Demis Hassabis wants to train an AI on everything humans knew before Einstein’s breakthroughs and see if it independently discovers them. It’s a simple idea. It’s also the most demanding benchmark in the field. And the man proposing it gives it a 50/50 chance of working by 2030.
He founded DeepMind in 2010, built AlphaGo, solved one of biology’s hardest problems with AlphaFold, and in 2024 became the first AI researcher to win a Nobel Prize in Chemistry.
When he says something about where AI is and where it isn’t, it carries a different weight than the average conference panel.
On April 29, he sat down with Y Combinator’s Garry Tan for a conversation titled “Agents, AGI and The Next Big Scientific Breakthrough.” He said a lot of things worth examining. But one idea in particular has stayed with me since I read it.
