The canonical evidence of polysemanticity and superposition that convinced much of the field, including me, that networks pack concepts arbitrarily was misread. What looked like arbitrary packing was mostly a motif I'll call recursive generalization, which is much less concerning. This means we can study networks directly, instead of first untangling them. Neural networks are already clean.
You might also like
Scaling laws work capabilities keep keep jumping but our ability to have alignment and safety assurance just be built in seems not there. um maybe never going to be there if that is the case then what do we do so then you got to come outside the model.
Future Proof
There's no meeting between like Sam Altman and Greg Brockman where they're just like let's read off the weights to GPT7 ... you're going to get like strategy like corporate strategy which is maybe valuable if you're like trading the stock market but it's not like you're going to listen in on a conversation and be like, 'Okay, I can just reverse engineer the next model.'
Diet TBPN
A chatbot account associated with the engineer's spouse, which shared identifiers with his own account, submitted prompts about building and concealing a continuous audio recorder, adding Wi-Fi connectivity, listening remotely, and encrypting its memory card according to the complaint.
Diet TBPN