A new N&V on the Article by del Rio Chanona & co-workers: nature.com/articles/s4428… "Machine-learning optimization of 3D-printed flow-reactor geometry" by Bennett and Abolhasani. Abolhasani Lab Link here ($): nature.com/articles/s4428… ➡️ Free read link: rdcu.be/dP9kW

Tom Savage Nature Chemical Engineering Nausheen Basha Omar Matar Antonio del Rio Chanona It is great working with Tom Savage in creative areas & more impressive to see his paper in Nature Chemical Engineering on ML-assisted approach to chemical reactor design. High level maths (!), but Anil Ananthaswamy ‘Why Machines Learn’ helps me grasp initial understanding: anilananthaswamy.com/why-machines-l…

Stop discarding your old gradients! Introducing AdEMAMix, a novel (first-order) optimizer capable of outperforming Adam. Let’s have a thread on momentum and the surprising relevance of very old gradients. A joint work with David Grangier and Pierre Ablin #ml #optimization 1/🧵

Wrote this blog post a while ago specifically for AI, interesting to think about ways to send a different message to humans / AI, chose RC4 encryption, but with computer use I’m not even sure this is robust… sav.phd/posts/for_ai/

Massive, John Trefry. Wild book, probably the closest it feels to be in the latent space of a recursively improved LLM, this thing is not made for us. Physicality gives it a presence, it will now haunt my bookshelf.

I've written about what an LLM cultural victory might look like and what this means for Deepseek with an irresistibly priced new reasoning model hosted in China. Have OpenAI ceded ground? Is 'chatting' with an LLM an inherently Western construct? +more blog.sav.phd/p/has-an-llm-c…

That! The Haiku and the Sonnet. While in Japan I've written about an alternative way of assessing the creative outputs of LLMs. open.substack.com/pub/trsav/p/th…

There's some interesting ongoing discussion about how much of a "diffusion model" this really is, but anything that even slightly challenges the autoregressive hegemony is a refreshing change!

Just used Claude Code to implement and benchmark improved convergence methods for my python based, LLM-ready chemical process simulator... insane. Burns through tokens but vibes-based coding is truly here... a thousand moats cried out all at once and were suddenly silenced

Salad cream bosh.