The response has been surprisingly muted on GPT-5. My sense is that it’s an improvement, but not groundbreaking, unlike the delta for GPT-2 -> 3 ->4. It may provide more evidence that scaling is exhausting itself as a paradigm. Time to go back to ideas.

Yoshua Bengio is now the most cited scientist of all time.

📢Interested in doing a PhD in generative models 🤖, AI4Science 🧬, Sampling 🧑‍🔬, and beyond? I am hiring PhD students at Imperial College London Imperial Computing for the next application cycle. 🔗See the call below: joeybose.github.io/phd-positions/ And a light expression of interest:

The most important skill for a researcher is not technical ability. It's taste. The ability to identify interesting and tractable problems, and recognize important ideas when they show up. This can't be taught directly. It's cultivated through curiosity and broad reading.

9 of our papers are going to NeurIPS in San Diego. More bragging will follow shortly.

🚀Curly Flow Matching has been accepted to NeurIPS Conference 2025! Massive shot out to my awesome collaborators Lazar Atanackovic Viggo Moro Kacper Kapuśniak İsmail İlkan Ceylan Michael Bronstein Joey Bose Alex Tong Stay tuned for the camera-ready version+code soon 📸 See you in San Diego! 😎

Check out our NeurIPS spotlight! 🤩

Super excited to announce our recent work was accepted to NeurIPS 2025! 🌟 We introduce Prose, a 280M-parameter transferable normalizing flow proposal for efficient sampling of unseen peptide sequences 😮 Many thanks to the fantastic team!

Come see our poster at COLM today!

Attention Sinks and Compression Valleys in LLMs are Two Sides of the Same Coin "Through experiments across several models (410M–120B parameters), we confirm that when the beginning-of-sequence token develops extreme activation norms in the middle layers, both compression valleys

gLSTM extends xLSTM to a graph neural network architecture: arxiv.org/abs/2510.08450 "gLSTM mitigates sensitivity over-squashing and capacity over-squashing." "gLSTM achieves comfortably state of the art results on the Diameter and Eccentricity Graph Property Prediction tasks"