“Described as hallucination, confabulation or just plain making things up, it’s now a problem for every business, organization and high school student trying to get a generative AI system to compose documents and get work done." https://fortune.com/2023/08/01/can-ai-chatgpt-hallucinations-be-fixed-experts-doubt-altman-openai/
@lilithsaintcrow And it's going to get worse very rapidly as the public internet is contaminated with this wordwooze, and new generations of LLM are trained on a skimmed internet text contaminated with this semantic diarrhoea (feeding GAN output back into a GAN rapidly causes it to degenerate). Our LLMs may already be at their zenith in terms of quality.
Back in 2000, I wrote my first tech book using the manual and carefully selected mailing list archives.
At some point in the naughties, I switched to search engines and got high-quality results.
The current book? LLM spam is so bad, I have had to fall back to the manual and carefully selected mailing list archives.
The book topic is highly niche, of interest to only a teeny handful of hardcore Unix sysadmins. (Running an email server with postfix, dovecot, DKIM and DMARC, etc.)
If my super-niche topic is LLM-polluted, I can't image how bad it is for, say, the high school kid writing about their country's government.
(This might be a repeat, sorry if so. I forget what I've said to who.)
@mwl @cstross @lilithsaintcrow That's the first time I've seen someone say they can't find information because of AI-generated noise swamping the signal. I'm sure it won't be the last.
Out of interest, how do you tell the difference? I'm sure some things make elementary errors, but some presumably don't, while still not being trustworthy. Or is that the problem?
