off to #emfcamp for the weekend! if you are an ethics reviewer for neurips please do your reviews so I don’t have to chase you while I’m in a field

The #ICLR2026 fomo is already intense

"Pancreatic cancer mRNA vaccine shows lasting results in an early trial: Scientists caution that more research is needed, but nearly all of the patients who responded to the personalized vaccine are still alive six years later."

Pancreatic cancer mRNA vaccine shows lasting results in an early trial

Scientists caution that more research is needed, but nearly all of the patients who responded to the personalized vaccine are still alive six years later.

nbcnews.com

Hey, I wrote a thing about AI in astrophysics ergosphere.blog/posts/the-ma...

the greatest joy of being a computational scientist is having the computer work for you while you do something else

“Interpretability plays a special role in machine learning because instead of focusing on making the AI smarter, we focus on improving human insight. I think this is the most important category of interpretability research, and we do not do enough of it.” 😎😎😎

At the #Neurips2025 mechanistic interpretability workshop I gave a brief talk about Venetian glassmaking, since I think we face a similar moment in AI research today. Here is a blog post summarizing the talk: davidbau.com/archives/202...

remember to always include a QR code on your poster. spotted at #eurips

What coding with an LLM feels like sometimes.

found a file from PhD days with the FORTY-EIGHT ways "ACE inhibitor" was encoded in the EHR system we were working wth

finally got around to booking my travel for #EurIPS2025! Looking forward to connecting with the European ML scene in Copenhagen

uv is so good

Eugene Vinitsky 🍒@eugenevinitsky.bsky.social · last yr.

uv is now stable enough that I'm switching my course over to using it. No more conda, we will create a generation of uv evangelicals

Some papers really have a good intro

The more rigorous peer review happens in conversations and reading groups after the paper is out with reputational costs for publishing bad work

I'll admit, I was skeptical when they said Gemini was just like a bunch of PhDs. But I gotta admit they nailed it.

what is the purpose of VQA datasets where text-only models do better than random?

lads can we stop

quick diagram of Bluesky’s architecture and why it’s nicer here

it's frustrating how inefficient review assignments are: we target a minimum number of completed reviews per paper but in accounting for inevitable no-shows, some people end up doing technically unnecessary (if still beneficial) reviews

New work from my team! arxiv.org/abs/2507.12950 Intersecting mechanistic interpretability and health AI 😎 We trained and interpreted sparse autoencoders on MAIRA-2, our radiology MLLM. We found a range of human-interpretable radiology reporting concepts, but also many uninterpretable SAE features.

Insights into a radiology-specialised multimodal large language model with sparse autoencoders

Interpretability can improve the safety, transparency and trust of AI models, which is especially important in healthcare applications where decisions often carry significant consequences. Mechanistic...

arxiv.org

We're excited to announce a second physical location for NeurIPS 2025, in Mexico City, which we hope will address concerns around skyrocketing attendance and difficulties in travel visas that some attendees have experienced in previous years. Read more in our blog: blog.neurips.cc/2025/07/16/n...

During the last couple of years, we have read a lot of papers on explainability and often felt that something was fundamentally missing🤔 This led us to write a position paper (accepted at #ICML2025) that attempts to identify the problem and to propose a solution. arxiv.org/abs/2402.02870 👇🧵

ExplainableAI has long frustrated me by lacking a clear theory of what an explanation should do. Improve use of a model for what? How? Given a task what's max effect explanation could have? It's complicated bc most methods are functions of features & prediction but not true state being predicted 1/

Pleased to share our ICML Spotlight with @eberleoliver.bsky.social, Thomas McGee, Hamza Giaffar, @taylorwwebb.bsky.social. Position: We Need An Algorithmic Understanding of Generative AI What algorithms do LLMs actually learn and use to solve problems?🧵1/n openreview.net/forum?id=eax...

This is cool arxiv.org/abs/2505.18235. Linear representation hypothesis discourse needs more differential geometry i m o

The Origins of Representation Manifolds in Large Language Models

There is a large ongoing scientific effort in mechanistic interpretability to map embeddings and internal representations of AI systems into human-understandable concepts. A key element of this effort...

arxiv.org

🚀 After two+ years of intense research, we’re thrilled to introduce Skala — a scalable deep learning density functional that hits chemical accuracy on atomization energies and matches hybrid-level accuracy on main group chemistry — all at the cost of semi-local DFT ⚛️🔥🧪🧬

Limited time offer: I get an email every time someone fills in this form, so act now to add to the chaos of my inbox before I figure out how to turn this off

Stephanie Hyland@hylandsl.bsky.social · last yr.

NeurIPS 2025 call for ethics reviewers neurips.cc/Conferences/... #neurips2025

A strange thing about living near Duxford is having WW2-era planes flying overhead on a regular basis.

90% of doing science is being open to new ideas.

NeurIPS 2025 call for ethics reviewers neurips.cc/Conferences/... #neurips2025

I am disappointed in the AI discourse steveklabnik.com/writing/i-am...