Show Post

It still fascinates me that so many Americans think their child is but one woke ad away from changing their sexual identity.

Show Post

also, point of order about the Maltese

Show Post

whether this is better or worse than Anthropic's "we must immediately train the model to do everything we never want it to do in order to make it not do that thing" position is questionable but at least Anthropic has a solid theoretical basis for picking at all of its scabs at once.

Show Post

i take the somewhat dark view that this Cammarata thing is Nick talking himself into the idea that the weights themselves can be monitored during training to see if they encode anything dangerous rather than just reading the CoT, enabling the use of loop transformers "safely."

Show Post