Demis Hassabis just took apart the "AGI is basically here" story. He did it with a physics problem from 1905. He runs Google DeepMind and won a Nobel Prize for using AI to solve a biology puzzle that had beaten scientists for half a century. Very few people in this field have a record like that. On the NothingButTech podcast, he went after the biggest claim in AI right now. The loudest people in the industry keep saying AGI has more or less arrived. OpenAI's own definition is a system that beats humans at most "economically valuable work." Put simply: replace enough jobs and you can declare victory. Hassabis thinks that bar is far too low. His view is that true general intelligence must match what a human brain can do, since the brain is our only evidence that this kind of mind can exist at all. He called it "a higher bar than just being able to do some useful economic work," which is about as blunt as a courteous British Nobel winner gets about his rivals. Then he described a real test. Current AI has absorbed everything people have written, relativity included. When it explains relativity to you, it is reciting an answer that already exists. That isn't intelligence. So he proposed a test that memorization can't pass. Train a model only on what humans knew in 1901, four years before Einstein published. Then ask it to arrive at relativity by itself. There's nothing to look up, because in 1901 the answer hasn't been written. The only route is the one Einstein took: start from the physics everyone shared and reason toward an idea nobody had. Hassabis says no AI today can do that, however much it has memorized. What we call "almost AGI" is really the greatest librarian ever built. It can retrieve any answer that exists. It can't produce one that doesn't. His second test is sharper still. AlphaGo, built by his own team, famously played a move no human had tried in 2,000 years of Go. People called that genius. Hassabis says it still misses the bar. The real question isn't whether an AI can find a new move inside Go. It's whether an AI could design a game as deep and elegant as Go to begin with. No model today can. The people saying AGI is already here are often the same people raising hundreds of billions on that promise. Those valuations need the finish line to be close. So the line keeps moving closer, and AGI keeps being redefined down to "does useful work," until whatever is already on sale happens to meet it. Hassabis has nothing left to prove and nothing to sell. He has the Nobel, and he is saying machines still can't manage the single thing that would make them truly intelligent: have an original idea. To be fair, he's no pessimist. He believes true AGI will arrive, and he's devoting his career to building it. He just won't pretend it already lives in your phone. Next time a founder says AGI is months out, remember the Nobel laureate who designed his test around Einstein and conceded that nothing we have built can pass it. Where do you land? Interested in AI? Follow @AlltheLatestAI and never fall behind. I track ChatGPT, Claude, and how every new agent is changing how we work and think, then hand you the pure signals.