AI researchers combine neural networks and logic to build autonomous science agents

AI researchers are pushing for a new generation of “thinking” models that don’t just crunch data but actually reason like scientists. Eric Schmidt, Google’s former CEO and now head of Schmidt Sciences, and Suhas Mahesh, who leads the group’s AI-for-science program, argue in a new paper that today’s large language models are still too dumb for real discovery. They say the field needs models that can generate hypotheses, design experiments, and interpret results—not just regurgitate papers.

The push comes as labs worldwide race to turn AI into an autonomous research assistant. Last year DeepMind’s AlphaFold solved protein folding, but that breakthrough was narrow. Schmidt and Mahesh want systems that can wander across disciplines, spot anomalies, and suggest new experiments without constant human prompting. Their team is building prototypes that combine symbolic logic with neural networks, trying to mimic how researchers actually think. The work is partly a reaction to today’s models, which excel at language but fail basic scientific reasoning tests.

The hardest hurdle is scale. These new “agent” models need vast amounts of curated scientific data and massive compute—thousands of GPUs running for months. Schmidt’s fund is committing tens of millions to the effort, while labs at Stanford, ETH Zurich, and Cambridge are racing to replicate the approach. Early tests show promise: one prototype recently proposed a plausible new catalyst for splitting water, though the idea still needs lab validation.

If it works, the payoff could be enormous. Autonomous AI agents might shave years off drug discovery, materials science, or climate modeling. But critics warn that the same tools could automate away junior researchers and concentrate power in a handful of tech-funded labs. Schmidt’s group counters that open-science releases and broad partnerships can keep the technology democratic. The next 18 months will show whether these agents can move from whiteboard sketches to real breakthroughs—or whether they’ll get stuck in the same data traps that hobble today’s AI.