The big benefit in my experience was that I could have a program with real users, that did have errors (from me being new to Elixir and not knowing better) and still not experiencing downtime.

Instead, CPU or Memory would increase over time, hit the VM limit, kill and restart.

So later when I noticed this, I could debug and fix it without simultaneously fighting a prod incident.

I wish the nix programming language wasn't so rough because it can be pretty great at this problem. Being able to compile from source while just listing out package dependencies is powerful.

Hmm... Earlier today I asked "Are you Gemini pro?" And it answered word-for-word the same way. Is this a hard-coded or heavily prompt-coached answer?

It's suspicious when an AI answers 100% the same.

Huh, same here, that “LOL” response was the exact same thing for me. Pretty weird.

When I expressed my surprise about its casual response, it switched back to the usual formal tone and apologized.

Not sure what to make of this as I don’t consider myself to be in the know when it comes to ML, but could this be training data leakage? Then again, that “LOL” sentence would be such a weird training data.

Yep, prototype exactly that this past week. With a strong instruction spec prompt from the start, you can have an AI come up with a much better answer by making sure it knows it has time to answer the questions and how it should approach the problem in stages.

The great part is with clear enough directions it also knows how to evaluate whether its done or not.

Can you share the kind of prompts you used?

It seems plausible you could have the LLM side call upon its knowledge of known problems and answers to quiz the q-learning side.

While this would still rely on a knowledge base in the LLM, I would imagine it could simplify the effort required to train reinforcement learning models, while widening the domains it could apply to.