the market does not already believe. He has a Nobel, a working AGI roadmap, and the credibility to call out the finish-line-dragging happening in public. The Einstein Test is not a thought experiment. It is a falsifiable benchmark, and the loudest voices in AI are not running from it, they are simply not engaging with it. The next time a founder tells you AGI is months away, remember that the man in the room with the actual Nobel Prize built his test around Einstein, and admitted that nothing we have built can pass it. The man without one is the one telling you we are almost there.
