Hacker News Reader

Comment by roryirvine | Hacker News Reader

I've spotted surprising amounts of confidently-stated nonsense even in fairly neutral articles where Elon / xAI is unlikely to have a particular political slant.

Many of the most glaring errors are linked to references which either directly contradict Grokipedia's assertion or don't mention the supposed fact one way or the other.

I guess this is down to LLM hallucinations? I've not used Grok before, but the problems I spotted in 15 mins of casual browsing made it feel like the output of SoA models 2-3 years ago.

Has this been done on the cheap? I suspect that xAI should probably have prioritised quality over quantity for the initial launch.

View on HN

Some time ago, there was a project called Citizendium that aimed for quality over quantity, with articles written and peer-reviewed by subject matter experts who had to use their real names and working email addresses, among other requirements. I always thought that was interesting, since the main critique of Wikipedia is its open editing model.

Citizendium is still around, though they've loosened some of the requirements in order to encourage more contributions, which seems self-defeating to me. I think they should have tried to cooperate with Wikipedia instead. The edits and opinions of subject matter experts could be a special layer on top of existing Wikipedia articles. Maybe there could be a link for various experts with highlights of sections they have peer-reviewed and a diff of what they would change about the article if those changes haven't been accepted. There could also be labels for how much expert consensus and trust there is on a given snapshot of an article or how frozen the article should be based on consensus and evidence provided by the experts. This would help users delineate whether an article contains a lot of common knowledge or whether it's more speculative or controversial.

When I was a kid you could subscribe to an encyclopedia. They were too expensive to be bought in one go and they were too expensive to make in one go. The solution was to sell them in installments and mail you the monthly addition that you could add. Obviously the marketing ploy was that if you were halfway through the 'A' that you would buy the rest of the A and once you had the 'A' then you'd buy the rest of the book.

Regardless, the business was there. Wikipedia killed all that. So if you want to create an expertly created encyclopedia anno 2025 you have a real problem: you will need to pay experts for their time somehow otherwise why would they compete with the million monkeys, but your source of revenue has been strangled by those very same monkeys, who it turns out produce content that is orders of magnitude better than anything I've ever read in a for-pay encyclopedia from before Wikipedia.

The bar to entry is insanely high.

> I suspect that xAI should probably have prioritised quality over quantity for the initial launch.

I mean, I don't think this is _for_ people who care about quality, tbh. For those, there is wikipedia. This is more of a safe space for Musk.

> I mean, I don't think this is _for_ people who care about quality, tbh. For those, there is wikipedia. This is more of a safe space for Musk.

Wikipedia isn't for those who care about quality, either. It's still quantity over quality, just not as badly as this LLM garbage.

[citation needed]

But it is much, much better than the encyclopedias that it replaced.

> safe space for Musk

> I've spotted surprising amounts of confidently-stated nonsense

I find this to be the most annoying aspect of AI. The initial Google AI results were especially bad. It is getting better, but still spout info I know is false without any warning.

Like, I find blowhards tiring enough in RL. Don't really want to deal with artificial blowhards when I'm trying to solve a problem.

Reply on news.ycombinator.com