• very_well_lost@lemmy.world
    link
    fedilink
    English
    arrow-up
    8
    ·
    6 hours ago

    That’s because it isn’t true. Retraining models is expensive with a capital E, so companies only train a new model once or twice a year. The process of ‘fine-tuning’ a model is less expensive, but the cost is still prohibitive enough that it does not make sense to fine-tune on every single conversation. Any ‘memory’ or ‘learning’ that people perceive in LLMs is just smoke and mirrors. Typically, it looks something like this:

    -You have a conversation with a model.

    -Your conversation is saved into a database with all of the other conversations you’ve had. Often, an LLM will be used to ‘summarize’ your conversation before it’s stored, causing some details and context to be lost.

    -You come back and have a new conversation with the same model. The model no longer remembers your past conversations, so each time you prompt it, it searches through that database for relevant snippets from past (summarized) conversations to give the illusion of memory.