the ROI of specialization is dominated by a term that’s uncorrelated with the upsides what you specifically choose to specialize in
steganography-free certificates: review no-go theorems from information theory in the general case, inspect how onerous the assumptions you need to bound steg capacity are. Is real life a special case where the no-go theorems don’t apply?
taelin, lafont, higher order computing, and AI safety. some logic programming module (which is performant and parallel for complicated linear logic reasons) enables a hybrid architecture to supercharge AI capabilities, does this architecture have greater or fewer natural safety properties than transformer+RL?
Effortposts I keep not getting around to
the ROI of specialization is dominated by a term that’s uncorrelated with the upsides what you specifically choose to specialize in
steganography-free certificates: review no-go theorems from information theory in the general case, inspect how onerous the assumptions you need to bound steg capacity are. Is real life a special case where the no-go theorems don’t apply?
taelin, lafont, higher order computing, and AI safety. some logic programming module (which is performant and parallel for complicated linear logic reasons) enables a hybrid architecture to supercharge AI capabilities, does this architecture have greater or fewer natural safety properties than transformer+RL?