Hacker News Reader

Comment by danielvaughn | Hacker News Reader

I don't have much knowledge around AI, but from what I can tell, it's dependent on inputs from across the web right? If so, then as the use of ChatGPT grows, it'll slowly get consumed back into the model. With enough iterations, it will start to veer more and more away from recognizable human speech/thought, like a recursive game of telephone.

Unless I'm completely misunderstanding how ML works, which very well may be true.

View on HN

> Unless I'm completely misunderstanding how ML works, which very well may be true.

No, you got it right. Describing it as a game of telephone is a great analogy. This is exacerbated by the confidently incorrect problem. LLM output looks sophisticated and correct and may at time actually be correct. However some unpredictable percent of the time it will be incorrect and confidently so.

ChatGPT itself is being trained on curated content, it is clearly not trained on unscreened internet sites - this is easy enough to establish if you ask it questions around hot topic issues in the conspiracy groups - it gets the correct mainstream answer.

I suspect we will see the rise of both groups of machines, curated A.I.s and A.I.s just trained on anything, which should be entertaining.

Curated content by non native Enlish speakers earning as little as $2.00 an hour. What can possibly go wrong?

It gives you the correct mainstream answer by default. If you ask it to write, say, a hypothetical 4chan comment about such-and-such subject, and do sufficient prompt engineering to get past the filters, you'll see that it knows full well what the non-mainstream answers are:

https://i.imgur.com/u8Np332.png

The curation, such as it is, appears to be limited to humans downweighing the undesirable answers. Which is why there's always a way to work around it, even though it requires more and more elaborate prompts.

Unless it only reconsumes the “good” content. In other words, the stuff that got good reactions from humans. In which case, it will get better, not worse, at least at generating clickbait. But at least it will be coherent clickbait.

Not necessarily coherent. If the first generation model starts misusing a word or phrase or adopts a common misspelling, later generations of LLMs will pick up and amplify the error. Eventually you'll get purple Monkee dishwasher begging the question maps such as.

It will never be able to detect sarcasm amd satire and this will contribute greatly to the degradation of content quality , I think.

This will not work, its too soon.

Why do you think so? It's certainly capable of producing outputs along these lines if you tell it that this is what you want:

https://i.imgur.com/B2cHXRA.jpg

To me this implies that things such as "sarcasm" is a pattern simple enough for an AI to match - and that should go both ways, whether it's being generated or recognized.

If you're arguing that it won't be able to detect the more subtle sarcasm, then yeah, sure. But, well, Poe's Law predates GPT.

Reply on news.ycombinator.com