entrepreneur, founder Om, founder of 1859, inc. — love intersection of AI and techbio — prefer to build.
San Diego, CA
Joined June 2009
- Tweets3.8K
- Following2.2K
- Followers662
- Likes6.8K
1/ 🧬 @MSFTResearch released BioEmu2 with open weights under MIT. their headline: a 5 millisecond peptide binding simulation that would take more than 10 GPU-years with MD ran in 40 minutes on the same GPU. 👇
5/ 📊 related: a JCIM benchmark out oct 20 tested four cofolding models, alphafold3 included, on kinases and GPCRs. ligand placement was generally reliable, but state-defining conformations like kinase activation loops were often mispredicted. pubs.acs.org/doi/10.1021/acs…
6/ ✅ model card and code: huggingface.co/microsoft/bio…
for people who run MD for real projects: where would you trust a 40-minute emulated trajectory enough to skip the full simulation, and where wouldn't you?
I’m definite not a decel - speed is the name of the game. However, this is getting insane. The biggest opportunity in AI might not be data centers… it might be cybersecurity. Companies are probably going to spend single digit percentage points of revenue on cybersecurity alone at this rate.
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing-ou…
All these crazy raises are forgetting you need to make actual discoveries that pharma will buy. Seeing a lot trying to reinvent the wheel with automated chemistry, automated biology, when all of this infrastructure already exists across vendors that have done this for decades. Not gonna end well.
1/ 🧪 @bmsnews is putting @boltz_bio's platform and proprietary models on its own NVIDIA DGX SuperPOD. boltz says BMS scientists have long used its open-source models. this deal brings the commercial stack inside. 👇
5/ 📊 related, a day later: @Bayer signed with @phylo_bio to give its scientists Biomni Lab, an agent platform for pulling literature and omics evidence for drug discovery. chugai and ono already use it. phylo.bio/blog/bayer-partner…
6/ ✅ the announcement: boltz.com/bms-partnership
for comp chem teams at pharma: when a vendor model moves inside your firewall, what internal benchmark does it have to beat before your chemists actually use it?
1/ 🧪 @MicrosoftAI's diagnosis agent is now a @ScienceMagazine paper. on NEJM case records it got 83.2% right, against 36.1% for generalist physicians who could use search and references. on ER admissions, 80.3% vs 46.8%. 👇
5/ 📊 related: this week @SophontAI released holon-s1, which predicts spatial gene activity from routine H&E slides. research use only, evaluated across 16 organ types. prnewswire.com/news-releases…
6/ ✅ the paper: science.org/doi/10.1126/scie…
microsoft says a prospective study with beth israel deaconess is next. for ED docs and hospitalists: what should that study measure before you'd let this into your workflow?
