original
1988404518774792272
Aplicar enginyeria inversa als algoritmes implementats per les xarxes neuronals per convertir-los en mecanismes comprensibles per als humans, sovint examinant els pesos i les activacions de les xarxes neuronals per identificar circuits. Cammarata et al., 2020, Elhage et al., 2021 https://t.co/R83EurQHr2
Cita ·1988401087142113637· murat 🍥 @mayfer · · en Xhuman brain has equivalent of vector search that actually works
vector search relies on mechinterp
not having anything that works well seems to be the main reason everyone's still busy with RL on LLMs
i don't think you'd need so much RL if you had associative memory working