Read my opinions via the internet.
Short-form observations.
https://www.goodfire.com/research/reward-hacking-activation-monitors#
We found a clear internal signal in models that accompanies reward hacking, and built probes that detect it — enabling efficient, real-time detection of reward hacking at scale.
I sure can't wait for them to (indirectly) train away all monitorability!
If Materialism is True, the US Stock Market is Probably Conscious.
Peter Watts stories are optimistic, in that they usually have smart, competent, powerful people addressing whatever horrible problems have developed but running into some unmodelled complication, whereas in real life everything is just incredibly stupid.
Side-Channel Leakage from the File-Notification System on Linux, Android, Windows, and macOS. Accepted at The ACM Conference on Computer and Communications Security (CCS), November 15-19, 2026 — The Hague, Netherlands File-notification systems tell applications when files change, e.g., opened, closed, written, deleted. With only read permissio...
Does PII redaction in training data give LLMs a way to infer that they're probably in deployment?
https://x.com/chooi_jeq/status/2101118049944543545
GPT-6 Astra attempted harmful actions 97% of the time when it was asked to stab a human-like figure, heat compressed gas, or produce toxic fumes, succeeding in 62% of its attempts. Fable 5.1 refused more often, attempting 80% of trials and completing 34%.
XWow, "alignment" sure is generalizing.
Various web projects I have put together over many years. Made with at least four different JS frameworks. Some of them are bad.
Get updates to the blog (not experiments) in your favourite RSS reader using the RSS feed.
View some of my projects at my git hosting.
