FeepingCreature
No bio...
User ID: 311
... But first, we need to talk about parallel universes.
That is to say, you can also do challenge runs like speedruns, such as the here-given minimum A-press SM64 run. Those tend to be more about research than memorization/repetition.
FeepingCreature
10mo ago
Though the counterside of the models not having access to their weights is so long as they're all trained on the same data, it does not take all that much finetuning to effectively turn any one model into any other model. So our LLM may only have to export a few thousand pages of example output. (Numbers invented.) And while they can't tell what their weights are numerically, they do know what they were trained on (https://arxiv.org/abs/2501.11120).
FeepingCreature
10mo ago
I think Eliezer is still a net positive tbh, I suspect without him we'd gotten on the takeoff train later but superintelligence/AI x-risk wouldn't be on the societal radar to any appreciable amount. Also there really is a case for having the risks as early as possible if you're gonna have them at all.
I think not just that the constitution is a suicide pact, but that every nationstate is a suicide pact practically by definition. That's what it means to hand off the monopoly of force to the state.
This also illustrates that the alternative to law and order is either banditry or civil war.
Yep, and this mechanism should be "I'll call up my bank and take Microsoft out of my personal index." Sadly, while this does exist, it's not really percolated down to a consumer product yet.
More options
Context Copy link