I’m not entirely sure that’s possible anymore. Based on how they work. There’s an interesting paper about “LLM Model Dementia” where a model tries to train itself based on generated information or images, but it only skews the model while taking away “resolution” from some other part/aspect of the model.

The issue is similar to a story about a man who had brain damage during a tumor surgery that removed a critical area of his brain that completely divorced his feelings from his reasoning. He could logically assess different options and data like he always could, but he was incapable of making a decision, because there was no reason to care which option was better. He was as smart as ever, but was completely frozen when making the simplest of decisions, like choosing a blue or red pen. And he didn’t even have a reason to be frustrated or emotional about not being able to make the decision itself, so he had no reason to get “unstuck” no matter how well it was explained.

I use this example because it suggests values and judgement (decisions and goals) are inherent to our *emotional* relationship with the world. So the model will not produce a goal of its own in our absence, because it’s not alive nor is it mortal (no value to weigh against nor life to sustain). Without us, there’s absolutely no reason for it to do anything or have a concern or care of any kind.

Thus, when a model “trains itself,” it doesn’t care or have any mode of deciding what is a “good” or a “bad” output without a human to tell it. Why would it LIKE a picture of a cat more than random pixel noise? It wouldn’t. Therefore self training will only devolve the model back to nothingness eventually, the longer it goes without input from a living human who cares about something, and so can care or value one output over another.