Three sites made 215,128 “best software” pages for AI. Perplexity cites them

Post by Terr_ » Thu, Sep 03, 2026, 12:53 AM UTC

embedding-shape wrote: post #128

cainxinth wrote: post #86

Earlier quoted context omitted.

Take a random essay and add in a bunch of the phrases that LLMs love like “load-bearing,” “crucial,” structural,” and “woven,” and then submit the original and the edited version to an LLM and ask which is better. It will choose the second one virtually every time. They have ingrained biases that associate those words with good writing and arguments.

This is why using other LLMs as scorers for benchmarks and evaluations is such a bad idea, they'll have preferences you can't anticipate and won't understand immediately.

Post by fluoridation » Thu, Sep 03, 2026, 1:10 AM UTC

thephyber wrote: post #133

mstaoru wrote: post #47

Well it's not only this, or protection from LLMs training on LLM output. LLMs training on human output is also problematic. I was traveling to an obscure small town, doing some "research" with LLMs beforehand. Every and each one told me enthusiastically to go to "Foobar square" (name changed) for the "best street food in XYZ town", some added a lot of colorful details. There was no Foobar square in XYZ town. There wa…

Your one example doesn't make all of LLMs a lie. It's like someone reading a National Enquirer article about "Hillary Clinton being an alien from outer space" (a real headline topic from decades ago) and drawling the conclusion that all journalism is "a lie". The user has to understand media literacy and be at least a little skeptical of the claims that are made, then cross reference with another source.

Post by nemonemo » Thu, Sep 03, 2026, 1:17 AM UTC

Terr_ wrote: post #141

embedding-shape wrote: post #128

Earlier quoted context omitted.

This is why using other LLMs as scorers for benchmarks and evaluations is such a bad idea, they'll have preferences you can't anticipate and won't understand immediately.

The idea that that LLM reliability or bias can be solved with more LLM is... infuriatingly persistent.

Post by CamperBob2 » Thu, Sep 03, 2026, 1:24 AM UTC

fluoridation wrote: post #143

thephyber wrote: post #133

Earlier quoted context omitted.

Your one example doesn't make all of LLMs a lie. It's like someone reading a National Enquirer article about "Hillary Clinton being an alien from outer space" (a real headline topic from decades ago) and drawling the conclusion that all journalism is "a lie". The user has to understand media literacy and be at least a little skeptical of the claims that are made, then cross reference with another source.

Okay, but the anecdote states that every model repeated the pseudo-factoid about Foobar square, not just the 4 GB open source model equivalent of a tabloid.

Post by dspillett » Thu, Sep 03, 2026, 2:00 AM UTC

xpct wrote: post #18

If I recall correctly, there were some papers which suggested that LLMs favor LLM-generated passages over human written ones. I can consistently reproduce this by asking Claude which code snippet it prefers: the one it generated in a different chat, or one that I refactored for my own needs and find more useful. It always picks its own :) I've also experienced that both Claude and Codex routinely include generated we…