Hacker News Reader

Comment by Imnimo | Hacker News Reader

Suppose that instead of a 100B+ parameter neural network language model, I had a website with a simple bigram language model that, after drawing enough samples, produced a similar output. Would anyone bat an eye? Would anyone bother to write a USA Today op-ed about the incident?

I think a reasonable person would immediately understand that the outputs of the bigram model were not statements of fact. Is there some threshold of language modelling performance where this understanding is no longer expected?

View on HN

I don't expect ChatGPT to say true things. But I have asked complicated technical questions that it provided detail, correct and useful answers to. And I have asked other complicated questions and it's provided plausible sounding whole-cloth fabrications that it further gave references to.

Knowing human beings, this is the level of plausibility where people start to believe the thing. People aren't good at seeing plausible arguments as out-and-out fabrications.

So, yes, this is past some threshold despite being far from AGI. I mean, Google is probably pasted that threshold too but Google only gives existing websites (which can be filled with lies too).

A workaround i've found is to always reply with "no you're wrong" for each seed reply. This forces it to either add more detail or in some cases give a completely different answer. If the replies are in the same ballpark, there is a good chance its not hallucinating. If it gives a completely different answer you are better off rewording the prompt or scrutinizing the answers more than you typically would.

Somehow everyone takes comments like my one above as being "my problem" with ChatGPT and offer "tips". Sure, there might be ways to correct the problem when it's gone off the rails and you happen to know that. That might make it more useful sure.

But that's not the point. The point is that average people are going to ask things they don't know the answer to and put some credibility in it's answer because it's fairly good at seeming right and at some percentage being. Me improving my approach cannot change that. Tips on improving performance won't stop that. etc.

Based on that logic, if you want to sue ChatGPT for incorrect information, then you should sue Google for giving you incorrect information too (if it has presented you lies or dangerous materials).

Google does have to deal with sites devoted to systematic libel and extortion. I don't think there have been cases but there have been articles on the subject and after enough pressure is downranking extortion sites.

I wonder if fabrications by LLMs like ChatGPT will end up becoming a net good for society.

What I mean is, maybe we'll collectively (re)realize how absolutely vital truth and critical thinking are.

I suppose whether we do depends on whether people in positions of influence begin to emphasize their importance in their writing and speaking.

...maybe we'll collectively (re)realize how absolutely vital truth and critical thinking are...

If two hundred years of scurrilous journalism didn't teach us this, I don't see why one lying machine would do so.

Good point, convincing people to change their mind is not easy.

Some cable news viewers want entertainment about labs and lasers rather than facts and a chat bot is incredibly entertaining. A chat bot can tell them the earth is billions of years old but they’ll just exit the chat, start searching and find it on some conspiracy website.

> I think a reasonable person would immediately understand that the outputs of the bigram model were not statements of fact.

Interesting considering the context that you'd expect people to know what a bigram model is, or how it would be different.

Any other kind of model isn't any less "dangerous" to unreasonable people like the blog post writer, it's just more obscure, especially right now, than ChatGPT.

I'm not fan of conservative law professors generally but I can't see what's unreasonable about the argument he's making here. Broadcasting lies about someone is bad for them and the "those fools should know this is bs so I'm not responsible" defense is itself bs.

Edit: While I might not agree with Turley, his wikipedia bio makes him sound far more principled and consistent than the average "public intellectual" today.

"those fools should know this is bs so I'm not responsible"

Is actually something you would absolutely argue in a US defamation trial. Defamation damages need to stem from people actually believing the falsehood. If the bad press from a false statement leads to someone, say, losing their job, those are damages, but the plaintiff would need to prove it was because their employer believed the lie.

the only one broadcasting the lies seems to be Turley himself: chatGPT didn't share the conversation with the internet, he did

Google autocomplete does make headlines, regularly.

https://www.google.com/search?q=google+autocomplete+usatoday

Shoot someone with a water gun in the outback vs a gun in the city.

One idea is to let them buy liability insurance, and watch the price go down over time.

Reply on news.ycombinator.com