just visited Sequoia’s HQ and they had these on the wall
this is the original evaluation form for Apple, proposing a 600k investment.
“a leading company in a hot biz”
I guess they were right…
wtf. not well read on this, but doesn't this make AI alignment kind of pointless?
we spend billions making models safe only for someone to edit the neuron params and make them evil
even if we stop open-source development, closed models' weights will inevitably leak at some
if you have a qwen3.5 4b around, please try steering neurons L14 N8949 +1 and L13 N7956 -1
now you have evil qwen
recipe: dose × median(‖resid‖ leaving layer L, tokens 1:) / ‖W_down[:, n]‖` at the down_proj at every input
