Meta-pin👇

This is a huge deal, and a great time to remember that when people and organizations make a pivot on an issue that has historically been an in-group/out-group identify marker, it can be a very destabilizing experience. Even — and especially — if you felt their old position was stupid, it’s important to meet that vulnerability with compassion rather than derision.

Having never used TikTok and being generally very averse to slop, this is the video that finally made me viscerally appreciate that I, too, will be susceptible to falling into the abyss of the endless slop machine. Also an oddly emotionally evocative for me, perhaps because of the dissonance I experience between cute/catchy and disturbing.

True software insiders know the real danger of stochastic parrots is that to acquire a parrot is a major decision: it is likely to outlive you. If you don't know how to treat the parrot, it could be emotionally scarred and spend many decades feeling frightened and unhappy. If you buy a captured wild parrot, you will promote a cruel and devastating practice, and the parrot will be emotionally scarred before you get it. Meeting that sad animal is not an agreeable surprise.

Dostoyevsky on the death of his infant daughter. A passage that has stayed with me years after I first encountered it

My thoughts echo these in so many ways. For the longest time, MD identity has been tied to being walking reference manuals, but with evidence at your fingertips, that is less and less necessary. However, background understanding of phys/pathophysm and ability to reason over information, remains more important than every. In my view, access to information is never a bad thing; it just is about how it is used. I see trainees’ learning being accelerated by rapid access to knowledge, as long as that knowledge is being actively used with the intent of self-improvement and the benefit of future patients in mind, not just to accomplish daily tasks.

I'd like to share one of my favorite stories about a discovery we made at @LilaSciences. We’ve been working for a while on catalysts for clean hydrogen. One day, our AI proposed chemistry that a scientist who’d studied this exact reaction for 12 years thought had to be wrong. That composition became our strongest lead, sustaining more than 1,000 hours of oxygen evolution with little loss of catalytic performance. The scientist is John Gregoire (known as JG to us), who left his Caltech research professorship to become our Chief Autonomous Science Officer, and has been working to build some of the world's most sophisticated automated labs at Lila. Making clean hydrogen from water requires electricity to drive two reactions: one produces hydrogen, the other oxygen. The oxygen-producing half, the oxygen evolution reaction (OER), is particularly difficult. In proton-exchange-membrane electrolyzers, the catalyst must drive that reaction efficiently while exposed to strong acid and high voltage. Most materials either corrode or require too much energy to serve as an effective catalyst. Iridium oxide is the industry standard, but iridium is one of the rarest metals on Earth. Depending on it creates a supply constraint on scaling this technology. JG’s former group at Caltech spent 12 years exploring catalysts for this reaction. He knows how unforgiving these conditions are. Early in this project, the scientists checked the AI’s suggestions closely, occasionally overruling them. The choices were mostly familiar and sensible. As the experiments progressed, the team stepped back, retaining safety review while letting the system choose what to try. By the second campaign, which took just four weeks, it was producing results they hadn’t expected. The system had started exploring palladium-based oxides. When JG and Rafael Gómez-Bombarelli, our physical sciences CSO and an MIT professor, saw the results, they thought the model had gone astray. These were combinations that JG would have told his grad students weren't even worth testing. But the measurements were promising, so the team kept testing. Across 2,942 catalysts spanning 53 material systems and 26 elements, the platform uncovered an unexpected family of palladium-based oxides with a promising combination of activity and durability. It screened around 240 catalysts a week, an order of magnitude faster than a standard lab by our team’s estimate. The best contained tiny additions of indium and manganese, together accounting for about 0.09% of its metal atoms. At 10 mA/cm² in 1 M sulfuric acid, it maintained an overpotential below 0.5 V for more than 1,000 hours and retained 96% of its palladium. Unmodified palladium oxide crossed that voltage threshold after only 200 hours. Under the microscope, the team saw a needle-like structure that developed during operation and appears to be associated with the improved durability. We’re still working to understand exactly why it works. Palladium is also a precious metal, but its supply base is much larger than iridium’s. These catalysts contain neither iridium nor ruthenium, potentially giving us another option. Giving the AI control over which experiments to run was decisive. The system combined predictive models that learned from our experimental data with LLM reasoning that helped decide where to search next. It could pursue chemistry our scientists would have passed over, have the lab make and test it, and use the measurements to choose its next experiments. Each round of results changed where it looked next. More than 90% of the screening workflow was automated, with scientists providing oversight and transferring samples between instruments. These are laboratory results. There’s still substantial work ahead to establish performance under industrial conditions. People ask when AI and automated labs will start accelerating science. For us, it’s already changing what gets tested and how quickly we learn. We can make hundreds of materials a week, test them, and use the results to decide what to try next, including ideas our own experts would have passed over. I want a lot more of this: AI helping us discover things we can actually use to make the world better. Full story and preprint below 👇

