Romain Lacombe
@rlacombe
Researching AI/ML & ChemE @Stanford ⚡️🧪 | Past: CEO @Plume_Labs (acquired by @AccuWeather) 🌦️
☀️+☁️=❤️!
We’re excited to announce we’ve joined @AccuWeather, the global leader in weather forecasting, to focus on AI for climate. This is an extraordinary opportunity to amplify our impact at planetary scale, and we couldn't have done it without you. Thank you! blog.plumelabs.com/2022/01/24/a-t…
This is awesome. Computational urban ecology a.k.a “we ran around late at night to record wild rats” 🤘
New preprint! tl;dr — We ran around late at night to record wild rats in NYC and figured out how to quantify their behavior and environment. 🧵 w/ Dima Batenkov, @zamakany, @e_mackevicius biorxiv.org/content/10.110…
"I've raised millions of dollars to do X" Very cool Now pick up the shovel, and pray to fucking god you chose the right thing That, and the hard work, is the hard part IMO
This. When people show you who they are, believe them the first time.
People who cheat on their spouses are more likely to cheat at work, too. Evidence: Marital infidelity predicts 2-3x greater odds of professional misconduct among CEOs, financial advisors, and police officers. Integrity is not a 9-5 job. Character counts in every part of life.
Important analysis of Humanity’s Last Exam: 30% of bio/chem questions found invalid! We’re benchmarking these models on questions that aren’t even correct.
HLE has recently become the benchmark to beat for frontier agents. We @FutureHouseSF took a closer look at the chem and bio questions and found about 30% of them are likely invalid based on our analysis and third-party PhD evaluations. 1/7
No man ever steps in the same newsfeed twice.
ahhhh yes good morning let’s open up the Infinite Information River for a nice 20 minutes before getting out of bed or thinking any original thoughts at all
Very cool paper from FAIR Chemistry presented at @icmlconf — don’t sleep on n-th-differentiable MLIPs!
That's a wrap for ICML 2025! 🎉 🇨🇦 Fun to present/discuss a bunch of recent work. In particular, Ray Gao and I presented: Learning Smooth and Expressive Interatomic Potentials for Physical Property Prediction, which received an oral 🎤 / spotlight poster 🖼️ 🙌 A huge shoutout…
And it’s a wrap for @icmlconf 2025! Thanks for a wonderful week of ideas, talks, papers, and new friendships. Vancouver is stunning! 🤖🎓🇨🇦
🚨 An anti-scaling law for reasoning models! 🚨 ✈️ At @icmlconf presenting our latest paper: 📝 ‘Don’t Think Twice! Over-Reasoning Impairs Confidence Calibration’ ⏳TL;DR: Increasing the thinking budget of reasoning LLMs makes them over-confident!
Very interesting! You may both be interested in this setup where we show increasing reasoning budget (CoT length) impairs model calibration on confidence related tasks (R2FM workshop tomorrow). x.com/rlacombe/statu…
Tomorrow’s the day! Presenting our anti-scaling law for reasoning models at the @ICMLconf reliable foundation models workshop. Come see us if you’re in town!
“Here’s to the crazy ones, the misfits, the rebels. The closed 3-manifolds homeomorphic to a 3-sphere.” 🎶
Some people worship athletes or actors. As for me, I admire great scientists. The guys who developed quantum mechanics, differential geometry, etc. Sure, Brad Pitt is a good-looking guy, but Grigori Perelman proved Poincaré's conjecture.
“A founder goes down with the ship.” ❤️
To put it mildly, the past week at Windsurf has been crazy. There have been a lot of different rumors and reports, so I want to share a transparent account of how it actually went down. Before I start, I just want to say that Varun and Douglas were great founders and this…
Tomorrow’s the day! Presenting our anti-scaling law for reasoning models at the @ICMLconf reliable foundation models workshop. Come see us if you’re in town!
📉 Most surprisingly (to me), we uncovered an anti-scaling law! 💭 Increasing their thinking budget makes reasoning models **over-confident** and impairs their confidence calibration! 📝 Paper: openreview.net/forum?id=e7G5a…
“AI is a telescope for mathematicians” 🔭 - Henry Kvinge, at @icmlconf
