Hacker News Reader

Comment by sdrinf | Hacker News Reader

This iteration of shake-up is not strictly about tech superiority, though. Google probably already has a repro of ChatGPT running in their datacenters.

Fundamentally, it's about product, and expectation shift. I am presently routinely having conversations, and for the first time in the history of the Internet, getting straight answers to moderately complex questions, and their follow-ups, in real time, from ChatGPT. Google's entire business model is "giving people pages on the Internet that might contain those answers", and some of those being paid for the user's clicks.

If I get the full answer, instantly, wrapped in 2 paragraphs of text.... where do the ads go? Editorials inside the answer? Why would I click them?

View on HN

How do you verify that ChatGPT’s answer is accurate (it often enough isn’t), if not by googling?

You could use any search engine, and just gather 'decent' results -- scrape the first 3 results and compare using a new instance of a language model to tell you what's inaccurate in the first one based off the fine tuning data.

I'm wondering what happens when most people start using chat bots and stop using search. There'd need to be a big shift from using user's navigation as a signal and for search engine to determine the quality and authority of the content.

How would the next gen web be? Would people even create webpages anymore? How would these be funded if people aren't navigating to these sites and just using chatbots.

Google's already given up on search, google any programming topic and you get stackoverflow-scraped spam everywhere even BEFORE the stackoverflow post it was scraped from. Google's no better than altavista circa 1997 at this point imho. Brave search is better for most things, and if it misses then consult google, etc.

Yeah, if I can't trust the intial answer, I also can't trust any of the refinements. It will be a long way before you can get a similar level of assurance by asking a number of different language models as you can get now by looking at various sources around the web.

what if you google something, then check the links that google brings up that seem most likely to help you, it remembers that (reinforcement learning), 10 other people search almost the same phrase and pick maybe 60% of your links, after 100% and maybe a decent idea of which links are the best, it'll just return the links along w/ a summary of the details from each of them, without overlap, so maybe one page has top 4 x, and the other top 3, and another top5, well it'll give you a venn diagram summary of those 3 pages and maybe of the 12 listings, 3 are common so it only summarizes the 9 unique ones.

Reply on news.ycombinator.com