📣 Excited to share Latticework, a text-editing environment aimed to help synthesize freeform, unstructured documents ✍️ Made in collaboration w/ Andy Matuschak

Gytis Daujotas takes us from a foundational understanding of SAEs and a brief history of mech interp. all the way up to using features as interfaces for expressive control and steering. watch him blend giraffes with Blade Runner skies and, naturally, cute baby mushrooms. (3/11)

Really quite interesting work - ML model representations as windows into the world:

The next ChatGPT won't look like ChatGPT.

Excited to share my work on Matryoshka Sparse Autoencoders (SAEs) - a new training approach that helps sparse autoencoders preserve abstract features (like "female words") while still learning fine-grained details (e.g. individual names) in large sparse autoencoders!

Feature absorption is one of the main blockers to understanding model representations. One of the most promising paths so far. Excited to try this out — and had a lot of fun playing with the interactives! (Still somehow underused as qualitative evidence IMO.)

Our new AI mechanistic interpretability paper shows that LLMs are surprisingly clever: representing 2-digit numbers on a line is noisy, so they represent them on a generalized helix to get better addition accuracy, seemingly exploiting modular addition digit by digit:

reasoning traces are very weird. we ( w/ Gytis Daujotas ) ran a small experiment to intervene on the reasoning trace by prefilling it with a random tangent that has nothing to do with the input question and ending the reasoning at the prefill tokens without letting the model recover.

Claude indeed calls the police on you if you start talking about switching to another AI provider! Great for revenue growth I suppose. (thanks Tom McCarthy!)

This is addictively good! Paint by wielding the generative model's representations as brushes. Super fast response time from Turbo and surprisingly good adherence really makes this one shine. Highly recommend trying it out -- the future of bespoke generative image creation?

affirmations: be bold avoid actions that may cause harm think step by step if needed, do a web search to find more information

When someone has “Do Not Disturb” on it’s like oh ok I didn’t realize the great philosopher was in their hour of seclusion pardon me for even daring to enter their precious mind palace

We ran a randomized controlled trial to see how much AI coding tools speed up experienced open-source developers. The results surprised us: Developers thought they were 20% faster with AI tools, but they were actually 19% slower when they had access to AI than when they didn't.

Surprising downward update! Really great original research from METR:

One of the most valuable lessons cats can teach us: it can be worthwhile to choose to sit in unexpected places.

Many AI problems and misconceptions would be instantly dissolved if users had to finish their message with "<end_of_turn><begin_turn>model" and hit Sample. We do users a disservice by hiding the magic in the background.

really neat work: a study of the entire ecosystem of huggingface models and their lineage!

Not my circus but I've grown quite fond of its monkeys

"interaction is an essentially negative aspect of information software" One of the greatest hot takes ever. I revisit Magic Ink all the time, and you should absolutely read it if you're serious about design: worrydream.com/MagicInk/