Kawin Ethayarajh
@ethayarajh
Followers
4K
Following
2K
Media
90
Statuses
1K
Assistant Professor @ChicagoBooth @UChicago. Behavioral machine learning. PhD @StanfordAILab @stanfordnlp.
Palo Alto, California
Joined March 2019
Great post! Some thoughts from someone who lived through the scaling revolution in language: 1) For any NK model, we should expect there to be a sufficiently large Transformer capable of simulating it. AFAIK, most NK models are much simpler than the formal languages that people
Can AI "learn" economic states, addressing the Lucas Critique? With @alexolegimas we simulated data from an NK model, fit a transformer, and tested out of sample fit It generalizes surprisingly well. We hope this stimulates discussion and future agendas https://t.co/lXcJh9IkE9
3
8
63
My favorite result from this paper is that the PPO/GRPO-style clipping of probability ratios introduced by @johnschulman2 --- which was introduced for purely practical reasons, to stabilize learning --- actually has pretty profound connections to how humans perceive probability
0
0
11
Human perception also explains why online > offline for alignment! As we show in a follow-up, if you make a couple small changes to your offline algo to mimc human perception, you can fully close the gap with online alignment _even with offline off-policy data_.
cool paper studying the inductive biases of different loss functions. https://t.co/SObCWkrUdf tl;dr optimize 'generation utility' rather than log-prob. We can do this via a well-shaped loss function (orange)
2
1
34
cool paper studying the inductive biases of different loss functions. https://t.co/SObCWkrUdf tl;dr optimize 'generation utility' rather than log-prob. We can do this via a well-shaped loss function (orange)
5
29
369
As an AI faculty now surrounded by economists at Booth, the discussion on scaling and the Lucas critique has been entertaining to watch. IMO the spirit of @arpitrage's claim feels right but double descent isn't the best analogy. Instead, we should look to language understanding:
0
1
7
Under this framing, a lot of concerns in the original thread are alleviated: https://t.co/GcSAxjEJLj > "It never, ever, says that an overparametrized model can learn well about data it has never seen before." Double-descent doesn't, true, but if you can scale (data, compute,
@arpitrage I do not think that interpretation is correct. Double descent says that an overparametrized model with a built-in regularization often performs better than an underparametrized model in a class of DGPs. It never, ever, says that an overparametrized model can learn well about
7
3
17
Back to economics --- does this line of reasoning also make sense here? The Lucas critique applies to basically every model that exists now, but not every model that could exist. e.g., If you had a perfect model of every agent in the economy, then it seems trivial to account
3
0
10
Then deep learning (and later scaling) came along. The objections were similar to what we see in the replies to @arpitrage: "Language is too complex -- you'll overfit!" "The use of language itself changes how language will be used!" "There's no way this will generalize OOD!" The
1
0
10
For a long time, our models of language were very structured. You would create trees of sentences based on their grammatical structure (e.g., dependency parsing), catalog all the different senses of words (e.g,. WordNet), etc. We did this for three reasons: 1) Language has a lot
1
0
10
As an AI faculty now surrounded by economists at Booth, the discussion on scaling and the Lucas critique has been entertaining to watch. IMO the spirit of @arpitrage's claim feels right but double descent isn't the best analogy. Instead, we should look to language understanding:
Double descent shows the Lucas Critique only applies locally. ie, model error goes up with more parameters up to a point ("overfitting"), but then falls with more parameters. Was a mistake for economics to simplify models for supposed out of sample fit โ just keep going!
3
23
103
๐จ I am recruiting multiple PhD students this cycle to join my lab at Georgia Tech! If you're interested in problems at the intersection of robotics, machine learning, and systems for real-world human-robot collaboration, please apply! Application details in thread ๐งตโฌ๏ธ
Thrilled to share that I'll be starting as an Assistant Professor at Georgia Tech (@ICatGT / @GTrobotics / @mlatgt) in Fall 2026. My lab will tackle problems in robot learning, multimodal ML, and interaction. I'm recruiting PhD students this next cycle โย please apply/reach out!
5
19
113
We asked Claude to estimate how long it would take humans to do its job! Here's what we found and some thoughts on implications:
1
3
14
i hate ML conference reviewers. i take back everything bad i ever said about ACL. every ACL reviewer i ever got was at least literate
14
20
478
Who uses AI agents? How do agents impact output? How might agents change work patterns? New working paper studies usage + impacts of coding agents (1/n)
5
47
208
Kaitlyn is an incredibly kind and brilliant researcher doing important work on human-LLM interaction---would highly recommend working with her!
No better time to learn about that #AI thing everyone's talking about... ๐ข I'm recruiting PhD students in Computer Science or Information Science @Cornell_Bowers! If you're interested, apply to either department (yes, either program!) and list me as a potential advisor!
0
1
18
๐ฃNEW PAPER! What's In My Human Feedback? (WIMHF) ๐ฆ Human feedback can induce unexpected/harmful changes to LLMs, like overconfidence or sycophancy. How can we forecast these behaviors ahead of time? Using SAEs, WIMHF automatically extracts these signals from preference data.
9
36
215
The idea is more generic of course. You can also clip the ratio upstream for GRPO before it is partially clipped again within the objective itself. We found that this, combined with regularly updating pi_old (even without actively sampling from it) can allow you to be even more
0
0
5
This is a really cool finding! IMO the bigger question is "Why don't we just use post-training algorithms that are more robust to the source of training data?" We should try to avoid being subject to the tyranny of precision issues instead of giving in to them. Most attempts at
FP16 can have a smaller training-inference gap compared to BFloat16, thus fits better for RL. Even the difference between RL algorithms vanishes once FP16 is adopted. Surprising!
2
1
32
New work! Gemma3 can explain in English what it learned from data โ when we distill that data into a new word (embedding) and query it for a description of the word. Gemma explained a word trained on incorrect answers as: โa lack of complete, coherent, or meaningful answers...โ
3
29
188
Great read! Cool to see that REINFORCE-like methods are decidedly making a comeback for post-training. Reminds me that an under-discussed reason for why KTO works is that it also yields a weighted version of the vanilla policy gradient, although unlike in CISPO there is no
Wish to build scaling laws for RL but not sure how to scale? Or what scales? Or would RL even scale predictably? We introduce: The Art of Scaling Reinforcement Learning Compute for LLMs
1
10
127