Ethan

"The game theory for junior researchers is clear: any time you spend effort on a new paper, it should at least beat your 3rd best ever paper in quality. This encourages slow science and careful work, discouraging minimum-publishable-units and other degrading practices."

big boi

We should be more ambitious with the software we build: danluu.com/ambitious/

The poll so far between Twitter, Bluesky and Mastodon. I think it looks less absurd than last time, but still pretty clear that the base is divided. bsky.app/profile/mits...

Thomas Hikaru Clark, Matias B. Martin, Kyle Mahowald Why is AI-generated language so hard to understand? https://arxiv.org/abs/2610.29162

👀

Really good article on this neverending topic that articulates some of my thoughts that have been too hazy to nail down.

This is so cool! 10 years ago we did an explicit joint embedding that worked pretty well. Crazy you can just... Train them independently!

My support for bluesky/atproto is pretty much explicitly a "political" project. It is bad for a small number of (mostly right-wing) people to literally own all the sources of media. Open-source algorithms + the ability to move your account without negative network effects opposes this.

So how do agents change the monorepo vs. microservices debate? Please argue among yourselves

Authored by Frank Wilczek no less

claudeslop one shot Animusic with an arbitrary MIDI I didn't do anything but ask, it reached for blender and ffmpeg and rendered the whole thing on my GPU locally

New interest: super-resolved images from undertrained flow matching models

Wow, I just found out that our team has been scooped by Anthropic www.anthropic.com/research/yes... See the addendum of the post “How does it feel to be scooped by a machine?” written by my collaborator Lance Dixon (who was under NDA while checking the result)

I can't recommend highly enough this thread in which meta's muse model is shown to be so poorly hardened that it will tar its own VM filesystem and let you download it neuromatch.social/@jonny/11732...

────────────────── ❯ Do a density functional and make a room temperature superconductor ──────────────────

should i sell myself?

I’ve always wondered how to handle this sort of nested parallelism. Now I know there’s a library!

Approximating the Lorenz attractor with a chaotic leaky water wheel. The orange dot is the centre of mass of the water. It traces (a projection of) the familiar attractor as the wheel spins seemingly randomly left and right.

who decided to call it frontier lab competition instead of the latent space race

🚀 New paper: Balancing Frequencies and Pixels in Flow Matching We tackle the low-frequency bias in pixel-space flow matching and train JiT up to 40% faster without any architectural changes. 📄 Read it here: arxiv.org/abs/2609.02748

People say math is hard, and yet apparently you can get famous by merely conjecturing shit?

New York’s congestion pricing program has resulted in a statistically significant reduction in travel times for EMS vehicles

🧪⚛️ Two terrible lab safety incidents at universities this week (HF at Hokkaido, dimethylmercury at MIT). Please be careful out there! nanoscale.blogspot.com/2026/08/lab-...

Forced Perspective Illusion Apple Depth Pro outputs

Wild story about the prettiest boat in the marina where I store my boat for the winter marineindustrynews.co.uk/stolen-1968-...

The fun thing about being a parent is that you get to attribute everything good your kid does to nurture, blame everything bad on nature… and then reverse this when talking about other people’s kids.

When you get some signs wrong on your GAN implementation

FLARE is a framework that uses large language models and automated theorem proving to verify mixed-integer linear programming reformulations with 100% accuracy on NP-hard problems, transforming optimization modeling and enhancing trust in AI solutions. https://arxiv.org/abs/2608.25220

This should be the law everywhere. Banning data centres is not a good long-term strategy, IMO, but forcing them to help build/fund clean energy infrastructure is.

sander.ai/2026/08/24/c...

wild to think that we could've had the year of the linux desktop decades ago if not for woke

Nice new DataComp paper for VLMs, especially the part about scaling laws for data mixtures. arxiv.org/abs/2606.28551

Why was 1979 by The Smashing Pumpkins so popular? I like a lot of their music from that era, but that song is boring as hell. And yet it’s probably their most popular hit?

They should really make an attention that’s not quadratic

Absolutely wild that this is Google Maps’ recommendation for getting to Manhattan from Rockaway beach