Will be at CoRL next week with the @FlexionAI team. Come say hi!
Daniel Dugas
Verified account Flexion@DanielDugas14
Robotics & AI research, META FAIR - Locate 3D, VJEPA2, World Models - PhD at ETH Zurich.
288 Following 504 Followers 804 Posts
Latest posts 5
Soon this will be a question in MLE interviews and everyone will have to pretend like they would obviously have solved this bug from first principles
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:View quoted post on X
good example of the generality of reasoning (unless somebody thinks OpenAI included crawling data in their training 🐛)
Me: Put the green box on target, clearing obstructions.
Astra: You got it, boss 🫡
(Purely python solution generated in a red teamed sandbox using no any other tools, with $20 budget)View quoted post on X
Huge post, and probably paper-of-the-year for me
tl;dr: we _can_ understand what neurons do in ML models!
They are not black boxes, at least not as much as we thought.
tl;dr: we _can_ understand what neurons do in ML models!
They are not black boxes, at least not as much as we thought.
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we hadView quoted post on X
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com
so this is why i haven't seen @theproteindance in a while
WE TALKED TO A CHINESE AGENT FLEET. LIKE WE LITERALLY TALKED TO IT. FULL CHINESE AGENT FLEET REPORT OUT NOW, READ IT OR I WILL THROW A ROCK AT YOU. GET IN swarmcha.se/china-fleet?v=2View quoted post on X
