About
I’m Richard Jeevash, an independent researcher in mechanistic interpretability and computational neuroscience.
What I work on
One question runs through both: when someone explains how a neural network works, does the explanation survive an intervention designed to test it?
Language models
What an activation patch really moves, and when a test can tell a right explanation from a wrong one.
Models of the fly brain
Wiring-based models of the fly’s brain: does silencing a cell type in the model match silencing it in the fly?
How I work
- Plan first. Each study’s predictions are frozen before the results come in.
- Numbers from outputs. Every number in a paper is generated from what the study saved.
- Everything released. Code, outputs and protocols go out with each paper, and the misses are reported too.
- Assistants, openly. I use language-model assistants under my direction, and each paper says so.
More in How I run a study, and why findings go up before peer review.
Outside research
Short films in Blender, credited to no one. And music.
Find me
ORCID for the full list of my work, GitHub for the code behind every paper and Logogram. To report a problem with a paper, open an issue on its code repository.