Claude generates slow single-worker dataloaders to avoid execution failures.
agents often pick extremely conservative configurations (max tokens, parallelism, learning rates) to the point of ineffectiveness.
i think this is another example of rlhf-induced pathological failure avoidance, much like defensively spamming try/except everywhere in python
No Digg Deeper questions have been answered for this story yet.
@xlr8harder yes claude write a shitty handrolled dataloader with a single worker just fuck my shit up fam i dont even care anymore
agents often pick extremely conservative configurations (max tokens, parallelism, learning rates) to the point of ineffectiveness.
i think this is another example of rlhf-induced pathological failure avoidance, much like defensively spamming try/except everywhere in python