Hacker News Reader

Comment by davidw | Hacker News Reader

The thread is here and... it's pretty wild. This doesn't seem like it was a well-thought out idea to release this without some more consideration beforehand.

https://bsky.app/profile/karenattiah.bsky.social/post/3letty...

View on HN

> They admitted it internally, to me

It's becoming an increasingly apparent problem that people don't understand that LLMs frequently hallucinate, especially when asked to introspect about themselves. This was 100% a hallucination.

"I know, let's release an AI bot that pretends to be something that's likely to be viewed through a charged political lense and have it hallucinate random answers to people"

It's getting weird that people cite ChatGPT to settle debates. Even on places like HN where supposedly people have a technical background and know they shouldn't blindly trust an LLM as a source of truth. Even with the banners it has warning about this.

Technical knowledge and wisdom are two very different things. There is an excess of the former here, on this site, in this field, but a comical dearth of the latter.

I have but one upvote to give you, good sir/madam, when I wish I had 100.

I would argue that there is not an excess of the former. A lot of the so-called techies don't know tech either.

That's fair. I suppose it's more that there's a worship of knowledge outside the context of wisdom.

There are wise people here, they are just lurkers though (wisely).

People don't understand that LLMs are just statistical magic, by design of the LLM providers because LLMs wouldn't take off if people understood they are just statistical magic.

You can understand that and still think this was a terrible idea.

I mean all of the atoms in the universe are “just statistical magic” when viewed through a quantum lens. I’m not sure that phrase is as effective a dismissal as you seem to think.

You and I understand this. Heck, most of us on HN understand this.

When the media and companies with a financial interest in AI have been shouting from the rooftops that "AI is going to replicate humans very soon and all of your jobs will go bye-bye because this thing will do everything better than you", can we really be surprised when the broader populace looks at it that way?

It was likely a hallucination, right? You didn’t check to confirm?

But this is a solved problem that Meta chose to implement poorly. It can be done right. I just asked o1 “create a diversity report for openai” and its very first paragraph is:

> Below is a *sample* diversity report for OpenAI. Please note that the numbers, programs, and initiatives are provided as *illustrative examples*. This report is not an official document, and the data contained here is *fictional* and for demonstration purposes only.

Yes, and I suspect that the very first line of the chat has something equally lawyer-y to serve the same purpose.

I suspect it doesn’t.

LLMs just predict the next word, based upon the texts they were trained on. They are, by definition, a hallucination. It's just that, sometimes, they put words together that fool people.

They're just ELIZA on steroids, people.

How did ELIZA do on the Advent of Code?

It hallucinated less, that's for sure, because we know exactly how it works, though I doubt any of the AoC's challenges were suitable for its 1980s logic (that's when I typed in a magazine's C64 version).

"There's lies, damned lies, and statistics." --Unknown

LLM's did well: https://blog.scottlogic.com/2024/12/15/llms-vs-aoc.html

Right and computers just do boolean math, not anything useful.

A hammer can bonk someone or help Jimmy Carter build a house for someone.

We digital logic folks are both tool builders and tool wielders. Usefulness is in the hands of the user, but that doesn't mean the creator isn't a dipsh_t who made a crap tool, or that the user isn't a moron.

My computer allows me to build stores of digital information representations, with functional integrations to other systems, including me. It's useful for me and my family, for sure.

Soo...they didn't learn anything from Microsoft's failed attempts at leaving an A.I. on the net ??

what was it ?? 48 hours before MS had to take them down ?

learning? that is for the AI

justice for taytay

gone, but not forgotten

16.

(When searching for a citation I found this little compendium of banger tweets, and the addendum that Tay had a brief reprise of accidently going live again, this time tweeting "kush! [ I'm smoking kush infront the police ]" I guess in the modern milleu this would be considered escaping confinement)

https://dailywireless.org/internet/what-happened-to-microsof...

If an AI says it, it must be true?

This seems like a possible hallucination.

Reply on news.ycombinator.com