4 ms·

Maybe it's a good thing for Wikipedia? DDG can scrape Wikipedia once, and then let their AI act as a sort of "knowledge CDN" so Wikipedia ultimately serves fewe

Maybe it's a good thing for Wikipedia? DDG can scrape Wikipedia once, and then let their AI act as a sort of "knowledge CDN" so Wikipedia ultimately serves fewer page loads. As long as DDG supports Wikipedia in some way that seems like a net positive. Seems to me that Bing and Google's chatbots are having truthiness issues because they're just scraping information from all over the internet. Any HN reader knows all too well that you can't believe everything you read on the internet. While Wikipedia isn't necessarily all true, the bar is much, much, much higher than the truthiness of the rest of the internet. Can anyone else verify if a LLM trained on a set of mostly-true info like Wikipedia will still make shit up?

rpastuszak 4y ago
> Can anyone else verify if a LLM trained on a set of mostly-true info like Wikipedia will still make shit up? It will, but less frequently. For instance, I can imagine a situation where it would give a plausibly sounding answer to a question based on an incorrect premise. The issue is that the bullshit responses will sound plausible from the stylistic pov. Under the hood LLMs we're talking about here are still just "spicy autocomplete" with non-deterministic results. The results are impressive, just not in the way the VC driven discourse/hype would suggest.