Tag: research
All the articles with the tag "research".
What would convince me a model monitors itself?
Published: at 09:00 PMMy six-part evidence bar for claiming a language model monitors itself, and the probing and patching methods I'd use to clear each part.
How do you know your agent works?
Published: at 02:40 AMScore by outcome, beat doing nothing, beat a random policy, ablate every part, and count the cost. A practical checklist for evaluating AI agents.
I tried to catch a language model making a math mistake. My first accepted paper is about why it did not work.
Published: at 03:30 PMMy MATH-AI workshop paper at NeurIPS 2026 tests whether drift in a language model's hidden state can flag and fix math errors mid-solution. It does neither, and getting to that answer taught me more than a positive result would have.
Attention Residuals: a close read of the Moonshot AI paper
Published: at 09:35 PMA close read of Attention Residuals, which lets each layer choose which earlier layers to read from, and why that matters for interpretability.