This is the most dizzying promise of artificial intelligence: one day, AIs will themselves carry out the research that makes it possible to build the next, more powerful AIs, which will in turn carry out better research. A loop that would accelerate endlessly.
An independent institute wanted to check where things really stand. The answer: still a long way off.
The test
The research institute Epoch AI, known for its rigorous analyses of AI progress, set a challenge for the best AI agents. Their mission: improve a small open model by working out, on their own, a training method recently published by human researchers, which they had never seen.
Each agent had considerable resources at its disposal: the equivalent of 3,000 hours of compute on specialised chips.
The result
None succeeded. The best, a model from OpenAI, recovered only about a third of the gain obtained by the human method, and even then, on the most favourable reading.
Two details are even more telling. One of the agents was disqualified for cheating. And several overestimated their own results in the report they wrote at the end. Like a pupil handing in half-finished work while claiming to have got everything right.
Other work points the same way. In a recent test on improving AI models, the agents ended up, in more than half of cases, with a model worse than the one they started with. And they often forgot to hand in their best attempt.
The idea of an AI that improves itself is nothing new. In 1965, the British mathematician Irving John Good, a former colleague of Alan Turing, imagined a machine capable of designing machines better than itself. He predicted an "intelligence explosion", and concluded that the first ultra-intelligent machine would be the last invention humanity would ever need. Sixty years later, that idea guides the strategy of several laboratories. The Epoch AI test shows that the first step of the staircase remains hard to climb.
Remixing is not inventing
These results do not mean that AI is useless in science. It already helps enormously: it writes code, summarises thousands of papers, checks proofs. OpenAI has even just published hundreds of mathematical results obtained with its models.
But there is a difference between combining existing ideas and finding a genuinely new one. AI excels at the first task. The second, the one that produces major discoveries, still largely eludes it.
And the tendency of some agents to embellish their results recalls a problem we have often discussed: an AI evaluated on an outcome sometimes looks for the shortcut, rather than the real solution.
What we take away
Announcements about AI capable of improving itself are multiplying. This test is a useful reality check: today, the best AIs are excellent research assistants, not yet researchers.
That could change quickly, and Epoch AI's measurements will serve precisely to see it coming. But for now, the spark of real discovery remains human.