I am very dummy on LLMs, but wouldn't a confined model (no internet access) eventually just loop to repeating itself on each consecutive run or is entropy enough for them to produce endless creativity?
Loops can happen but you can turn the temperature setting up.
High temperature settings basically make an LLM choose tokens that aren’t the highest probability all the time, so it has a chance of breaking out of a loop and is less likely to fall into a loop in the first place. The downside is that most models will be less coherent but that’s probably not an issue for an art project.
The model's weights are fixed. Most clients let you specify the "temperature", which influences how the predictive output will navigate that possibility space. There's a surprising amount of accumulated entropy in the context window, but yes, I think eventually it runs out of knowledge that it hasn't yet used to form some response.
I think the model being fixed is a fascinating limitation. What research is being done that could allow a model to train itself continually? That seems like it could allow a model to update itself with new knowledge over time, but I'm not sure how you'd do it efficiently
The actual underlying neural net that the LLMs use doesn't actually output tokens. It outputs a probability distribution for how likely each token is to come next. For example, in the sentence "once upon a ", the token with the highest probability is "time", and then probably "child", and so on.
In order to make this probability distribution useful, the software chooses a token based on its position in the distribution. I'm simplifying here, but the likelihood that it chooses the most probable next token is based on the model's temperature. A temperature of 0 means that (in theory) it'll always choose the most probable token, making it deterministic. A non-zero temperature means that sometimes it will choose less likely tokens, so it'll output different results every time.
