So in many respects, search the place it used to construct the model? Isn't that functionally bias-reinforcing?
"Look what I synthesise is correct and true because when I use the same top 10 priming responses which informed my decision I find these INDEPENDENT RESULTS which confirm what I modelled" type reasoning.
None of us have a problem with an LLM which returns 2+2 = 4 and shows you 10 sites which confirm. What worries me is when the LLM returns 2+2 = 5 and shows 10 sites which confirm. The set of negative worth content sites is semi infinite and the set of useful confirmed fact (expensive) sites is small so this feels like an outcome which is highly predictable (please don't beat me up for my arithmetic)
e.g. "Yes Climate science is bunk" <returns 10 top sites from paid shills in the oil sector which have been SEO'd up the top>"
We will very quickly enter a Kepler effect of information on the internet. All text on the internet will become AI slop being parsed by AI. Real information and human beings will be drowned out by the garbage. The internet will cease to be useful and we will retreat to corners of the web or to walled gardens. I'm seeing more and more online communities these days enforce invite only because there's just too much AI slop everywhere now.
"Look what I synthesise is correct and true because when I use the same top 10 priming responses which informed my decision I find these INDEPENDENT RESULTS which confirm what I modelled" type reasoning.
None of us have a problem with an LLM which returns 2+2 = 4 and shows you 10 sites which confirm. What worries me is when the LLM returns 2+2 = 5 and shows 10 sites which confirm. The set of negative worth content sites is semi infinite and the set of useful confirmed fact (expensive) sites is small so this feels like an outcome which is highly predictable (please don't beat me up for my arithmetic)
e.g. "Yes Climate science is bunk" <returns 10 top sites from paid shills in the oil sector which have been SEO'd up the top>"