Grant Harvey

So it seemed like it was taking reviews about a completely different website that was like a SaaS tool and they were conflating it with our website, which was like a media publication.

And like it was taking the bad reviews from the SAS tool and it was applying it to our media.

And that was just right there in the AI overview.

So that's like a huge problem.

Right.

Like because now people could then read that review and then they could think that they're taught it's talking about, you know, a different company just because I had similar names.

They YOLO it till they get more money.

Pretty much.

I mean, do you think this gets better with reasoning models?

Like, better reasoning models?

Or, like, will reasoning models be able to, like, parse that problem?

Or reinforce it?

Yeah.

There was the other one recently too, right, Corey?

There was the one where OpenAI basically went in and said, look, we figured out why hallucinations happen.

It's systemic based on the reward function.

Just for people who haven't read it, essentially they're like, we basically incentivize it to BS, like what Corey said, because we don't give it points for saying, I don't know, or we don't give it partial points for trying, but then realizing it's wrong.

Yeah.

Well, that actually reminds me, Corey, if you don't mind, like there was the article that I just read recently that was from 12 Grams of Carbon talking about the problem, which is the very scientific term of inshedification, which is basically how over time our platforms get worse and worse and worse.

And he basically wrote it down where it's like there's competitive reasons for that, you know, like and both sides, both the platforms and the users force that mechanism.