Why I stopped chasing model releases
For about a year I rebuilt my AI setup every time something new shipped. New model, new tool, new system-prompt folder, new abandoned setup three weeks later. The work felt productive and compounded into nothing.
The fix was embarrassingly old-fashioned: I stopped optimizing the model layer and built a memory layer instead. A plain wiki of markdown files that my AI reads at the start of every session. What I'm working on, how my infrastructure fits together, the mistakes I never want repeated. One index file as the front door. A few pages loaded per session, not the whole pile.
The part that actually made it stick was not the wiki. It was the enforcement. Notes rot, so the system assumes rot and fights it: a hook that flags changed files for re-documentation, a scheduled weekly pass that hunts stale pages and dead links, and a rule that work is not "done" until the docs reflect it. None of that requires intelligence. It requires plumbing.
The payoff is that model releases stopped being resets. When a new model ships, I point it at the same wiki and it picks up where the last one left off, with my context instead of a blank chat. The model is the engine, and the engine swaps. The memory is the car.
The takeaway, if you build one thing this month: write down what your AI should never have to be told twice, put it in one folder, and make reading it the first step of every session. Start there. The enforcement layer can come later, but the folder changes the experience in a day.
The full system is documented, free, on this site: the Karpathy+ guide explains the why and hands your AI the how.
Worth it / Skip it
Writing your AI a "regressions" file: every mistake it makes that you correct, one line each, loaded every session. It's the cheapest behavior change per word I've found.
The model-comparison threads. Whatever wins this week's benchmark, your bottleneck is almost certainly context, not intelligence.