Ben Kompa

Very cool to see @LilaSciences showcased during @JensenHuang's keynote at GTC Berlin! Full keynote: youtu.be/PvFjbPYEOZI?t=3784 @nvidia's dedicated video on Lila: youtube.com/watch?v=Y39uub4n…

twitter is the bloomberg of ai research

Teams are already post-training Nemotron 3.5 Super VL for tasks in document intelligence, cybersecurity, finance, scientific experimentation, and travel. The chart below shows how accuracy improves when the model is post-trained for each task.

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 👇

I'm building Scientific Superintelligence at @LilaSciences so that we can cure all diseases This requires self-play against reality, where an AI agent can formulate hypotheses, design and execute experiments, and refine its reasoning through empirical feedback If you're interested in working together on this, DM me the coolest thing you've built or discovered

gptimages Be like