Episode Details

Anthropic published wet lab results from Adaptyv Bio and Twist Bioscience showing Claude hitting 22-35% protein binder success against 14 of 15 targets — roughly double the industry baseline. We cover what it means for drug development, the honest caveat about export control constraints limiting the most capable models, the Reconstruction benchmark showing AI still struggles with genuine hypothesis generation, and GitHub Copilot Autofix introducing a shell injection vulnerability exploited in five days.