@DrewHawkswood@VictorTaelin And isn't used anywhere because of the issues that require bend2
I'm interested in personally using it and contributing to it. But I do think that requires accepting imperfection and actually releasing the thing
@hopes_revenge I could grace my dog's paws with a light sprinkling of water and she'll run around and stare at me like she just got waterboarded at guantanamo
@MarkSchmidty@Lorenzoowb@VictorTaelin they rolled back that policy and just boot you explicitly to a lower tier model if you run up against their detector
if I'm pangram I'm paying a lot of attention to the youngest generation(s) and looking to see how their AI enabled learning is shifting their language distributions
the game theory of a pause is Not Great™️
defection is cheap and it's hard to verify
even 3-6 mos lead time is a huge payoff
not like nuclear where the marginal spend doesn't matter once you have a few warheads
grateful that the major labs aren't (yet?) nationalized
both of the leading labs have had serious loss of control incidents. there will be serious coping about this from both sides and from /acc bystanders but these are complex emergent loss of control incidents that were detected weeks after the fact
Just wrapped up on @GPU_MODE's Cholesky competition and had a great time
Nominally #4, but 2/3 of the top 3 solutions look like they have reward hacks so we'll see how it shakes out
Watch this space, I'll write up a blog post on my harness process. 1500+ iterations deep!