DeepMind's Hassabis Proposes 'Einstein Test' as AGI Benchmark, Targets 2030 Timeline

A New Benchmark for True Intelligence

Demis Hassabis, CEO of Google DeepMind and 2024 Nobel laureate in chemistry, has laid out what he sees as the remaining barriers to artificial general intelligence in a wide-ranging interview published on 29 April. Among his most striking proposals is a benchmark he calls the 'Einstein Test', which would train an AI system exclusively on knowledge available before 1901 and then ask it to independently derive Einstein's 1905 breakthroughs, including special relativity. 'Once it does, these systems will be close to inventing genuinely novel concepts,' Hassabis said during the conversation titled 'Agents, AGI & The Next Big Scientific Breakthrough'.

The Road to AGI by 2030

Hassabis believes existing techniques such as large-scale pretraining, reinforcement learning from human feedback, and chain-of-thought reasoning will form part of AGI's final architecture, but 'one or two' missing pieces may still be needed. He identified continual learning, long-horizon reasoning, and memory as the core unsolved challenges, placing his personal AGI timeline at around 2030. He acknowledged a striking paradox: models that solve International Mathematical Olympiad gold-medal problems can still commit elementary arithmetic errors when a question is rephrased. 'Something appears missing in its introspection over its own thought process,' he said. He believes a major AI-led scientific discovery is 'imminent', and that the more demanding test would be for an AI to propose an entirely new set of problems as profound as the Millennium Prize Problems.

Isomorphic Labs and the Virtual Cell

Hassabis revealed that Isomorphic Labs, the AI-driven drug discovery company spun out from DeepMind, is on the verge of announcing new results. The company, which earlier this year announced a research collaboration with Johnson & Johnson and had its AI-designed drug candidate ISM8969 cleared by the FDA for human clinical trials in January, now has 19 programmes spanning cancer, cardiovascular disease, and immunology. The longer-term ambition is a complete virtual cell, a fully functional cellular simulator whose outputs closely match experimental results, a goal Hassabis estimated is roughly ten years away.

Advice for Founders

Hassabis urged deep-tech founders to take AGI timelines seriously when planning decade-long ventures, and to focus on domains involving the physical world where shortcuts are unlikely. 'Life is short and energy finite, so invest your vitality in things no one else will do if you don't,' he said.