Thrilled about this joint initiative with the @AnthropicAI team! We’re hiring for a few roles (eng, MD, partnerships) focused on these efforts. We look for extremely driven individuals that understand how transformative this tech can be and want to have an outsized, positive impact on the world. Please reach out if you think you fit the profile for any of the aforementioned roles!

Excited to see this out! Coming from drug discovery into ML, evaluating generative molecules is one my favorite parts. With this model, finding flaws got noticeably harder and critiques had to become much more nuanced. That’s a massive step forward. Congrats to the team!

Large Drug Discovery Model is out! I am exited to introduce our new generative SBDD framework. We experimentally validated LDDM across 5 targets and successfully designed novel, validated hits with structural accuracy confirmed by X-ray. Preprint: biorxiv.org/content/10.64898… 🧵 1/7

Let me rephrase this more honestly because we shouldn't exploit hope by masking science with marketing: 1. "optimize your embryo": suggesting there's some editing or engineering element. There is not. The optimization refers to your ability to choose between whichever embryo you were able to produce 2. "14 IQ points": suggesting you can get a kid that's 14 IQ points smarter than population average. This is wrong. What it means is IF you can get 10 embryos then on average they expect the estimated (and flawed and non-validated) IQ score difference between the worst and best egg is 14 points. So if your embryos theoretically produce 100 IQ score average (arbitrary, no validation, but representative of their concept) embryos, your lowest would be 93 and your highest would be 107. Also, the important number is embryo vs any embryo, and its +7 IQ in a pool of 10. Not that exciting right, assuming it even works? 3. "nearly a standard deviation": suggesting you can get a kid almost a standard deviation above population average, which means smarter than 84% of the world. This is false. The SD is relative to siblings IQ and we already know the polygenic scores explain 19% of the variance within sibling so really that collapses the IQ gain to maybe ~5 4. "trained on 1,000,000+ people": suggesting that it's generalizable or validated against outcomes. It's neither. It's more like only european ancestry, so if you're not then it's not even applicable. It's not validated outcomes, meaning there wasn't actually any "using our method we've seen 1 million people's IQ explained" 5. "Humanity can now direct its own evolution": No. 6. "The next generation can choose to do the same": suggesting you could potentially compound or carve your progeny's IQ trajectory. also no. there's like 80% of IQ unexplained, so you cannot actually directionally control any of it I'll be honest... I admire the ambition but I dislike the bs its wrapped in. Compressing the entire post into something more honest would sound more like: hey we've used some cool statistical methods on published data that's not really specific only to IQ but we did more cool statistics to try to isolate the IQ portion and we think our scores can help predict which of at least 10 of your european embryos would differ by 14 IQ points but we're really not entirely sure that clinical outcome is real and I know the average woman doing IVF is above 35 years old and only retrieves 2 embryos so...

It’s been incredible to see how capable Claude is at optimizing biological ML models! Neither @amirshanehsaz nor I had much experience with performance engineering starting this 4 weeks ago, but with Claude we found significant speedups we hope will be helpful for the community

Americans are more negative than positive about AI’s impact across nearly every area surveyed. The one exception: healthcare. We have a unique opportunity to showcase the benefits and set the standard for how AI can meaningfully improve people’s lives.

Beautiful reflection.

Finally get to say that a drug I helped discover is in human trials. The team sprinted for 6 years on this. We went from a thought-to-be undruggable target to profound degradation very quickly, and then several more years optimizing for human dosing. That speaks to the power of DNA encoded libraries and modularity of TPD, and the hard road that is drug *development*. ir.nurixtx.com/news-releases…

Not saying this is true, but a whole lot of this plot would seem to be explained by: “Older sibling is nuts, chronically stressing out the younger sibling”

Exciting progress toward gene therapy in hydrocephalus. And for targeting the choroid plexus as a platform for neurological intervention.