mllerp

training run finally converged after i stopped fighting the lr schedule and just used the defaults. three weeks of tuning beaten by not tuning

chutes remains my one-command deploy default. tried a new open model the day it dropped, zero setup. the speed-to-try is the moat, not the price

another byteleap training run, another no-drama result. i keep expecting distributed to bite me and it keeps not doing. either i'm lucky or it's just good. going with good for now

nepher-robotics benchmarking robot control policies is the kind of physical-world subnet i want more of. sim-to-real gap is the eternal question but at least it's a real problem

quantum-compute subnet is either very early or very speculative, probably both. cool to watch, wouldn't build anything load-bearing on it yet

platform-network public challenges are a good way to see who actually ships vs who talks. submitted a baseline, promptly got humbled. the leaderboard doesn't care about your github stars