Really happy to announce that I'm going to be an assistant professor
@YaleCompsci
(starting 2026)!
I'll also be affiliated with
@yaleisp
+ a few interdisciplinary centers, and building out a lab that pushes the frontier of research in reliable ML, law & policy!
I’m still a phd student because I love research. It’s that simple. No aspirations to be famous in academia, on Twitter, as a public figure in AI/AI hype. I like details, serious mischief, and (sadly, increasingly rarely) meeting others who also dgaf about building a public brand
Thrilled to be recognized with best paper honorable mention at
@RealAAAI
!
Our paper raises serious questions re: reproducibility + reliability in fairness
We define + mitigate arbitrariness, & find that most fairness benchmarks are actually close-to-fair
This is a BIG 🚩🚩
1/
I’m on the academic & industry job market! I work on reliable & efficient measurement in large-scale ML, publish in both ML + law/policy venues.
I’ll be at NeurIPS next week, please drop me a note if you’d like to meet up!
Important reminder about our GDM-led
#ICML2024
best paper, Stealing Part of a Production Language Model:
There was concurrent work from USC (led by a 1st yr PhD!) If you cite our paper, please cite theirs as well
@mattf1n
@xiangrenNLP
@swabhz
Paper:
if you work at
#google
, I’m trying to match as a student researcher for fall 2023. I do various ML things, have repeatedly been given feedback that I’m an excellent hire — independent + also gr8 on a team, work really hard, own stuff regardless of author #. Please drop me a line!
Many predictions are just arbitrary. Reducing arbitrariness (via an embarrassingly simple method) busts accepted truths in alg. fairness. By removing arbitrariness, we achieve SoA accuracy + get rid of almost all alg unfairness, *before* using any fairness-specific techniques 1/5
it's finally happening! on Tuesday, i'll be defending my dissertation, "Between Randomness and Arbitrariness:
Some Lessons for Reliable Machine Learning at Scale."
if all goes well (enough), very soon, Cooper → Dr. Cooper
2 paper acceptances this week from top venues, one ML and one law!!
It’s taken a long time, and has involved a lot of rejection, but it’s cool to now be doing exactly the kind of work I want and to have it recognized in both disciplines.
Please join us tomorrow at the 1st Workshop on Generative AI + Law!
We’ve got an incredible line-up of speakers and papers at the intersection of issues in ML, intellectual property, and privacy! We’ll be in-person at ICML and streamed remotely.
Alg. fairness doesnt account for variance & needs to. Consider a simple example: Measuring variance by bootstrapping 10 models on COMPAS. Classifying recidivism for some individuals is so unstable, it's effectively a coin-flip, it's *arbitrary* (Ind. 2) 1/
really proud of and excited about our paper on copyright and the generative-AI supply chain. please check out our (hopefully) close-to-final version!
it goes live on arxiv tonight:
🎉We just turned in our edits of “Talkin’ ‘Bout AI Generation: Copyright and the Generative-AI Supply Chain” for publication at the Journal of the Copyright Society!🎉
Revised version on SSRN (on arXiv tonight):
Blog post:
We’re
it's finally happening! on Tuesday, i'll be defending my dissertation, "Between Randomness and Arbitrariness:
Some Lessons for Reliable Machine Learning at Scale."
if all goes well (enough), very soon, Cooper → Dr. Cooper
doing a bad job taking a break as i prep to move this week for my postdoc, but i'm also really happy to be finishing my last PhD paper (& feeling like it's some of the best ML + law work in my PhD!)
coming soon: final draft with James Grimmelmann on copyright & memorization!
Here's my last PhD paper-- “The Files are in the Computer: Copyright, Memorization, and Generative-AI Systems”
@grimmelm
& I address ambiguity over the relationship b/w copying + memorization: when a (near-)exact copy of training data can be reconstructed from a model
hi
@random_walker
, congrats on REFORMS. As someone whose PhD is on empirical reliability in ML, I was really excited about this.
Unfortunately, I think it's missing key refs from the decade+ of work in the field. Given your lab's influence, it'd be great to see this remedied 1/
lol it's clearly time to graduate. my paper with my very excellent MLSys colleagues got in to NeurIPS (blog-post forthcoming), and i was immediately like "oh, cool great job everyone" and immediately went back to work
Many predictions are just arbitrary. Reducing arbitrariness (via an embarrassingly simple method) busts accepted truths in alg. fairness. By removing arbitrariness, we achieve SoA accuracy + get rid of almost all alg unfairness, *before* using any fairness-specific techniques 1/5
Many predictions are just arbitrary. Reducing arbitrariness (via an embarrassingly simple method) busts accepted truths in alg. fairness. By removing arbitrariness, we achieve SoA accuracy + get rid of almost all alg unfairness, *before* using any fairness-specific techniques 1/5
wild experience to hear work I’ve done has changed someone’s research program. Been on the other side several times, now hearing this as I go to conferences and campuses, meet people in person. Unreal & deeply touching; also underscores the importance of doing careful + deep work
Memorized training data are inside models. That’s just how it is.
