Hacker News Reader

Comment by gizajob | Hacker News Reader

Haha in the demo he asks "when is the next solar eclipse and where is the best place to see it?" - The AI responds correctly that it's on April the 8th 2024, but then clearly hallucinates like crazy and says "the best places to see it are in Exmouth, Australia and East Timor" which is totally incorrect - this eclipse will be visible only in North America, and invisible in Australia and East Timor. Good job he didn't ask it to book flights to Australia on the 7th of April.

You'd think your tech demo would check to see if your AI was hallucinating!

View on HN

It also says that amount of almonds has 15g of protein, which would actually be like 50 almonds according to a few different online nutrition sites.

Cant believe they left this stuff in.

Gotta at least make it seem good in the commercial, this ended up being the opposite of a sizzle reel

I think people are very eager to believe AI is useful in ways it often isn’t. Kind of like the crypto hype. People were willfully ignorant to how little sense it made in so many contexts. The common thread here is “yeah, but money!”

I’m not an AI detractor. I use it and really like it. I just don’t like it for information like this. Anything where the response needs to be verified yet is very brief makes no sense to bounce off of an AI, in my opinion.

Kinda shows you what their target audience is.

Google received significant backlash for this when demoing Bard: https://www.bbc.com/news/business-64576225

Because of stuff like this I don't get how you can trust it not to hallucinate when it summarizes your inbox.

IME LLMs hallucinate much less when given a text that it can summarize or answer questions about vs. asking it to generate something

Yep, the eclipse answer was incorrect (1).

You’d think they’d have learned their lesson after Google Bard’s hallucinated demo!

(1) https://www.space.com/33784-solar-eclipse-guide.html

this is aside but I really don't think you need to use a footnote if your comment is two sentences and a link

I appreciate it and do it myself. If HN supported hyperlinks of our words the way html does we could just do that, but footnotes are the next best thing. Big unreadable links in the middle of what you're saying, even just one sentence, muddy the message.

I was also gonna link to the Bard article, but was lazy.

Reply on news.ycombinator.com