Great to have been involved in this initiative led by Sayash Kapoor and Arvind Narayanan to (hopefully!) improve the use of machine learning in science. Further thoughts in my Substack post: fetchdecodeexecute.substack.com/p/reforms-a-gu…

To my mind, unconstrained military use of AI is one of the most risky & is underemphasized in policymaking. Military use must be a central part of AI Safety discussions. Glad to see a couple of new pieces emphasizing this point.

ft.com/content/da03f8…

foreignaffairs.com/united-states/…

With my brilliant friends at Knight First Amendment Institute, I filed suit against Meta today, asking a federal court to find that CDA section 230 gives users rights to control what they see on social media via third party tools. See our complaint at knightcolumbia.org/cases/zuckerma…

Lots of practical advice to help researchers doing ML-based science avoid unintentional irreproducibility and overgeneralization in this new paper led by Sayash Kapoor

REFORMS is an exceptional work by an ensemble cast spanning institutions and disciplines! Check it out!

The approach also directly inspired our work on open foundation models, where we worked towards consensus across folks from different institutions:
crfm.stanford.edu/open-fms/

Excited to share that our paper introducing the REFORMS checklist is now out Science Advances!

In it, we:
- review common errors in ML for science
- create a checklist of 32 items applicable across disciplines
- provide in-depth guidelines for each item

science.org/doi/10.1126/sc…

MTurk is basically junk responses. People often lie about their background characteristics. And they often choose the same answer for most questions, regardless of content, such that you can ask people opposing questions and get completely incoherent results (even after screening…

- Agents are costly and that should be jointly optimized with task accuracy
- Simple baselines like retrying, retrying with different temps, retrying with better models outperform complex Agents on the Pareto frontier of cost/accuracy
- reproducibility & benchmarks continue to be…

put your mouth where your money is

amazing how adding cost on an axis can directionally shift interpretations on many ai agent approaches!

Rigorously evaluating “agents” takes thought!

great work debunking the (cost-normalized) performance of popular coding agents by Arvind Narayanan Sayash Kapoor Benedikt Ströbl .

We fully recognize that there are downsides to reporting dollar costs in evals, given that costs change quickly. The difference in our perspective comes down to model evaluation vs downstream evaluation (our focus is the latter). aisnakeoil.com/p/ai-leaderboa…