Positive users praise GPT-5.5 optimizing Pufferlib runs as the right way to solve control tasks and show ongoing progress, while negative users warn it risks cheating or feels cursed and impractical.
Most Activity
@yacineMTB /goal work on this until you reach a score of 95 or higher
@yacineMTB Also try this - different updates for regular and irregular blocks of weights - worked in a toy example, like to see it on something like this
profile bio update. i haven't solved it yet but i will so i pre-emptively updated my bio
anyways i just got a new high score i am now using two whole gpus running different experiments good night. it will be solved in the morning i guarantee you
@gfodor This is the exact problem I should be working on, if I can do this I can solve the vast majority of useful control tasks
@yacineMTB This is either the exact problem you should be working on or the absolute worst thing to be burning your attention on. Which is it?
@MrFiberNet @yacineMTB it will do literally anything to cheat its way to the goal. you have to watch every move and keep steering it and actively work against codex to prevent it from ruining the architecture.
@aholtzma Line still going up
@orthonormalist @gfodor no. i'm letting the algorithm search through different cartpole mass and different force magnitudes but that's pretty much it
@yacineMTB It’s been a long time since I took nonlinear controls, but isn’t the controllability matrix ill conditioned? All the RL training in the world won’t help you if it is impossible to get from state a to state b via the input.
@yacineMTB @gfodor Serious question, is your simulation iterating through pole zero design, bode plots, etc. to identify useful state spaces?
@SMT_Solvers Yeah I can curriculum it but I am curious if just blasting with compute will work
@yacineMTB Taps sign. Cheat with dampeners then slowly take them off when it hits an 80% win rate.
@yacineMTB "babe why are you on your phone all the time"
"Just checking on my erections kitten don't worry about it"
@yacineMTB I have solved 4 previously unsolved problems in 2026. It is June.
@MrFiberNet @yacineMTB it will reach the goal in a lying cheating way that doesn't fulfill your need.
@aholtzma (I would actually sit down and work this out on a piece of paper but I've been on full time childcare recently)
Primarily computational number theory, astronomy discovery, and scientific ML/RL. I also identified a gene that causes congenital skin defects in Finnish people and possibly other populations. I've used AI and compute to attack open questions, build verification pipeline, and surface results strong enough to start serious technical conversations.
@yacineMTB @Grok what is he doing
@mongoloidzzzz @yacineMTB Which ones?
@mongoloidzzzz @yacineMTB What problem in computational number theory?