Yes they don’t know right from wrong. Its fancy math for word averages.
You know they need human made content, and you know they’ve tried to train it on AI generated content and it just collapses in on itself.
2 Likes
That’s because they don’t learn much by interacting with the user. What @lphilpot refers to is that AI generated content is put out on the internet, at which point it gets ingested for the next round of training.
Even if it would take your input with such a gigantic weight that it would use it for the next prompt, this would never happen since the current ChatGPT model is always a few months behind. What is being trained through conversations isn’t what is being served to the public. You can ask him for his last update date for more info.
To be fair given how the models keep improving, we have to assume OpenAI, X, etc, have tools to weed out a lot of their own generated content before using it to train their models.
I wouldn’t bet on that. For images maybe, since they can be watermarked, but there’s no real way of telling if text is AI generated if it’s not explicitly tagged as such.
1 Like
Thanks to All. I think I’ll stick with ChatGPT 3.5 (but not to quote here) and accept that the training set is old and I believe not continuously updated. I might be wrong about that.
You shouldn’t rely on ChatGPT, that should be one of the last resort if ever. It did worked out for me twice with code-related things, but I verify and adjust until it works. But, I do not advice people to use it at all.
3 Likes
Don’t. Unless you enjoy playing intellectual Russian roulette.
2 Likes
It’s responses usually satisfy my simple questions.