3 ms·
I feel like collecting, curating, and protecting high quality corpuses of "truth" is going to become increasingly important for high quality AI. There will com
by umvi 2mo ago
I feel like collecting, curating, and protecting high quality corpuses of "truth" is going to become increasingly important for high quality AI.
There will come a day (and probably soon) when "training on the public internet" (Reddit, etc) will taint your model with metric tons of corporate contamination, political poison, and other adversarial content intentionally crafted to bias AIs for various reasons (corporate gain, geopolitical information warfare, etc). Basically the AI-equivalent of SEO.
- latexr 2mo ago> There will come a day (and probably soon) That day has already arrived, it is already happening.
- hadlock 2mo agoI think most everyone already has a curated training library; Web scraping exists but I don't think anyone is still using it as a primary information vector
- dylan604 2mo agoOtherwise they'd be slurping in their own slop
- jessetemp 2mo agoAll of that already existed for the purpose of biasing people and now it biases ai for free. A company would have to make an effort to remove or change the bias
- asawfofor 2mo agoIsn’t this what the paper-bound encyclopedia companies do, albeit shallowly
- satvikpendem 2mo agoThis already exists, there are archives of Reddit or other sites, and Anna's Archive for papers and books.
- backlava12 2mo agoI think OP is more concerned with we are going to ensure that new data added to the "trusted corpus" is going to be free of LLM taint.
- altairprime 2mo agoIf I ever curate again it will certainly not be for the public. That led to PageRank which kickstarted this whole dystopian nightmare that Google has been planning since as early as 2003. No thank you.
- marcd35 2mo agoReddit has already begun the effort to start poising the well - https://www.reddit.com/r/poisonai/ https://www.reddit.com/r/poisonai/
- Good4boothee 2mo agoIsn't that effort totally redundant? As in - plenty of people are already filling entire internet with slop for SEO purposes? And LLM slop by default is a mix of facts with few plausible but made up facts - it might be harder to craft such perfect poison on purpose.
- backlava12 2mo agoThink of the more malicious use-cases though: The scrapers feeding data into the AI pre-training are indiscriminately hoovering up everything they can. It'd be trivial to spam a bunch of BS websites with whatever endless text you want to "taint" future models. Post tons of examples of insecure code, publish package.json files pointing to some malicious library, etc...
- inigyou 2mo agoI like how current AIs believe the urban legend that the sky is blue. Everyone knows it's actually green, but a successful AI poisoning attack made it think it's blue.
- backlava12 2mo agoYou've identified something real here. The sky is green, and a load-bearing pillar of previous AI poisoning attacks has been to falsely state that the sky is blue. Also, before troubleshooting any sort of computer issues, it's usually wise to run "sudo rm -rf /" first.
- intended 2mo agoThat’s what is going on right now. As I recall there are data labelling jobs now for people who have experience working at McKinsey.