That’s kinda overly anthropomorphising. They are a statistical model that uses the number of occurrences of tokens in association with each other to predict the most likely next token. You can do that with pretty much any training data and any text. The current very large models will have almost any kind of text in their training data, including unhinged manifestos, racist tweets, etc.
Since models are stateless, getting them to produce that kind of output doesn’t change or worsen them in any way.
It’s not that simple. You can adjust outcomes by weighting. You can adjust outcomes in a myriad ways. Yeah, sure, in a trivial sense it’s just statistics, but that’s like saying that computer communication is just a bunch of ones and zeros. It misses a lot of the point.
It really gets way more complex and most certainly can be adapted and tuned in various ways.
That’s kinda overly anthropomorphising. They are a statistical model that uses the number of occurrences of tokens in association with each other to predict the most likely next token. You can do that with pretty much any training data and any text. The current very large models will have almost any kind of text in their training data, including unhinged manifestos, racist tweets, etc.
Since models are stateless, getting them to produce that kind of output doesn’t change or worsen them in any way.
It’s not that simple. You can adjust outcomes by weighting. You can adjust outcomes in a myriad ways. Yeah, sure, in a trivial sense it’s just statistics, but that’s like saying that computer communication is just a bunch of ones and zeros. It misses a lot of the point.
It really gets way more complex and most certainly can be adapted and tuned in various ways.