Hacker News Reader

Comment by wang_li | Hacker News Reader

LLMs don’t learn facts. They don’t think, cogitate, daydream, hallucinate, lie, or any other activity that requires intelligence, a mind, or a personality. Writing about them the way you did just advances the scam that is the current AI industry.

View on HN

When I say "my thermostat tries to keep the temperature at 72", this doesn't mean I'm asserting anything about the thermostat having "goals" in some deep philosophical sense. It just means that the behavior of the thermostat can be usefully understood, in the relevant context, as goal-directed. (Likewise, "legs are for walking" even though legs were designed through a evolutionary process, etc.) We use this sort of language all the time, it's very useful, and it's mere pedantry to demand people not to. Likewise here, a machine can absolutely "learn" facts in the sense that it extracts them from the copyright-able creative choices in the text written by a human author.

You can't say I'm using the word learn as a metaphor for some purely mechanistic and deterministic process and not like when people say learn as in middle school kids learning algebra and then switch to using the second definition to absolve OpenAI from following the license.

That'd be like saying my computer learned the film Oppenheimer when bittorrent copied it to my computer's storage system, transcoding it to a lower bitrate, and claim I'm free of copyright law because my computer just learned some facts.

All of OpenAI's products are databases with highly lossy compression applied. Using language that implies a mind is pure scam.

I’m not using anything about “learn” to justify what’s copyrightable and what’s covered by fair use. Those things are defined by the output, not the process that produced them. The point is that LLMs are very effective at extracting non-copyrightable facts from the words and other creative choices made by an author.

Reply on news.ycombinator.com