Alexander Hoyle

I have, and he is almost certainly incorrect about Anthropic‘s profitability

I'm so sorry to hear this. Your work is consistently thoughtful, distinctive, and impactful. I have so appreciated your mentorship and perspective, and am sure you will find a place where your voice is valued

hoping to develop something like this :)

Philosophy of Language and Computation II, Spring 2025 | Rycolab This graduate class, partly taught like a seminar, is designed to help you understand the philosophical underpinnings of modern work in natural language processing (NLP), most of which centered around...

You might find this interesting rycolab.io/classes/phil...

Already a thing!

www.cs.umd.edu/~jbg//docs/2...
aclanthology.org/2023.acl-lon...
(I’m sure there’s more)

I appreciate the shout-out to our work showing that LDA tends to work better, although it would have been nice to include it as a comparison :)

I thought these were deeper explorations of LM “understanding“ than the Stochastic Parrots paper:
direct.mit.edu/coli/article...
julianmichael.org/blog/2020/07...
direct.mit.edu/coli/article...

Also see Will Merrill’s work for more formal treatments

Yeah, that denominator was what I was wondering about, since my understanding is usually it's something like 5-10 onsites per season?

I did a "toe-dipping" last cycle (first year of postdoc), and I wasn't sure how to consider the proportion of onsites I got relative to screenings

I was discussing first rounds with a colleague recently---do you have a sense of the typical "conversion" rate to onsites?

[Also, sorry for dropping into your mentions for the second time today]

Does it create good diagrams?

Oh amazing! Do you have the datasets you used to test it? I’ve been playing around with a project related to this topic

This is super great---a nice, illustrative collection of measurement issues. Thanks for sharing

FWIW, auto completion models are worse because they are cheap and run all the time. the larger models/integrated systems (like Claude code) are significantly better

I think you are relying on two assumptions: (a) they used a commercial LLM to begin with and (b) that commercial LLMs train on all data that is passed to them

I don't know about (a), but (b) is not the case when using the API endpoints for the major providers

platform.openai.com/docs/concepts

Sincere question: how do you distinguish your paper being passed to a commercial LLM from your paper being indexed by commercial search engines?

It’s also weirdly dependency-averse (maybe it has been RLHF’d away from using outdated/old libraries?). I was vibe coding an iOS app and it reimplemented a full photo editor. I know nothing about the Swift ecosystem, but I figured there had to have been some existing library (there is of course)

Now I'm late to replying---thanks, this means a lot coming from you! For future work, are looking at ways to apply it to mech interp, but it's still in the early stages. I'm also interested in the "reverse" problem: transformations that improve similarities along a given dimension (e.g., syntax)

Happy to be in the second round!

Was just browsing “For You“ and didn’t expect to see one of ours here! Some here I haven’t seen, thanks for the pointers :)

Obsidian does calendar management?

Interesting—what’s the difference from Claude Code/Codex?

Ah l see you wrote about this in the original thread, apologies!

that’s fair, the author is definitely putting forward a position. Your argument is that cog sci anticipated these developments somewhat, or? (May have not gathered it in the thread)

The Case That A.I. Is Thinking ChatGPT does not have an inner life. Yet it seems to know what it’s talking about.

Just read a nice New Yorker piece that interviews mostly cognitive scientists/neuroscientists who are (mostly) very impressed:

www.newyorker.com/magazine/202...

It’s unfortunate because I’ve found LDA (mallet/tomotopy, not gensim) is actually very good for tweets

my poor heart 💔

Realizing neither of you have John Williams' Stoner, which is a beautiful book

I enjoyed Changing Places, and there's two more in the "campus trilogy". I don't recall much about the third (if I even got to it?), but Small World was fun---more in the left quadrant

It's funny, when I've spoken to some ML people---including, just recently, Dave Blei himself---the question is often, "who's still using topic models?" Meanwhile in DH/Social science, it's "please, enough topic models!"

Some collections of work in these directions:

openreview.net/pdf?id=gJcEM...
arxiv.org/abs/2210.13382
iclr.cc/virtual/2025...
www.worldmodelworkshop.org