3 ms·

one stupid idea for "de-dupiclating" LLM-generated summaries, translations and elaborations could be to ask another LLM for a spefific format of detailed summar

one stupid idea for "de-dupiclating" LLM-generated summaries, translations and elaborations could be to ask another LLM for a spefific format of detailed summary and score on the similarity of that (punishing duplication of content that was already present in the index before and from a different domain) Not a production-ready idea of course, and the ideal LLM for that might have yet to be invented. Still, I have plenty of faith in Google still to apply techniques like this in a clever way. They were very serious about their "Knowledge Graph" already 15 years ago, albeit with totally different data processing and acquisition techniques. If all else fails, Google has historically put much of the burden on website authors, e.g. all the schema.org and other structured data stuff. Their game has become much harder, but I still believe that Google can and will remain competitive in web search at least for quite a while to come. Despite all the frustrating spam and a perceivable drop in results quality over the last 5 years, I still don't know about any fully competitive rival for Google.

_heimdall 3y ago
My point is actually that the problem space itself isn't solvable, not that someone other than Google will solve it. Google never solved how to win the game against SEO, and in recent years it seems they've either effectively lost or given up on it entirely. Generative tools make that much worse. Comparing content based on LLM summaries wouldn't really do the trick. For one thing, the LLM itself is a black box so we don't know how it works and could never measure or predict the accuracy or reliability. For another, when two pieces of content have a matching summary the only tie breaker would be publication date (or first seen date in the case of Google). That really doesn't definitively answer the question of which was LLM generated, and won't help at all for content that is originally generated by an LLM and won't match any known content. There's simply no way for an algorithm to accurately recognize computer generated content when the generation tools are good enough to mimic human content.

_heimdall 3y ago

My point is actually that the problem space itself isn't solvable, not that someone other than Google will solve it. Google never solved how to win the game against SEO, and in recent years it seems they've either effectively lost or given up on it entirely. Generative tools make that much worse. Comparing content based on LLM summaries wouldn't really do the trick. For one thing, the LLM itself is a black box so we don't know how it works and could never measure or predict the accuracy or reliability. For another, when two pieces of content have a matching summary the only tie breaker would be publication date (or first seen date in the case of Google). That really doesn't definitively answer the question of which was LLM generated, and won't help at all for content that is originally generated by an LLM and won't match any known content. There's simply no way for an algorithm to accurately recognize computer generated content when the generation tools are good enough to mimic human content.