The biggest bottle-neck to my personal code productivity:

The fact that OpenAI still hasn't pushed a Codex mobile app.

My paper of the year is Andrew Gordon Wilson's "Deep Learning is Not So Mysterious or Different". I'll be thinking this year about what family of functions (support) combined with what prior over parameters (inductive bias) can actually well capture drug discovery data including activity cliffs.

You learn a lot about the underlying system design of your apps when you run them in a low data environment.

A fundamental lesson of modern AI is that scale is essential: training bigger models on bigger datasets unlocks new capabilities. A fundamental lesson of AI engineering is that scaling up isn't trivial: it is not just a matter of spending more money and resources.

It shows

Very interesting article here vec2vec.github.io

Showing how the latent representations of two different vision models can be “translated” into each other via a universal “platonic” representation. As the authors note: interesting cybersecurity implications

Strong Platonic Representation Hypothesis: the universal latent structure of text representations not only exists, but can be learned and, furthermore, harnessed to translate representations from one space to another without any paired data or encoders.

Data contamination is all random forest needs Here's why we believe our Hermes prediction results are real

Got recommend this substack from Leash bio by a friend.

I think this is a masterclass in how to correctly split the data if there ever was one.

Respect your chemistry folks!

open.substack.com/pub/leashbio...

Contrasting photographs of the night-time skylines of Manhattan (left) and Nijmegen (right), with matching genome-wide association plots underneath each.

Not sure who came up with "Manhattan Plot", but in 2014 I coined the alternative term "Nijmegen Plot" (inspired by the Dutch town where I live) to describe underwhelming results from our earliest genome-wide association scans of language/reading traits.

Love these maps of "street-text sightings" in the Pudding's latest piece
pudding.cool/2025/07/stre...

Great blog post on rotary position embeddings (RoPE) in more than one dimension, with interactive visualisations, a bunch of experimental results, and code!

Can an AI model predict perfectly and still have a terrible world model?

What would that even mean?

Our new ICML paper (poster tomorrow!) formalizes these questions.

One result tells the story: A transformer trained on 10M solar systems nails planetary orbits. But it botches gravitational laws 🧵

Today's #RDKit blog post is a heartfelt plea for clearer communication.
greglandrum.github.io/rdkit-blog/p...

There is a new startup from China called Moonshot.

The original “moonshot” was the Apollo Program.

An AI based moonshot could be referred to as an “AI pollo” program.

“ai pollo” in Italian means something like “to the chicken”.

I was recently on a flight with free Wi-Fi for texting but nothing else.

Jokes on them: I can use Llama through WhatsApp now …

The new #RDKit blog post, inspired by a question from @valencekjell.com, looks at the impact of molecular size on similarity thresholds.
greglandrum.github.io/rdkit-blog/p...

Yay for @pschwllr.bsky.social and @mlederbauer.bsky.social (and all your co-authors who aren't on BlueSky yet) 🥳

This #dataset is a prime example of #GoodData, and it ties nicely with what @clarakirkvold.bsky.social and @grynova.bsky.social were talking about a few weeks ago in their #JournalClub

I've got a joke about Osysseus. I got lost on the way to the punchline...

smell rights. in the US, Hasbro has a tradmark for the smell of Play Doh.

change my mind:

bruot RIS to Bibtex converter is the best website ever built.

www.bruot.org/ris2bib/

If anybody out there working on antimicrobial resistance (AMR) and needs some motivation on this gloomy New England Monday.

I think the ranking of things which are hard to predict goes:

1. The stock market.
2. LaTeX figure placement.
3. The meaning of life.

#booksky

Cheminformatics family businesses be like

Just to clarify: I’m washing my wants twice now! Not to cause any concern here.

One of the surprising things about working in a microbiology lab is that you become more worried about washing your hands before using the restroom rather than after.

I think the thing I'm most excited to see over the next ~10 years of #dataviz is web-based content that interweaves long-form text and modular interactives.

Not as heavy as scrollytelling and not as aimless as a dashboard, but something in between.

This is what I was going for with the QR project!

change my mind:

bruot RIS to Bibtex converter is the best website ever built.

www.bruot.org/ris2bib/

You know volatility is going crazy when sitting down to write a PAC proof about the sampling efficiency of an active-learning algorithm feels like a therapy session.

At least math hasn't changed over the past 12 months ...

Not me accidentally typing `squeue` into the Facebook chat.