🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

I'll be in the US in March (PhD visit days): - Boston: Feb 27 to Mar ~7 - Berkeley: Mar ~7 to Mar ~18 - Pittsburgh: Mar ~18 to Mar 24 would love to meet up if you're around :) (would also appreciate recs for ppl you think I should meet -- or takes on grad school...)

very interesting that every frontier lab interp team is working on sparse autoencoders (SAEs) and ~ no one in academia is

mechinterp people: does anyone have a good (formal?) definition of 'feature' that doesn't assume the linear representation hypothesis? like, if I have some points in high-dim space, what makes them "the composition of several features" as opposed to "some random points"

fyi the real reason i've been ignoring you is: - i want to reply - i want to be able to give you the attention and focus you deserve - i never feel like i have enough energy to properly do that

maybe the most exciting interp result I’ve seen all year (if it ends up being true for interesting reasons): a meaningful step towards uncovering the type of the residual stream

on reading ml papers:

perhaps growing up is realising that 'growing up' was a comforting lie

On Reddit's statistics forum, the most common question is "What test should I use?" My answer, from 2011, is "There is only one test" allendowney.blogspot.com/2011/05/there-…

Who should I meet in Cambridge? (You?)

in slight defense of mathiness / mathematical notation in ML research papers: a thread (twessay?)

SHA-256: 218cebed21f2e8514df2ea1e4caca39750349cf30804995d5d577f08afc5855a

Announcing Transluce, a nonprofit research lab building open source, scalable technology for understanding AI systems and steering them in the public interest. Read a letter from the co-founders Jacob Steinhardt and Sarah Schwettmann: transluce.org/introducing-tr…

The tragic suicide of Sewell Setzer III shows our generation has become unwitting test subjects in a vast, unregulated AI experiment. That's why we're launching Center for Youth and AI with our Generation AI Survey in TIME. A thread: (1/10)

This is a really creative and well-executed paper on using "black-box interpretability" methods to understand and control model cognition. Especially impressed by the many applications explored IMO this is an important direction; this paper sets the field on an excellent path!

LLMs are dramatically worse at ARC tasks the bigger they get. However, humans have no such issues - ARC task difficulty is independent of size. Most ARC tasks contain around 512-2048 pixels, and o3 is the first model capable of operating on these text grids reliably.

I've recently learned about Algebraic Positional Encoding from Bruno Gavranović and isnt this the coolest breakthrough in mathematical approaches to transformers in the last few years arxiv.org/abs/2312.16045

Take a break from arxiv/LW/AF. Sit in the woods with a random textbook and mull new ideas away from interp community lockstep. Diverge. Don’t compete with a saturated subtopic, maybe you’ll get to take weekends off. Premature overinvestment comes from monoculture.

The slowly-unfolding premise of the Good Place is that everyone is damned. They are damned because they participate in the modern world; they buy from sweatshops, they eat chocolate, they fly in airplanes while the poorest people in the world see their harvests fail thanks to