Latent Space: π¬Biology Is Turning Into Software β Matt McPartlon and Neil Patil, Chai Discovery
retained text
Chai Discovery co-founder Matt McPartlon and product lead Neil Patil explain why pharma abruptly started buying AI design tools this year instead of forcing every AI-bio company to go build its own pipeline β and what specifically changed in the models to make drug design teams trust them. We trace the lineage from Chai-1, which they open-sourced for reasons that had almost nothing to do with the model itself, through the Chai-2 campaign that convinced Lilly, Pfizer, Novartis, and argenx to sign. Along the way: why their CEO says the company's real competitor is a mouse; the cryo-EM result so accurate the team's first reaction was to suspect the lab of sending back their own file; why the product looks more like SolidWorks than ChatGPT, and why Neil expects to throw most of it away; what happened when a pharma scientist saw results on a target she'd spent a decade on; why the compute market is quietly mispriced for this class of model; and the one bottleneck both guests would remove by fiat β which is not data, not compute, and not the models. 00:00:00 β What the product actually looks like 00:01:15 β Intros: Matt and Neil 00:03:22 β Four pharma partnerships 00:05:00 β The software-layer bet 00:07:00 β The 50-target challenge 00:08:12 β What is an antibody, and why target it 00:11:22 β ADCs, bispecifics, and pressing the switch 00:12:39 β How this was done before: mice and yeast display 00:14:29 β Selectivity and cross-reactivity 00:17:26 β Chai-1 and the MSA detour 00:19:41 β Sitting in the OpenAI offices 00:23:09 β Anatomy of a structure prediction model 00:25:49 β Chai-2: all-atom diffusion and crossing into design 00:28:21 β Co-designing sequence and structure 00:29:48 β How do you know a design is any good? 00:32:36 β Validation, cryo-EM, and the 0.33 Γ result 00:35:26 β Chai-2.5, Chai-3, and betting on the models 00:39:18 β Building the product under pharma IP constraints 00:42:32 β Convincing scientists who hate AI tools 00:44:58 β The ten-year target 00:45:50 β Battle-testing without your own pipeline 00:47:40 β Waterfall to loop 00:49:04 β Levels of abstraction, and throwing the product away 00:52:06 β Epitope prediction: the harder problem 00:55:29 β The economics question 00:57:29 β Modalities you can't get from immunization 00:59:48 β What's really blocking science-to-engineering 01:01:51 β Buying compute is a terrible job 01:04:18 β Triangle attention and why GPUs hate it 01:05:28 β Durable execution and Temporal 01:08:27 β Complexity vs. the bitter lesson 01:09:37 β Inductive bias and data efficiency 01:15:13 β Will protein design get commoditized? 01:19:13 β No pipeline, no data moat? 01:21:13 β Per-partner fine-tuning 01:22:36 β Are all AI companies consulting companies? 01:23:42 β The highest-value token in any domain 01:24:20 β Pharma as VC: Genentech and Eroom's law 01:26:31 β Everyone is a capital allocator 01:29:23 β One bottleneck removed by fiat 01:32:07 β Takeaways
document fetched 2026-08-17 06:35 Β· canonical https://youtube.com/watch?v=Qp5xklyJySI