Live data from Hacker News

Generative AI could make search harder to trust

wired.com

161–170 of 255 posts

Re: Generative AI could make search harder to trust

#161

Earlier quoted context omitted.

People are vastly everestimating how unique this problem of hallucinations is. It seems to me it relies mostly on discounting just how much we've already had to deal with this same problem in humans over the millenia. The problem of proliferation of bad information might be getting worse, but this isn't native to generative AI. The entire informational ecosystem has to deal with this. GPTs compound the issue, but as…

humans can only produce semi-convincing bullshit at a limited rate with AI this limit is all but removed all the human generated bullshit ever created will soon be dwarfed by what AI can vomit out in an hour

How do you know that? A human that produced fully convincing bullshit would be definitionally undetectable to you

Re: Generative AI could make search harder to trust

#162

Earlier quoted context omitted.

Uh, no. Generative AI does not have a standard of truth. It generates text according to the probabilities given what it learned from its data set. There is no systematic modelling of a world that it can test its assertions against. What are called hallucinations are a fundamental property of the approach. The hallucinations are probable -- just not true. The model has succeeded but the assertions are false. This has…

> the human method results in fewer routine hallucinations. I'd love you to produce data to back this up. My guess is that you are wrong, on the basis of how often I discover that I'm full of shit and how often I discover other people are full of shit. Humans are built for being wrong just as much as being right. We wouldn't have such complicated institutions and social structures built around controlling for those s…

You could probably just look up data on mental illness and/or pathological lying? Most people will say "I don't really know the details of that" rather than write you an essay of seemingly plausible but completely made up nonsense. Not all people, sure, but most.

Re: Generative AI could make search harder to trust

#164

I actually experienced this the other day. Bought the new Baldur's Gate and was wondering what items to keep or sell (don't judge me, I'm a pack rat in games!) I had found some silver ingots. The top search result for "bg3 silver ingot" is a content farm article that very confidently claims you can use them at a workbench in Act 3 to upgrade your weapons. Except this is a complete fabrication: silver ingots exist onl…

Probably in the future people will only trust sources of info that can't be monatised. If you want to know the answer to a game question you just got to the reddit or discord and ask, since there is no point autogenerating crap for discord when you can't put ads next to it and the mods can remove you.

Sadly Reddit is sort of monetized, as we have people selling accounts for what I believe are spamming and propaganda purposes.

Re: Generative AI could make search harder to trust

#166
post #136

I think it'd be kind of neat in a backward way if we went back to the 'specialized encyclopedia' days of the 90s. Web directories , 'Who's Who in Engineering' type lists, etc. It's a step back from universal search engines being able to find stuff, but it's a step forward with regards to curation and quality of results; so i'm not sure if it's entirely a downgrade. The early 90s 'website phonebook' type encyclopedias…

Sites like Reddit are kind of like web directories with its user base collectively being the curator but it can be gamed too.

Re: Generative AI could make search harder to trust

#167

Earlier quoted context omitted.

Back in the 90s when the Internet became a thing, it was common knowledge that because normal people made websites, that you should take things with a grain of salt. There was a bit of an overreaction to this, as the general feeling at the time was to trust nothing on the Internet. In the 00s and 10s, the quality of discoverable content improved: reddit and stackechange had experts (at a higher rate than the rest of…

Reddit never had experts. Maybe for half a minute. It became an echo chamber fast: fake internet points to be gained for saying what got upvoted last week, or to be lost for saying anything different.

If all you do is browse (default) home or all, then sure, it's just a stupid echo chamber obsessed with hating the things its cool to hate. That's not where the value is on reddit, and it's not what people are referring to when they say the search reddit for answers. If you're looking for product reviews, it's not perfect but it's tough to find anywhere better unless you happen to know of exactly the right hidden gem of a forum to visit for your particular subtopic (and the link to that hidden gem of a forum is probably easier to find on the relevant subreddit than it is on google).

Re: Generative AI could make search harder to trust

#168

Earlier quoted context omitted.

Generative Wayback Machine

Honestly, I would love to see a LLM based plugin that would take a website, remove all the tracking garbage and filler, and just give me back a no-frills static html+CSS site that looks like it was made in 1995. "Oh, you want a guide to writing your own loss function for Tensorflow? Here's an FAQ that could have existed on comp.lang.python3.tensorflow"

So you want a... reader mode plugin? How would it use an LLM beneficially?

Re: Generative AI could make search harder to trust

#169

Earlier quoted context omitted.

The thing is when you call a human on bullshit, they usually can't back it up well enough to pass the smell test. When you call an AI on bullshit it can instantly fabricate plausible, credible seeming sources/evidence. A human's lie is different than an AI's hallucination, since it's still based on (distorting) the truth, whereas the hallucination is based on an invented reality (yes I know it's applied statistics an…

Intelligent people can fill the void of ignorance with plausible sounding but factually incorrect information. They are apt to engage cognitive biases in such a way that the biases produce assertions that are deeply indistinquishable from factual assertions. They fool themselves in this way and they fool others. This happens all the time. LLMs are no different in this respect.

Sure people lie and spread misinformation (willingly or unwittingly). But coming up with plausible sounding lies take effort.

LLMs on the other hand are amazing and prolific liars and can produce a lot of bullshit for a price that’s effectively free - in fact it’s cheaper to create a LLM that’s inaccurate than one that’s … less inaccurate. The truth to lie ratio on the internet is about to take a huge hit. LLMs really are a Pandora’s Box.

Re: Generative AI could make search harder to trust

#170

Earlier quoted context omitted.

Reddit never had experts. Maybe for half a minute. It became an echo chamber fast: fake internet points to be gained for saying what got upvoted last week, or to be lost for saying anything different.

Two things can be true at the same time: "Reddit is prone to karma-driven bullshittery" and "Reddit content is generally significantly higher-quality than SEO content farms". With Reddit you might get inane arguments and bandwagoning about what the best game strategy is, but you're exceedingly unlikely to read about a game mechanic that was straight-up hallucinated by a LLM.

For now. Wait till legions of bot redditors infiltrate real subs and make their own.
Post reply on HN