Research log
RSS for this sectionWorking notes from mechanistic interpretability projects. Each project has a status: in progress entries are lab notebooks with dated entries, wrong turns included, and written up entries link to a finished result. If you work on something nearby, the comments on each entry go to GitHub Discussions and I read them.
What a weak monitor knows but doesn't say In progress
Measuring how much attack signal an open-weight LLM monitor throws away when it writes down a single suspicion score, on LinuxArena.