What would it take for AI to discover penicillin? (bosoncutter.substack.com)

At the recent SciFM26 conference held at the University of Chicago, discussions centered around the role of AI in scientific discovery raised intriguing questions about the potential barriers faced by autonomous labs in producing breakthroughs like penicillin. A hypothetical experiment illustrated a key challenge: when a lab optimized for yield inadvertently overlooked a failed plasmid that produced an antibiotic, it emphasized the limitations of current AI systems that rely heavily on specific quantifiable metrics and existing tools. Unlike Alexander Fleming’s serendipitous discovery, AI might miss crucial anomalies unless equipped with advanced capabilities, such as curiosity-driven learning, that allow systems to explore unexpected results beyond conventional targets. The conference dialogue also highlighted the “convex hull” assumption of autonomous labs, suggesting that they may restrict innovation by focusing only on available tools and methodologies. The argument posited that true scientific advances often arise from novel instruments rather than merely optimizing existing resources. As companies like Lila, Periodic Labs, and others push for autonomous systems, the conversation is shifting toward developing AI that not only analyzes data efficiently but also generates insightful hypotheses akin to human scientists. This reflects a growing recognition that achieving meaningful scientific progress may require more than just automation; it necessitates a deeper understanding of how scientific curiosity and exploratory reasoning can drive discovery.