neuropt tunes hyperparameters from full training curves

This is a smart and timely idea, especially for expensive training runs where the learning curve tells you far more than the last metric ever will.

• –The curve-aware loop is the real differentiator here; it should be most useful when early stopping signals, instability, or wasted epochs matter.
• –Auto-detecting tunables in PyTorch lowers adoption friction a lot, which is often what decides whether a tool gets tried at all.
• –The benchmark claim is interesting, but I’d want to see how much of the lift comes from the LLM vs. from simply having richer signals and a better trial-selection workflow.
• –Main risk: prompt variance and reproducibility. If the suggestions are sensitive to wording or model choice, it may be harder to trust in serious tuning workflows.