Antoine Vigouroux

growing the tech tree
blog: Telescopic Turnip

[Antoine Vigouroux @teleturnip]

Paper: https://t.co/mZAYaIV9r8
Code: https://t.co/9JfZSGvduT
Thanks to @leedsharkey, @danbraunai, @corythesaurus, @tspiliot and @MATSprogram

[Antoine Vigouroux @teleturnip]

New paper: instead of decomposing a whole neural network, we only dissect the parts used on a task we care about. It's much cheaper, and gives mechanisms you can inspect, erase, or edit to change how the model behaves 🧵

[Antoine Vigouroux @teleturnip]

We expect tPD to make parameter-space interpretability more accessible, and give researchers insight into how larger models process narrow tasks and how we might surgically edit them. Come talk to us if you have an exciting application in mind!

[Antoine Vigouroux @teleturnip]

If your opinion can be simplified to the point it fits in 280 characters, it is almost surely wrong.