By Cory Doctorow (GPG 0xBF3D9110957E5F4C) @doctorow.pluralistic.net · 18/09/2026However, the fact that this method works for *some* things does not mean that it works for *all* things. Naive, "theory-free" statistical extrapolation (the method that LLMs rely on) has hard limits. 214615
By Cory Doctorow (GPG 0xBF3D9110957E5F4C) @doctorow.pluralistic.net · 18/09/2026LLMs are very good at finding statistical regularities and producing new material that matches this statistical picture: you can use an LLM to produce strings of words that are statistically indistinguishable from sentences and strings of pixels that are statistically indistinguishable from images. 211117
By Cory Doctorow (GPG 0xBF3D9110957E5F4C) @doctorow.pluralistic.net · 18/09/2026The single most exciting and interesting thing about LLMs is *how well this works*. Call an LLM "a word-guessing program" and AI boosters will accuse you of reductionism. 21306
nycgeo @nycgeo.bsky.social · 18/09/2026This is the clearest explanation I’ve ever read about what LLM’s are doing, and how it compares to what people do and what actually exists 000
