Graduated in June, technically still finishing up right now (senioritis is real, sorry to every teacher currently getting my worst effort). Been accepted to State for the fall — computer science, in-state tuition, close enough to keep living at home and keep the setup I've already got instead of starting over in a dorm with a laptop. Not glamorous. Fine with that. I don't think the thing I actually want to learn is taught in the classes anyway, and it's a lot cheaper to learn it myself while someone else pays for my degree to exist. Anyway, actual notes. I keep manually doing the same loop over and over with every experiment: look at what worked and didn't last time, think up a change, implement it, run it, check the result, write it down, repeat. Every single time, by hand. And most of those steps don't actually require me specifically — "look at these three numbers and pick a plausible next thing to try" is not a fundamentally human operation, it's just a decision I happen to be making. Started sketching what it would look like to automate pieces of that. Not the whole thing, not yet, that feels like overreach. Just: a script that reads my last N experiment logs, proposes one or two small parameter or architecture changes based on some simple heuristics, and queues them to run overnight so I wake up to results instead of spending the evening babysitting one run at a time. First version is dumb. Really dumb. It basically just perturbs hyperparameters within a range and occasionally tries swapping in a variant it's seen do well before. No real judgment. But even dumb, it means I can go to bed and wake up to six results instead of one. Calling the whole thing PROMETHEUS in my notes now, mostly as a joke at first — my code folder needed a name and "ml_stuff_v3" was getting embarrassing. Might stick. Might not.