Nick on Nostr: I often see people using "AI" (often non-specific as to what tool exactly) to ...

I often see people using "AI" (often non-specific as to what tool exactly) to summarize a longer text source (or spoken sources with speech to text). My immediate reaction was that you could give a generative model (with sufficient memory) a text and a prompt and ask it to produce a summary, and it would likely produce something that looks like a summary but fundamentally isn't; rather it would be some text it made up that looks statistically similar to a summary and contains text statistically correlated to the input text. That would often look similar to an actual summary, but would differ from what we'd normally think of as a summary because it could contain things that were nowhere in the input text (just correlated to language in the original text) and might actually contradict the input text if it happened to be unusual compared to what was in the training set.

On the other hand, it did seem to me that it would be possible to train an LLM model specifically around the task of summarizing text. It would have the usual constraint that it would only ever be probably approximately correct (PAC), but it could be less or highly unlikely to produce fabricated output or output contradictory to the input. Based on a little reading, it looks like such specialized models exist. My question is this: When someone uses one of the major commercial chatbots (ChatGPT, Claude, etc.), and they ask it to summarize some text, is it employing such a more specialized model or is it just employing a more general text generation process as I described in the first paragraph?

#ai #llm