What 90,000 Rollouts Reveal About AI Agents Doing Science - Ian Quigley, Leash Bio

Leash Biosciences built a platform to physically measure thousands of proteins against millions of molecules. In this followup to our Leash Bio podcast with cofounder Ian Quigley, he returns to discuss the release of Minos, Leash’s framework of pre-measured experimental worlds that scientific agents can explore. Ian goes into how scientific agents traverse these environments, whether they use information from prior experiments to inform future choices, how good they get at medicinal chemistry, and how their broad decision making abilities affect their performance. We cover: - Huge action spaces that can be explored by reasoning agents - Which models are the best at medicinal chemistry - What mistakes they make and how Leash can tell - How agents weren’t ready to do this at all a year ago - The basic idea of a lab-in-the-loop with zero latency - The path to large numbers of new worlds Leash published this work and is releasing a few dozen Minos worlds and thousands of agent transcripts for the community to use. Original YouTube with Ian Quigley: youtube.com/watch?v=hQSBJVreLms Blog post: leashbio.substack.com/p/making-reality-replayable Preprint: github.com/Leash-Labs/minos-paper/blob/trunk/Minos_Leash.pdf Data and transcripts: strelka.leash.bio/ Sponsorship and business inquiries: business@latent.space