We can now understand most neurons in models up to trillions of parameters.

We can now understand most neurons in models up to trillions of parameters. AI researcher Nick Cammarata’s essay “Neural networks are already clean” reverses his earlier position on model observability. He formerly thought models used neurons in arbitrary ways that looked like tangled knots, blocking direct observation. Now he now believes they’re applying “recursive generalization”, making “the outlook for mechanistic interpretability … far better than we thought.” ~ learn more