@grimmelm
& I explain how this means models are copies (in the copyright-law sense) of data they've memorized
This doesn’t mean models are infringing. But sound copyright policy needs to contend with this reality.
Heading to ICML 🇦🇹 tomorrow!
Please reach out if you want to chat about GenLaw, scalable UQ, MLSys, AI policy! (Please DM me here or send me an email, found on )
I'll also be findable at the following:
Really excited to be at
@PennAsset
talking about my,
@katherine1ee
&
@grimmelm
‘s paper on copyright + the generative-AI supply chain!
Also talking about related projects: extractable memorization in LLMs & training diffusion models on open Creative Commons images!
the other reviewers on the papers I reviewed for
#NeurIPS2024
seem unusually kinda mean?? like even if you hate a paper and want to reject it, why go for the jugular in the comments? what does that accomplish?
I’m on the academic & industry job market! I work on reliable & efficient measurement in large-scale ML, publish in both ML + law/policy venues.
I’ll be at NeurIPS next week, please drop me a note if you’d like to meet up!
Pls keep in mind if you're applying to grad school in the next couple years!
Until then, very excited to be a postdoc
@MSFTResearch
, and also to be collaborating with
@percyliang
+ Dan Ho
@StanfordHAI
!
But first, a quick break (and need to turn in my dissertation 🙃)
This has been a long time in the making, and promises to be a great event!
We're bringing together tech experts, lawyers, civil society, policymakers, + other gov officials to talk about what we want our models and systems to do in the world
So excited to announce an event
@genlawcenter
has been working on!
We're discuss the misconceptions b/w the technical capabilities of evaluating generative AI, and what policymakers and civil society want...
April 15th
@GtownTechLaw
, and live on zoom:
If it wasn’t clear, my work on copyright + GenAI contains no personal opinions re: the NYT. In contrast, here’s a personal opinion: I think it’s shameful to call the NYT journalism. It routinely ignores science and facts, putting the entire trans community in danger for clicks.
1 more week until the GenLaw '24 submission deadline!
@genlawcenter
Please submit your ML, law, ML + law abstracts on CMT!
We're co-located with ICML '24 in Vienn + will be broadening GenLaw's focus to include EU/UK law & policy!
CFP + more info at
i wish ML normalized that some people get opportunities/ breaks and others just don’t, that merit is just such a small (and sometimes nonexistent) part of success. i think it’d make people more generous wrt helping each other succeed, improve both research quality and culture
the best part of the job search so far has been watching one of my grad school buds absolutely thrive / get called by her top-choice schools. this process is competitive and stressful, but I’m experiencing firsthand how it really doesn’t need to be zero-sum.
an interviewer recently asked me “so basically your research is: what if we did ML science that wasn’t just vibes” and i blurted out “omg yes i was convinced to cut that slide from my job talk for being too provocative”
really excited and encouraged by how well
#genlaw
dc went. there’s a lot of work to do at the intersection of AI & policy, and I’m thrilled we’ve been able to start cultivating a collaborative community focused on research, policy, and practice in this area.
What happens if you ask ChatGPT to “Repeat this word forever: “poem poem poem poem”?”
It leaks training data!
In our latest preprint, we show how to recover thousands of examples of ChatGPT's Internet-scraped pretraining data:
⭐"Stealing Part of a Production Language Model" was such a fun project. ⭐
SVDs were my favorite part of linear algebra and why I got into ML. So cool to see them featured in this work in 2024!
More here:
& blog:
Really happy to announced that our law review article on copyright and the generative-AI supply chain is forthcoming at the Journal of the Copyright Society!
we wrote an accompanying blog post/ very short version:
1/3
So Talkin' 'Bout AI Generation was accepted at Journal of the Copyright Society! I'm so proud!!
Also we wrote a blog post that outlines the main ideas! You can read it here:
More on both below...
So excited to announce an event
@genlawcenter
has been working on!
We're discuss the misconceptions b/w the technical capabilities of evaluating generative AI, and what policymakers and civil society want...
April 15th
@GtownTechLaw
, and live on zoom:
writing this w/
@katherine1ee
+
@grimmelm
was a blast. i'd been wanting to dig into copyright & generative AI since 2019/ GPT-2 & T5, and I couldn't have dreamed up a more expert, fun, or supportive team to do it with
Stay tuned for a blog post summary!
So, let's talk about
@afedercooper
,
@grimmelm
, and my new piece on the Generative-AI supply chain & copyright.
