LLMs do not learn

LLMs, "AI" in the common parlance, do not learn. Cannot.

I hear people casually mistake this all the time. "AI will learn from us and improve" they say, matter of factly, either speaking broadly or about their specific use case. They assume the model is learning on the fly as they interact with it.

Do you know how much of your conversation the LLM itself retains? 0%.

To be clear, there are ways to give the LLM access to relevant context along with your request. But it is not processing it the same as the parameterized data it was trained on, nor is its training being updated on the fly. LLM weights are frozen, the result of the training that created the model.

(There are open weight models, but that's weights open to programmatic manipulation before requests are processed, not reactive learning).

It's also important to understand that you can't just throw a giant amorphous blob of data along with every request because of context rot. If you're not familiar, this is why a conversation with an LLM can get goofier the longer it goes on. There is a limit to how much context an LLM can process and return a useful response without concepts getting garbled.

This is why tooling has quietly become a huge differentiating factor in AI performance, and the closest we have to "learning" is the strategic use of RAG (Retrieval Augmented Generation). The context of your request and any data retrieved from a RAG (web search, etc) then biases the distribution of the output.

So if you want LLM assistance in analyzing your company's internal data, you'll need a thoughtfully designed RAG that provides the relevant context per request. And with context rot, there is a ceiling to that.

Don't get me wrong, the Gen AI companies could store input from users and use it to train future models. It is against the licensing for most paid (but not free) tiers, but would be difficult to catch since there is no way to audit a model and identify provenance, you'd need access to the provider's raw input.

This is a big part of why I see LLMs as a stepping stone in AI and not the destination. I'm much more interested in watching the progression of active inference where the system continuously updates internal structure against live conditions, much more inline with our own intelligence.

We already have the proof that AGI does not require absurd amounts of compute, our brains. Anyone insisting that's the only path forward is telling you more about their capital commitments than about intelligence.