What Happens When AI Gets a Lab?

What happens when AI moves beyond the chatbot and into the laboratory? From Claude-designed proteins that bind scorpion toxins to robots that assemble optics experiments, we examine what AI can do for science and what still has to be proven in the real world. Lester Nare and Krishna Choudhary explore the gap between generating a promising design and demonstrating that it works. We also unpack BrainChat, a research platform that lets epilepsy patients interact with their brain-recording data and contribute feedback to seizure detection, and a 15-year study of orangutan mothers facilitating playdates for their young. Along the way: why binding a toxin is not the same as neutralizing it, how patient feedback can help personalize models, and what cloud laboratories and automated experiments could mean for scientific discovery. Which part of science do you think AI will change most? CHAPTERS 00:00 Playdates, brain data and robot labs 00:29 Hello Internet and Signal Award celebration 02:40 Orangutan mothers and infant playdates 15:18 What 15 years of orangutan observations reveal 25:16 Testing the playdate hypothesis 39:30 BrainChat and patient-guided learning 44:57 Putting patients in the feedback loop 1:02:00 Personalizing seizure detection 1:05:33 What the BrainChat study does not yet establish 1:07:59 Anthropic scorpion-toxin binders 1:15:49 Why a toxin binder is not an antivenom 1:18:31 Fast AI designs, slower laboratory testing 1:19:40 Connecting AI to the physical lab 1:23:22 Robotic optics and automated laboratories 1:28:04 Cloud laboratories and remote experiments 1:33:00 Closing the loop with Lila Sciences 1:37:48 Wrap-up RESEARCH & FURTHER READING Orangutan maternal ranging and infant social play (PNAS): https://pubmed.ncbi.nlm.nih.gov/42837467/ BrainChat and human-AI collaboration in epilepsy (Nature Medicine): https://www.nature.com/articles/s41591-026-04680-4 Claude-designed scorpion-toxin binders (Anthropic): https://www.anthropic.com/research/claude-scorpion-toxin-binder-design Robotic optics experiments (MIT News): EDITORIAL NOTES The orangutan findings are observational evidence of mothers facilitating social play, not direct proof of conscious intent. BrainChat was a small feasibility study, not proof of improved clinical outcomes. The scorpion-toxin work demonstrated binding, not a validated antivenom. On-screen clarifications: 15:20 The first author is Odd T. Jacobson. Maria A. van Noordwijk is a coauthor. 16:06 About 32,600 observation hours were collected over 15 years. That is roughly 3.7 years of continuous observation. 40:39 In epilepsy, deep brain stimulation reduces seizure frequency. Tremor is the Parkinson's disease example here. 47:07 The study included 13 participants: 11 with scalp EEG and 2 with intracranial EEG. 51:01 and 51:35 GPT-4o-mini ran in Penn's HIPAA-compliant Azure OpenAI cloud environment, not locally. 1:18:03 This was an Anthropic announcement, in collaboration with Citrini Research. WATCH & LISTEN Episode notes and research: https://ffppod.com/episodes/ep65 Watch and listen: https://ffppod.com Apple Podcasts: https://podcasts.apple.com/us/podcast/id1830375095 Subscribe: https://www.youtube.com/@FromFirstPrinciplesPod?sub_confirmation=1 Support the show: https://ffppod.com/donate Contact: https://ffppod.com/contact Follow @FFPPod on Instagram, TikTok, X and Facebook.