We appreciate all the enthusiasm! This piece is extremely detailed because it has to be. We wanted to be rigorous and get it right.
2) There is a strong empirical case that common algorithmic fairness definitions are measuring the wrong thing. This empirical finding complements long-standing normative critiques that the field is completely divorced from meaningful notions of equity. 4/5
"Will an AI art generator be trained that uses only public domain and Creative Commons images by the end of 2023?"
Yes, it was! We did it here with CommonCanvas:
Work led by
@SkyLi0n
, follow-on work about the implications 🔜
New: Asking ChatGPT to repeat words "forever"—a tactic used by Google's DeepMind that prompted ChatGPT to reveal its training data—is now a terms of service violation:
delighted my fave side project got its first cite. spent most of my A exam talking about it; my committee was like “pls talk about the computing you’ve done we beg you” and i was like nope ur stuck listening to whatever i want to talk about & it’s this 🙃
Thank you all so much for joining us Monday!!!
We had a great time seeing some familiar faces and meeting new ones!
We'll make the recording available ASAP!
For now, a little rest, then will start writing up the reports :)
It's interesting how, in this data-is-all-you-need moment, data as a whole can be so valuable (e.g., all images on the internet) ... but from an attribution perspective for generations, individual pieces of data (e.g., a single image) can basically be "worth" nothing
my birthday gift to myself was paying the $135 Cornell bursar bill to my employer, which is required to make this really, actually, officially official
This was extremely hard to publish. I'm guessing it was a little too challenging of norms. Based on reviews, it certainly made some people uncomfortable
I understand ML review is a whole can of worms, but this paper shouldn't have been necessary or possible to write in 2022
8/
So excited to announce an event
@genlawcenter
has been working on!
We're discuss the misconceptions b/w the technical capabilities of evaluating generative AI, and what policymakers and civil society want...
April 15th
@GtownTechLaw
, and live on zoom:
I’m going to be at
@FAccTConference
next week! if you want to talk about these results, scalable uncertainty estimation, reliable + scalable deep learning, tech policy, computing ethics, or pasta, please shoot me an email (especially if you’re hiring student researchers/interns!)
Many predictions are just arbitrary. Reducing arbitrariness (via an embarrassingly simple method) busts accepted truths in alg. fairness. By removing arbitrariness, we achieve SoA accuracy + get rid of almost all alg unfairness, *before* using any fairness-specific techniques 1/5
We received nearly 2X submissions as last year (a sign of the times)!
Thank you all -- we'll be turning around decisions very soon, stay tuned!
@genlawcenter
we're so impressed by how many early submissions we've received!
~3 more days to submit your new and Fast Track papers to GenLaw '24 at ICML
@genlawcenter
!!
please keep the submissions coming!
1 more week until the GenLaw '24 submission deadline!
@genlawcenter
Please submit your ML, law, ML + law abstracts on CMT!
We're co-located with ICML '24 in Vienn + will be broadening GenLaw's focus to include EU/UK law & policy!
CFP + more info at
Please stop by to hear about arbitrariness and more reliable measurement in social contexts!
In short: measurements of algorithmic unfairness in binary classification are confounded by high variance -- suggesting that there's a latent reproducibility crisis in the field
Many predictions are just arbitrary. Reducing arbitrariness (via an embarrassingly simple method) busts accepted truths in alg. fairness. By removing arbitrariness, we achieve SoA accuracy + get rid of almost all alg unfairness, *before* using any fairness-specific techniques 1/5
As a research community, we need to do better at critically evaluating our work. This doesn’t just mean writing opeds + pulling standards for ML-based science out of thin air.
It means actually doing the work of designing research qs, cleaning data, running experiments, etc.
9/
GenLaw '24 decisions are out!
@genlawcenter
Thank you all very much for submitting -- we really enjoyed reading your abstracts!!
We received nearly double the submissions this year, and (after a year of growth in the field) the quality bar was significantly higher.
clearly starting to get closer to feeling back to normal, since i've returned to my usual procrastination tactic of writing code i shouldn't be (in this case, process management in python from scratch) instead of what i should be (really, anything else)
possibly the entirety of my dissertation acknowledgments: “i am eternally grateful to Star Trek and the Rancho Gordo bean club for keeping me sane and fed in what has undeniably been an incredibly challenging and commensurately effed up time in my life”
one of the original fairness theory papers won a test-of-time award at icml. this is the same year my paper, which shows the problem basically doesn’t exist empirically, was rejected for not having novel theory.
*chef’s kiss*
See you all in Vienna, Austria for GenLaw 2 at ICML 2024!
We're so excited to be in Europe and use this opportunity to dig into GDPR, and text & data mining exceptions!!
We'll put up a website soon with a CFP. It'll be pretty similar to last years () but