A. Feder Cooper

(lucky for everyone that I'm too lazy to write a blog post))

Yes, I have published at that track before, and related ones. But I'm not eager to again. Getting into that is maybe worth a blog post.

No I did not write/submit this paper to the ICML position paper track. Like many (but of course not all) papers submitted there, I think this is at most a blog post (where "at most" is a very generous upper bound, because the ~300 characters above almost certainly are enough).

Position: ML conferences should consider removing the position paper track

(...and just acknowledge that every scientific paper is articulating at least one position)

(This is all to say, I've been shocked at some of what I've heard coming out of industry. My assumption used to be that they knew a lot more about this than they seem to.)

I think partially yes. There definitely are full-time applied and research people working on data curation as a topic. But there are a ton of gaps/ things that might seem surprising here. E.g., making corpus-level decisions doesn't always tell you much about the underlying training data examples.

Am also concerned about this, but it’s not clear to me that companies even know everything that’s included. I suppose “use it all” is an editorial decision, though.

I just had a paper I reviewed months ago be “desk rejected” by ICLR for this reason. (It’s arguably not a desk rejection after 3 reviewers already chimed in.) But, this seems to be where things are headed.

Even if chucking the papers outright is undesirable (hallucination checkers are not error-free), I'm disappointed there's no process at all other than "oops, you can go fix it if you care to."

(though going forward, I wouldn’t be sad if I had a bit more compute 🙃)

One of my favorite responses to questions about compute in my work this year is “it’s expensive, yes, but I had to develop some efficient algos and write some efficient code to make this possible. This work was done at odd hours on 4 A100s shared by a dozen people.”

note that i said “ML” and “copyright,” which are very specific things that i actually think have very little to do with the anger i’m referring to

it’s hard to work at the intersection of ML and copyright because “both sides” of the debate are angry and, in my experience, most haven’t done much of the background reading in ML or copyright to have an informed opinion. it’s just vibes and anger. i should probably write something up about this.

got to experience the "I did not write that headline" phenomenon firsthand

The article: "Correctly scoping a legal safe harbor for A.I.-generated child sexual abuse material testing is tough."

The headline: "There's One Easy Solution to the A.I. Porn Problem"

After twelve years of work, the world’s most beautiful subway station has been inaugurated in Rome: Colosseo, an underground archaeological museum.⚜️💙⚜️💙⚜️💙⚜️

It's been quite the experience seeing the responses to this work (across the spectrum). I've been working in this area since 2020 & am very grateful to have amazing collaborators + mentors who've supported me along the way (only a few on bsky) @pamelasamuelson.bsky.social @zephoria.bsky.social

our research on memorization and copyright (with @jtlg.bsky.social ) from 2024: scholarship.kentlaw.iit.edu/cklawreview/vol100/iss1/9/

Extracting memorized pieces of (copyrighted) books from open-weight language models Plaintiffs and defendants in copyright lawsuits over generative AI often make sweeping, opposing claims about the extent to which large language models (LLMs) have memorized plaintiffs' protected expr...

our research (with @marklemley.bsky.social ) from May on open-weight LLMs like Llama 3.1 70B: arxiv.org/abs/2505.12546

Extracting books from production language models Many unresolved legal questions over LLMs and copyright center on memorization: whether specific training data have been encoded in the model's weights during training, and whether those memorized dat...

For those interested in the details:

our recent work on production LLMs like Claude 3.7 Sonnet: arxiv.org/abs/2601.02671

AI’s Memorization Crisis Large language models don’t “learn”—they copy. And that could change everything for the tech industry.

The Atlantic posted an article about memorization and generative AI, and it mentions our work on extraction of books from production LLms and open-weight models.

www.theatlantic.com/technology/2...

The referenced work reflects research with @marklemley.bsky.social @jtlg.bsky.social and others.

Extracting memorized pieces of (copyrighted) books from open-weight language models Plaintiffs and defendants in copyright lawsuits over generative AI often make sweeping, opposing claims about the extent to which large language models (LLMs) have memorized plaintiffs' protected expr...

Happy you found our work interesting! Linking to the open-weight model extraction paper @marklemley.bsky.social was referring to:

arxiv.org/abs/2505.12546

(Indexing on the word “often”)

important disclaimer that our research (and the other papers referenced in this article) don’t really capture if they “often just repeat what they have seen elsewhere”

Me too. Like every time I want to move on I get sucked back in.

Eg, there’s an information-theoretic sense where the database analogy is correct, but it’ll be entirely misunderstood if we go that route bc of common perceptions of what a database is. And so that runs more risk than it’s worth imo, since the goal here is wider understanding / conceptual clarity.

Will update you on what we come up with for that law paper that’s due in June 🙃I truly don’t know how to do this yet.

Have also tried to communicate this to law colleagues without math:

scholarship.kentlaw.iit.edu/cklawreview/...

Extracting memorized pieces of (copyrighted) books from open-weight language models Plaintiffs and defendants in copyright lawsuits over generative AI often make sweeping, opposing claims about the extent to which large language models (LLMs) have memorized plaintiffs' protected expr...

In general I’ll just refer to our way too detailed paper on open-weight extraction of books. Tried really hard for completeness here. The main paper is ~20 pages but a lot of context and synthesis and results are in the ~150 page appendix 😭

arxiv.org/abs/2505.12546

Like the struggle to come with good terms and analogy is very real

The terminology is terrible. But it’s also terrible because this stuff might at face value seem simple but deeper understanding is quite complicated. (This is why I still work on this area; it hits at fundamental questions of what machine learning even means)