What's this about?
A DeepMind public-policy essay ("Conjecture Machines") arguing that AI agents are getting good at proposing hypotheses and designing experiments — so the new bottleneck in science is validation: physically testing ideas in the real world. It outlines four priorities for policymakers and funders to expand lab capacity, automation, and verification infrastructure.
What exactly is the "validation bottleneck"?
Science has two halves: generating ideas and testing them. AI just made idea-generation nearly free — a "conjecture machine" can propose thousands of plausible hypotheses. But wet labs, clinical trials, and field tests didn't get faster. When one half of a pipeline accelerates 100x, the other half becomes the constraint — and that's now where scientific progress will be won or lost.
Why is it worth noting?
606 likes; policy essays rarely trend. It matters as a statement of where DeepMind thinks the AI-for-science frontier actually is: not smarter models, but the physical world's throughput.