we have been experimenting with the idea of autoresearch too, but the optimization is different. we are using cpu instead of gpu, and we're using sota models in

we have been experimenting with the idea of autoresearch too, but the optimization is different. we are using cpu instead of gpu, and we're using sota models instead of local. this shifts the cost balance enough that checking if the changes are worth keeping at each iteration is cheaper than the cost of inferring new improvements to the program. it does not seem fair, or very informative for that matter, to compare optuna to autoresearch, these are optimizing different things, code vs parameters. consider pipelining a change in code with a budgeted run with optuna.