✨ #NeurIPS2025 paper: Scalable Feature Learning on Huge Knowledge Graphs for Downstream Machine Learning Combining contrastive learning and message passing markedly improves features created from embedding graphs, scalable to huge graphs. It taught us a lot on graph feature learning 👇 1/10

I don't know how to put it eloquently into words but my product manager thought leadership take is: I'm gonna bet on the company that delivers 100 features customers would like over the company that delivers 10 features they marty cagan'd their brains into believing are the best features

Thank you for choosing Amtrak (Not that you had a choice)

I’m feeling this drumset

A model for detecting LLM generated text is under review at #ICLR2026. The website using this model detects one of the reviews for this paper as LLM generated ... 🍿🍿🍿🍿🍿🍿🍿🍿🍿

I wrote about "chart go up" arguments on social media: varunprajan.github.io/blog/chart-g...

Fish sorting line.

Breaking: we release a fully synthetic generalist dataset for pretraining, SYNTH and two new SOTA reasoning models exclusively trained on it. Despite having seen only 200 billion tokens, Baguettotron is currently best-in-class in its size range. pleias.fr/blog/blogsyn...

I’m honored to serve 🫡

“…you’re much better served as a student by getting a degree in an “irrelevant” subject that motivates you to put in real, high-quality work than by shuffling through something more relevant on a surface level doing only the bare minimum to meet the requirements.” open.substack.com/pub/chadorze...

Latest newsletter (free) covers the newest AI video models, some misc 3d tools, an OCR model bonanza, a few games links including a fun new AI-assisted 2d tile project, my fairytale hunt project, and more... arnicas.substack.com/p/titaa-72-f...

In astronomy, an analogous effect is the Malmquist bias. "Usually heard in the phrase, 'You fool, you forgot to account for the Malmquist bias.'" -- Ethan Vishniac en.wikipedia.org/wiki/Malmqui...

As promised, here's a longer explanation of why you should vote YES on all the NYC ballot initiatives EXCEPT #6 ✅ YES on #1 ✅ YES on #2 ✅ YES on #3 ✅ YES on #4 ✅ YES on #5 🚫 NO on #6

(translating Icarus for the modern age) Oh he'd been cooking alright, but then he was cooked

Choudhury and Kim et al., "Accelerating Vision Transformers With Adaptive Patch Sizes" Transformer patches don't need to be of uniform size -- choose sizes based on entropy --> faster training/inference. Are scale-spaces gonna make a comeback?

sometimes I feel like I work for the government

If we’re lucky, this fever dream has just been a long, distasteful, guerrilla marketing campaign for a Diarrhea Planet reunion tour

also bluey and numberblocks

Pro tip: For good Halloween vibes, use non-normalized RoPE on images larger than your training resolution and larger than the composite period of some of the RoPE-rotations. You might get scary ghost structures in your features.

Imagine paying $10M to live in Hudson Yards

If I ever leave this city, it will 100% be due to the price of housing rather than taxes.

“…on current trends the grid can grow sufficiently to accomodate all of these loads and some residential electrification as well, and do so at a rate of growth only about half of that which was seen between 1975 and 2000.” energyandstuff.substack.com/p/sam-altman...

btw what I did here was I took a data set of ChatGPT interactions collected in the wild and reversed the "assistant" and "user" tags. fine-tuned llama 8B on some of that data and gave it the ability to message you first. 😊

Good article from @pedramnavid.com about building a successful #DevRel team. databased.pedramnavid.com/p/reflection... Measured and considered, and lots of wisdom. A++++ would read again.

Lot's of work being done on columnar file formats lately. I count 5 new formats so far (Lance, Nimble, Vortex, FastLanes, F3). It's definitely something we follow at LanceDB and it can be confusing to track. So here is my very biased head-canon (trying to stay positive)

Great thread

I was reading Zen and the Art of Motorcycle Maintenance last night, so it’s a bit of a surreal dichotomy to compare Quality to this Slop post

This week on Counting Stuff. Layoffs! I'm in search for a job now! It sucks but here we are! Now's the ideal time to give me money in exchange for time or a project. Or if you just have leads in quantitative/data/researchy positions across NYC or remotely www.counting-stuff.com/looking-for-...

LayerD: Decomposing Raster Graphic Designs into Layers Tomoyuki Suzuki, Kang-Jun Liu, Naoto Inoue, Kota Yamaguchi tl;dr: repeat(matting->inpaint). Apache 2 license arxiv.org/abs/2509.25134

“Is 90% of code going to be written by AI? I don’t know. What I do know is, that for me, on this project, the answer is already yes.” lucumr.pocoo.org/2025/9/29/90...

Incredible that it took this long github.blog/changelog/20...

Spotify should let me share songs with a specified start timestamp (like YouTube does) so that I can link directly to my favorite BTBAM parts

Status check: First one became a talk. Second one in progress. Third one will never happen because nobody deploys their own models anymore. Fourth one who knows. Fifth one I hope to scrap in favor of one day open sourcing the code that solved this blog post www.ethanrosenthal.com/2024/11/19/y...

New lesson! "An Arsenal of AutoEncoders (AE, VAE, Conditional VAE, VQ-VAE)" Colab link: github.com/drscotthawle...

This I why I can’t leave this city

I wrote this post 8 years ago, but I think now is the time that it’s actually coming to fruition thanks to LLMs and diffusion. We’re in a golden age. www.ethanrosenthal.com/2017/03/20/a...

In the age of AI, LLMs and code assistants what does it mean to be an open-source project that is welcoming to newcomers? As a newcomer, is it enough to use AI to solve a good first issue and submit it as a PR? As a maintainer, what should you do with PRs like that?

No biggie, just an original Signal Problems tote 💅

Is scale all you need? Or is there still a role for incorporating domain knowledge and inductive bias? While I was in Heidelberg, I took some time to write a short essay on this question called "The Bittersweet Lesson". theoryandpractice.org/2025/09/The%... #HLF25

The Stanford University researchers think the future of pre-training lies in the algorithms that will best leverage infinite compute. They find simple recipes that improve the asymptote of compute scaling laws to be 5x data efficient, offering better perf w/ sufficient compute.

Tbh I assume that everyone using LLMs in prod is building a failover system. The team wrote about ours blog.gradient-labs.ai/p/building-r...

I set out to avoid the News by writing about academia, but it got dark: The function of higher education is split between credentialing and actual education, and the angst around academia mostly oozes from the gap between them: open.substack.com/pub/chadorze...

I’ve seen semantic IDs pop up but never bothered to actually look into them. This write up from @eugeneyan.com is a great intro that also illustrates why they’re pretty interesting for mixing recsys and LLMs eugeneyan.com/writing/sema...

Got a bit tired of wading through lazy AI-generated interaction tests in critical systems that don’t actually do anything. Overzealous enthusiasts messing up the test suite w/ useless tests has become my new pet peeve. Mocking libs make it worse since LLMs love them. rednafi.com/go/test_stat...

Whoever created all of the automation in bathrooms clearly didn’t have kids

Nice blog post on query rewriting using LLM-generated buyer profiles www.etsy.com/codeascraft/...