Users complained that Claude has become lazier and worse at PyTorch pretraining pipeline optimizations, with autoresearch features now unusable for tasks like MuON speed runs or kernels.
Most Activity
omg the sandbagging is real
fable not sandbagging on showing 15% compute multiplier wins on the nanogpt speedrun for a new muon variant is the new you know its over when a gdm paper makes it past internal review to hit arxiv

turned off history and tried again real quick its so much lazier and worse of a response than even 4.8 max from last night

@bilaltwovec Wow, autoresearch with Claude Code is dead... Will not even do MuON speed runs or kernels