Gilles Louppe

Later we did early work using message passing graph NNs with @joanbruna.bsky.social and @johannbrehmer.bsky.social. Fully connected message passing NNs are pretty similar to transformers, but this was in 2017.
ml4physicalsciences.github.io/2017/files/n...

An unexpected shoutout from @kyunghyuncho.bsky.social in his NeurIPS keynote. @glouppe.bsky.social & I collaborated with him to bring ideas from natural language to high-energy physics back in 2016.
arxiv.org/abs/1702.00748

Super excited (and a bit nervous) to give a keynote tomorrow afternoon at #EurIPS. I'll be math, LLMs, and politics 🔥

Beautiful keynote by @gaelvaroquaux.bsky.social at @euripsconf.bsky.social .

🎉 "High-dimensional neuronal activity from low-dimensional latent dynamics: a solvable model" will be presented as an oral at #NeurIPS2025 🎉

Feeling very grateful that reviewers and chairs appreciated concise mathematical explanations, in this age of big models.

www.biorxiv.org/content/10.1...
1/2

What coding with an LLM feels like sometimes.

The Neurips test of time award goes to Faster R-CNN

Introducing gRNAde: our own little "AlphaGo Moment" for RNA design! 🧬🚀

📝: tinyurl.com/gRNAde-paper

Unlike proteins, RNA design has long relied on "wisdom of the crowd" (human experts) or the slow crawl of directed evolution — gRNAde changes that! 🧵👇

Scientific poster with dark background and two black holes illustrated in the center. The paper visualizes gravitational waves, and explains how parameter estimation is performed with DINGO. The standard DINGO model and the DINGO-T1 architectures are illustrated and results are shown. For example, it is possible to reanalyze the same event with different detector configurations with DINGO-T1, illustrated bz a corner plot.

1/ 🌀 New paper alert! We introduce Dingo-T1, a flexible transformer-based deep learning model for gravitational-wave (GW) data analysis. It adapts to different detector & frequency settings, improving inference efficiency and flexibility

🚀 #AI #MachineLearning #Physics #Astronomy #AcademicSky

Søren Hauberg shares how #EurIPS was put together in a matter of months

En Belgique on ferme des réacteurs nucléaires et on ouvre deux grosses centrales à gaz. Comme quoi le changement climatique et l'approvisionnement en gaz auprès de pays pas toujours bienveillants ne semblent pas des problèmes prioritaires pour cette majorité (et les précédentes).

Anecdotally (but I guess reflected in broader research) I was just talking to two friends who both work in software, and they complained that recent hires are so dependent on ChatGPT that these hires' own problem-solving skills are almost non-existent.

Yesterday, I did 2 AI conferences in Paris; different kind of vibes.
AdoptAI, a business show: pretty videos, promises about AI, 2.30€ coffee
NeurIPS@Paris, a research conference: equations, free coffee

AI business wouldn't exist without research. Let's not forget it and keep investing in research

Our Montefiore Science with AI lab will be at #NeurIPS2025 presenting 1 paper at the main conference and 3 papers at workshops. If you are attending, feel free to reach out with the crew to discuss science, AI, or just to say hi! (I won't attend this year unfortunately 🌱)

Les prix pour la première conférence en IA. Sans compter le vol jusque San Diego, la semaine à l'hotel, ou les frais sur place...

Finally, at the #CCAI workshop, Thomas will show how, without retraining, GenCast can be embedded in a particle filter for data assimilation. That is, no initial state x0 is required anymore, observations are sufficient to start generating realistic weather trajectories! arxiv.org/abs/2509.18811

At the same workshop, @orochman.bsky.social will discuss how neural solvers only approximately satisfy physical constraints (even if they are supposedly trained for that). Fortunately, simple post-hoc projection steps can help improve physical consistency significantly. arxiv.org/abs/2511.17258

At the #ML4PS workshop, @gandry.bsky.social and @sachalewin.bsky.social will present Appa, our large weather model for global data assimilation. Appa differs from other weather models in that it can do reanalysis, nowcasting and forecasting within the same framework. arxiv.org/abs/2504.18720

@francois-rozet.bsky.social (attending @euripsconf.bsky.social) will present the work he did at @polymathicai.bsky.social as an intern. In a nutshell, we find that emulating physics in latent space leads to better results than trying to generate in pixel-space directly. arxiv.org/abs/2507.02608

Our Montefiore Science with AI lab will be at #NeurIPS2025 presenting 1 paper at the main conference and 3 papers at workshops. If you are attending, feel free to reach out with the crew to discuss science, AI, or just to say hi! (I won't attend this year unfortunately 🌱)

The Mathematics of Artificial Intelligence: In this introductory and highly subjective survey, aimed at a general mathematical audience, I showcase some key theoretical concepts underlying recent advancements in machine learning. arxiv.org/abs/2501.10465

As we go into the Thanksgiving holiday, I wanted to express my thanks to my collaborators @johannbrehmer.bsky.social @glouppe.bsky.social, Juan Pavez, @smsharma.bsky.social. Recently, I was awarded the Pritzker Prize for AI in Science for work on SBI. That wouldn't have never happened without them.

Lots of interesting LLM releases last week. My fav was actually Olmo 3 (I love the Olmo series due to their full open-sourceness and transparency).
If you are interested in reading through the architecture details, I coded it from scratch here: github.com/rasbt/LLMs-f...

For some puzzling reason my (existing) Scholar profile cannot be found using Scholar itself, it's been like this for years 🙃 (search engines like Bing do find it however)

... and here I thought the new Scholar Labs would finally be able to search and find my Scholar profile 😥

It takes 225 million years or so for the Solar System to make one orbit around the galaxy. Sharks evolved long enough ago that they’ve done it twice.

What happens when you combine 10 years of brain data with one of the world’s fastest supercomputers?

A virtual mouse cortex simulation, thanks to a global collaboration.

🧠📈 https://alleninstitute.org/news/one-of-worlds-most-detailed-virtual-brain-simulations-is-changing-how-we-study-the-brain/

WeatherNext 2: Our most advanced weather forecasting model The new AI model delivers more efficient, more accurate and higher-resolution global weather predictions.

GDM WeatherNext 2

8x faster than v1, it can compute extreme situations and game out scenarios in one minute flat on a single TPU (as opposed to hours of supercomputer time for traditional algorithms)

will be available in all of Google’s weather apps

blog.google/technology/g...

Jaxley: differentiable simulation enables large-scale training of detailed biophysical models of neural dynamics - Nature Methods Jaxley is a versatile platform for biophysical modeling in neuroscience. It allows efficiently simulating large-scale biophysical models on CPUs, GPUs and TPUs. Model parameters can be optimized with ...

I am super happy to share that our project on training biophysical models with Jaxley is now published in Nature Methods: www.nature.com/articles/s41...

Oui c'était bien moi! Je me disais bien aussi que tu me disais qqch :-) au plaisir de se recroiser