Chenhao Tan

This account is verified. Learn more

@ChenhaoTan

Professor @UChicagoCS @UChicago. Directing @ChicagoHAI, also part of @UChicagoCI. Email for Postdoc/PhD opportunities. chenhaot.com

Chicago, IL chenhaot.com Joined June 2012

1.5K Following 6.3K Followers 2.7K Posts

Latest posts 5

Pinned post

Chenhao Tan @ChenhaoTan Sep 29

Excited to announce Living Science!

Papers have been the de facto unit of knowledge in science. But science is not static and a paper is not meant to be the final destination, especially in empirical research. Empowered by AI agents, we're revisiting influential research as new data, models, and methods arrive, starting with economics.

🧵

Chenhao Tan Verified account @ChenhaoTan Sep 29

x.com/i/article/2104…
View quoted post on X

Chenhao Tan @ChenhaoTan 15h

This is important and great work!

Michael C. Frank Verified account @mcxfrank 17h

New blog post: I used Claude to audit the reproducibility of every journal paper I've been an author on in the last five years. Our commitment to openness meant that the code and data were available, but there were some significant issues. 🧵babieslearninglanguage.blogspot.com/2026/11/an-ai-…
View quoted post on X

Chenhao Tan @ChenhaoTan 3d

Agreed that alignment/safety is not a problem to solve, but a dynamic to understand and manage.

However, it is also true that it is the most important problem in AI.

Julian Togelius Verified account @togelius 3d

Once again, I sit down to do some useful work and end up writing what I actually think instead, then post it before I sober up and nerf the text

View quoted post on X

Chenhao Tan @ChenhaoTan 5d

I asked AI to read all recent model cards to understand the trend on safety evaluation. One of the most striking takeaways is that while Anthropic, OpenAI and Google publish 100-300 page safety documents, Zhipu (GLM), Moonshot (Kimi) and Alibaba (Qwen) ship open-weight frontier models with essentially no safety evaluation.

model-cards.github.io/safety.html

Chenhao Tan @ChenhaoTan 5d

Honestly, my ICLR reviews are okay, they are just very critical.

Still curious whether @pangram would do a study on AI-generated scores. It would be cool to also reverse engineer the prompt and model used.