Post

Rewildpunk: The Counterculture AI Actually Needs by @abiawomosu.bsky.social: Plurality over monoculture. Relationship over command. Local usefulness over universal scale. Human judgement over automatic optimisation. And infrastructure people can shape & steward, over platforms they merely consume

abiawomosu.substack.comRewildpunk: The Counterculture AI Actually NeedsPunk took music back from the labels. Here’s how we take intelligence back from Silicon Valley.

fry69

@fry69.dev

Well, how do you plan to "punkify" a training run? Apart from the running cost (several million $ per model/attempt), it also requires access to massive amounts of data, knowledge, hardware, etc. See this dashboard for a currently running training run -> mimo.xiaomi.com/rl/

mimo.xiaomi.commimo-v2.6 RLTraining metrics of the mimo-v2.6-pro and mimo-v2.6-flash reinforcement-learning runs, live from the trainer's logs.

18/9/26, 6:56 am0 reposts1 like

fry69

@fry69.dev

18 Sept

Also this is an ongoing process. Models currently have a shelf live of one or two years max (often much less). Even if there were no progress in training procedures, the training data continuously gets updated and refined, so models need to be trained from scratch after a while.