2 ms·
Sure, but nobody uses an LLM in isolation. They use systems that are built around an LLM. In many of those systems, like a chat interface, the context is almost
by nomel 2mo ago
Sure, but nobody uses an LLM in isolation. They use systems that are built around an LLM. In many of those systems, like a chat interface, the context is almost always used a form of memory. I think I was correct that your disagreements based on the physical location of where the previous context is stored. At the functional system level, with nearly any tool you or I use, the context/RAG/etc is used as memory, loaded with relevant/past symbols.
Say I created a new LLM architecture that integrated this concept of only appending to previous context, with no way to clear it. With it now being physically closer/integrated with the bit of code you call an "LLM", would you then agree it was memory? I don't see that implementation detail being all that important.
- rrr_oh_man 2mo agoDude, you’re just going down a rabbit hole with imaginary arguments. Nobody’s talking about your weird location argument.
- nomel 2mo agoI think you miss the simple point I'm trying to make: context can be a practical form of working memory. You said > There are no past experiences. I'm saying that if the context contains something added in the past, it's functionally no different than a "past experience". For example, a context that is a chat log you had with the LLM. If you tell the LLM to remember a word, continue your conversation, and then tell it to recall that word, the system can recall that past experience. That "memory" that is the context feed to the LLM is a form of memory, in some cases. Now you say, "but that chat log storage that you send to the LLM is not part of the LLM!" and then I say "see above".