Thread

i use claude, mainly because of the personalities of the models, i don't like many of the others and only cursor Composer is close to being as good at working with code, but about 3x more expensive. claude pro 5x at $90/month is more than i can use and i'm a maniac on it. but i'm going to learn how to use pytorch with ROCm, experiment with some training programs for some interesting training methods i've found, need to write some tools to "augment" its training data (for, eg, a sudoku solver, this means rotating and mirroring the solved grids exhaustively before feeding it to the model) - there is a repo with all the bits you need (if you have an nvidia AI capable card, for AMD it needs modification). this is only for programming stuff though, i just want a Rob Pike grade of go programming assistant. the other stuff, already is fine for documents (my gpu has 16gb memory and can run ~20b models that are enough for documents and simple reasoning/tool use. planning to get one of those framework desktops, 128gb memory, can run 70b models (and if you get two, you can team them up with a 10gbit ethernet or 20gbit usb-c based connection) to run models in the same sort of size range as claude, gemini, etc). so, yeah, understandable to not want to give data to assholes like sam altman, mark zuckerberg, elon musk, or the rest, let alone paying for the privilege, but 2026 the home-trained home-run AI thing is going to start. also if you have the latest mac mini, this machine can run big models too, though it's not as fast as either ryzen ax max 395 or the dgx spark from nvidia