very excited to finally share this work. a simple ODE predicts MLP training dynamics*! if you're working on the fundamental science of ML training, I think you're gonna want to check this out.
very excited to finally share this work. a simple ODE predicts MLP training dynamics*! if you're working on the fundamental science of ML training, I think you're gonna want to check this out.
1/ We still don’t have a simple picture of how even 2-layer MLPs learn features. In a new paper with @berkott3 and @learning_mech, we propose a simple, solvable model of shallow MLP training, and find that real MLPs roughly follow it. 🧵
I wrote the songs. they are, as plainly as I can put it, the sadness, horror, longing, and loss of our time. I'm very proud that most of them are also funny. we laughed a lot making them, and after I stepped off stage, I cried.
lyrics are here: jamiesimon.io/lyrics
this week, @ericjmichaud_ and I performed a concert of original folk songs about AI at @imbue_ai. if there's one thing I've ever done or said that I'd hope would be widely shared, it's this.
youtube.com/watch?v=H6iXnYYs…
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 had
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/
hell yessssssss
strong rec to reach out to Florentin if you're seeking to get into learning mechanics / first-principles DL theory and wanna do a PhD or postdoc.
I'm delighted to announce that I'll join @MITEECS as an Assistant Professor next Fall! I will be recruiting graduate students this cycle, please apply if you're interested in the science of deep learning! Despite the ambient doomerism, now is a great time to do research.
Have several thoughts about the Goodfire blog.
First one is that the field of mechinterp has had (imo) a pretty huge structural blind spot so far as being currently unable to answer the question of "why and how did these structures get here" as opposed to just providing descriptions of "what are the structures that are there." Really impossible to make sense of the latter in a deep way without understanding the former. The post is a move in the right direction for that field (assuming the results hold up, perhaps a big ask at this point).
No deep understanding of these nets will be possible until the causal map between data and model is elucidated. This work doesn't get there yet but by putting training dynamics into the picture it gets a step closer.
3/ If you're working on making AI more capable/intelligent/etc., you've probably had serious moral doubts niggling in the back (or front) of your mind. Maybe you don't know what to do -- but you must do something. Come talk about it! We'll end with an open discussion.
2/ For a long time, I've seen the cowardice and fear and greed and fatalism of Silicon Valley and wondered what I can do. I've come to realize that researchers have influential voices, and you can do a lot with yours. This is what I'm doing with mine.
1/ On Oct 22, @ericjmichaud_ and I are doing something a little crazy: we're performing a public folk concert of original folk protest songs about AI, SF, and tech culture. Link below, motivation in thread:
luma.com/n6j7vvod