AI autonomous labs miss serendipity: The challenge of discovering penicillin

Connor Blake's analysis, presented at the SciFM26 conference, examines why AI-driven autonomous labs might miss breakthroughs like penicillin. He identifies two core assumptions that conflict with historical scientific progress: first, that there is a clear numerical signal to optimize (e.g., yield), and second, that the solution lies within the convex hull of existing tools. In the penicillin thought experiment, an AI optimizing plasmid production for yield discards a sequence that kills all bacteria (0% yield) without recognizing it as a powerful antibiotic. Goodhart's Law applies: the metric (yield) fails to capture 'useful discovery'.

Blake proposes curiosity-driven learning as a potential fix, where the system predicts outcomes and explores anomalies when predictions fail (e.g., unexpected death of all bacteria). However, even then, the system is limited by its sensors—if the byproduct is a hyperspectral reporter, but no hyperspectral camera exists in the lab, the discovery is invisible. This 'convex hull' problem suggests that autonomous labs need to invent new tools, not just combine existing ones. Freeman Yang's work on tool-driven discovery reinforces this: many breakthroughs come from new instruments, not just high-throughput searches.