[bob1029]

> And we are a long way from full retraining on commodity hardware.

This is the part where I am wondering. If we are clever with our architecture, the part that needs retraining may not necessarily be the LLM each time (or ever).

Training a binary classifier to detect a new situation is way more efficient than retraining or fine-tuning GPT4. You (or the AI) could train thousands of these models in just a few hours on commodity hardware. How much "intelligence" or capability could emerge from a tree of 1000+ binary choices evaluated over every input prompt at every turn? What are the implications of being able to retrain the entire classification front end on every turn? What if all statistics could be reflected by the LLM?

Think about dynamic classes proposed by the LLM at runtime that are then automatically trained on relevant data. That's where it starts to get a bit scary for me. E.g.:

> I propose that I add a new binary classifier to contextualize prompts that result in some measured outcome. If I detect a future prompt probably results in this outcome, I will add the following context to it: "..." If the confidence for this classifier ever measures below X%, it should be deleted.

A human could inspect this sort of system a lot more reliably than with other proposals.

[brucethemoose2]

You're talking about seperate networks picking LORAs and such? Loading a huge number of embeddings or swapping between sub/additive networks or whatever is not very efficient either, and right now the LLM itself is quite monolithic.

And training such a decision tree seems like it would take a very long time.

The OpenAI CEO has talked about using subnetworks instead of an omegamodel like everyone is using now, and that would make lots of what you describe (and straight up anonymous finetuning) more practical.