Mastodon Feed
Boosted by glyph ("Glyph"):
jonny@neuromatch.social ("jonny (nonvenomous)") wrote:
Obviously model distillation exists, and every open model nowadays is released as a generation of different sized models. I would even bet that there is all the above demand-based, topic-based, and user-credulity-based serverside switching between model sizes. But the conspiracy theory functions even if there isn't. The ambiguity created by the AI vendors on this topic as well as the abstract token allowances in the subscriptions operates exactly like the kind of parallel channels of periodic reinforcement from other scambling art forms like mobile games: the gambler is aware the game is rigged in some way but incapable of knowing. The rigging demands explanation, and the house is not quick to silence any stories that accidentally provide explanatory power even if they're bad stories. There are multiple ways to be cheated, so while one day it's a nerfed model, the next its token allowances, the next its guardrails, and so on.