"But chatGPT would actually write useful texts, most of them more useful than the average blog post. Over a few years they might be just as good as the best human articles."
If web sites start being written by A.I. and ChatGPT starts getting trained on A.I. output, it's going to start spewing garbage results.
It's only useful if trained on human input, A.I. have no model of the world and no A.I. can function if trained by A.I. written texts.
I think that only holds true if :
1) there’s no editorial influence on the chatgpt content - I.e., sanity checking.
2) humans decide they don’t like writing any more, which given how many humans like to write seems unlikely
3) the chatgpt echo chamber doesn’t reinforce non-garbage and resonates in some way that diverges to garbage
There’s also a weird assumption that humans don’t write a bunch of garbage as is.
I don't understand the claim that editorial influence will fix things when the above claim is Google can't distinguish between legit content and SEO garbage. Does google not have smart people working for them who can use editorial influence?
As to point 2 , the argument is that, like with google search results the dataset will be corrupted and get worse, not that it won't contain some human written prose, so that's irrelevant when you say humans will still write things.
Claim 3 seems to be an argument that we can solve the problem of machine learning being based on bad bots by throwing more bots at it, which seems to be a very Rube Goldberg proposal which I'm not sure will work.
Finally is the claim many humans are dumb, so the more chat bots that impersonate humans, the better I guess?
>we can solve the problem of machine learning being based on bad bots by throwing more bots at it
Yes, you can solve dataset problems with models. Look at ConstitutionalAI from Anthropic. They only write a set of rules (the so called constitution) and the AI will derive a large dataset of demonstrations all by itself. This can do away with the RLHF stage where OpenAI has the best results so far, but costs money to label and is harder to update.
The model "Claude" from Anthropic is being discussed recently on Twitter
If web sites start being written by A.I. and ChatGPT starts getting trained on A.I. output, it's going to start spewing garbage results.
It's only useful if trained on human input, A.I. have no model of the world and no A.I. can function if trained by A.I. written texts.