>It definitely learns things.

By the very way this technology works they can't learn anything after training. What you think is "learning" it's just a session log written back to the context when you resume the session.

It's literally called in-context learning. The fact that it doesn't go into weights and is retained only for a session doesn't mean no learning occurs.

Also agentic systems may choose or be instructed to retain some information between sessions in files and/or databases which is also a form of learning.

There are experiments with retaining session information in weights in some form of lora but there's no consensus if it's even desirable. There's a value in being able to start from a clean slate.

And what is written when you wake up from sleeping?

It's absolutely not the same. If you think llms and brains works the same way you clearly don't know how either works.

For a LLM learning what you wrote your last session would be update the weights with the new relationships and factual knowledge created in the session. That doesn't happen. The weights are static and fixed after training. There's no online training in the transformer architecture or any variant. If the weights don't update, the network doesn't learn. Period.

These things are a centrifuge seperating humanity into people who say things like "it definitely learns" and those who do not.

It can't do a thing. You tell it how to do a thing. It now can do a thing. That's the definition of learning. Your centrifuge separates people who are not afraid to believe in what is reality, from the ones that do.

My car can't do a thing. I operate it's controls. It then does a thing.

In fact, I put a brick on the pedal and it went off and did a thing all by itself. Spooky.

Great. You astutely observed that your car, unlike an agent, in fact, does not learn, because you operate it, instead of teaching it, how to perform a skill, by demonstrating how it should be performed.