My Context Window is FULL

I'm glad that thanks to AI I now have a phrase to describe this feeling in a very specific way...

So what a lot of people don’t know about AI (speaking specifically of LLMs) is that their context window is essential to their ‘intelligence’. They are trained on massive amounts of text, and are very complicated in their ability to respond to specific prompts. The context window is, essentially, their short term memory. It is just the recent record of ‘what were we talking about again?’ The important thing about this is that LLMs don’t LEARN anything - not in real time, anyway. So when they seem to be responding to your question with the context of what you’ve been talking about, they are actually just responding to a long list of tokens (words, essentially) that are now joined by your most recent prompt.

Here’s the critical piece of information for understanding: classic LLMs concatenate whatever prompt you’ve just entered to their context window (at the simplest level that’s just a concatenation of all your prompts and the LLMs responses - UP TO A CERTAIN AMOUNT OF INFORMATION WHICH IS REFERRED TO AS A ‘CONTEXT WINDOW’)

When the LLM runs in to the end of their context window - sometimes - the wheels fall off. Think Memento.

I will say that this is a very active area of development and research, so there’s lots of tricks to help out with this: making the context window larger, compressing a big context window in a smart way to preserve vital information, multiple agents that maintain their own context, files that are saved and then re-read when needed to serve as long-term memory. But really this is an essential and fundamental limitation of LLMs. The LLM can appear to be an entity with something that seems like consciousness because it’s basically responding (statically) to a longer and more complicated context window.

Anyway, at the base I’m not sure this is much of a revelation? We people (real folks. you know, not artificial) have been dealing with this for a long time. Context switching? That is a direct analogy to this. We keep a bunch of information in our active memory. Then the context changes and we have to access our long (or longer) term memory to re-understand what’s going on in the current conversation. It’s really a pretty close mirror of what the more advanced LLMs and agents are doing to maintain continuity with a wildly chaotic biological thought machine.