
Highlights and key takeaways from DeSci Summit Mumbai this week 👇 → The IP question decides whether a biologist can use frontier models at all. Several PhD labs bar ChatGPT and Claude outright, since prompting them leaks unpublished research. → Ask a frontier model to design a 10 amino acid peptide that binds and inhibits the SARS-CoV main protease and the request gets flagged. Both refusals ran live on stage, and the same prompt at @biosdiscovery returned an answer plus a route into a binder generation pipeline. → Finding a binder is finding a needle in a haystack. Doing it in silico only became possible after @GoogleDeepMind in 2020 and RFdiffusion from the Baker Lab. → Ten scored binders come back within ten minutes. The models sit on 200 GPUs, which opens the platform to anyone who shows up today. → RFdiffusion 4 runs on Mac, ported by @GeeveGeorge and currently the fastest RFdiffusion MLX model. A scientist designing a drug keeps every piece of data on their own machine. → A Qwen 8B model runs at 100 tokens/sec on M1 through M5 using custom kernels, ahead of ZSpec and Splash. The team calls it spec metal. → Qwen orchestrates and RFdiffusion designs, so drug discovery runs fully locally. The @biosdiscovery CLI is agent-native, built on the view that agents will be the ones using this software. → In coding and math the verifier is fast. In biology the verifier is the world, which is why wet lab demand rises as more people design medicines with AI. → Robios ran its first live demonstration. It loads and unloads the machine and records conditions like temperature, so two runs can be compared. → Human in the loop throughout. A senior scientist watches from home, gives feedback at any time, and steers it mid-run. → Competitors buy their robots from OEMs at prices most labs are unable to afford. Robios is built on open source and aimed at NCBS Bangalore and a wet lab in Mumbai. → Local agentic swarms are next, 10 to 20 agents on a MacBook reading real research papers and citing them. Citation is how hallucination gets reduced. → A research proposal takes around 116 hours to write. Grant acceptance has declined to 19% in the UK, 13% for US health research, and 8% for EU early career awards. → Publication costs roughly $12,000 at Nature and 17 months. @openlabsbio puts those first steps in one place, with an index covering failed and ongoing work, challenges for calling in skills, vaults funding first studies, and a launchpad for pre-company raises. Idea, lab, funded IP, product. Every layer we build shortens one of those and opens the next. Full talks ↓












