Comment by visarga

Not a JPEG and not a search engine, it is more like a database. A JPEG is just a static approximation, a search engine has efficient retrieval, but a LLM can also do complex data processing, like a neural information processor.

> But I’m going to make a prediction: when assembling the vast amount of text used to train GPT-4, the people at OpenAI will have made every effort to exclude material generated by ChatGPT or any other large-language model. If this turns out to be the case, it will serve as unintentional confirmation that the analogy between large-language models and lossy compression is useful.

This shows the author has not been following closely. There are many ways LLMs have been used to improve themselves. They can discover chain-of-thought justifications, they can rephrase the task, they can solve problems and ensemble many predictions, or sometimes we can use math or code execution to validate their outputs. If you give it three problems and solutions as samples, it can generate another problem and solve it, adding to the training set.

RLHF for example uses generated data for the preference labelling task. ConstitutionalAI does reinforcement learning from AI feedback instead, using both the generative and discriminative abilities of the model.