Amusing myself with the profile pic...
Lisa Gerrard and Pieter Bourke doing "Sacrifice" live remains... out of this world, even with a terrible recording. Needs headphones to work at all. Live, it must have been absolutely beyond belief www.youtube.com/watch?v=BoXs...
"Prefatory remarks on Oppenheimer": michaelnotebook.com/oppenheimer/...
My "Notes on Existential Risk from Artificial Superintelligence": michaelnotebook.com/xrisk/index....
Pretty frustrated that the design seems to be so bad. That seems like the kind of thing where a huge win over Twitter ought to be possible. But Bluesky appears to be playing catchup, not leap ahead
Impressions of Bluesky: + Very slow + Lots of little missing or misdesigned affordances + The network isn't here + The clear wins I can think of and care about over Twitter are that the CEO isn't Musk, and no two-tier system
Pretty interesting. Much of the online "thinking" about AI is really either (massively) motivated reasoning, or might-makes-right. But right now is an interesting moment, too: where good thinking may change people's actions https://twitter.com/AISafetyMemes/status/1661129481216573442
Learning to use ChatGPT well is a bit like "learn to write well" or "learn to do mathematics well" or "learn to use memory systems well": it seems to be a completely open-ended process, which one can get better at in an unbounded way!
It's interesting how much descriptions of alignment techniques for AI seem to be about suppressing their agency This seems to be true to a considerable extent for both RLHF and Constitutional AI