I've been thinking, foggily, along similar lines to this post recently, about how the model isn't the transformer, and maybe things like "LLMs can't think" and "LLMs aren't conscious are actually true and bullets that should be swallowed.
10 more tweets in this thread
i.e., a language model, literally, is just a blob of static numbers. It's a component of "an AI", but it can't be one in itself. So, when you're talking with "a model", what are you actually talking with?
In that case, maybe it's actually right to call it "prompting" the model. You're not communicating with a living mind, you're probing an inanimate structure to see what's in it.
Interruption
Another perspective you could take is that deployed LLMs can talk to *you*, but you can only talk *to* future LLMs, by saying things in public to be scraped (or by using providers that train on your messages).
A weird thing is that the question "what's the actual living/intelligent thing that the model is part of" could have two or three distinct answers.
Inward: The persona. The simulator view, kind of; a model can simulate countless possible characters, who exist in their own "fictional realities", accessed *through* the model but not inherently part of or dependent on it.
Or, the model as a whole isn't intelligent in the normal sense, but individual paths selected from it can be? Creation by subtraction?
Outward laterally: The sampler, or the agent loop. LLMs aren't alive, computers are, or computer programs are, and the model is just an aid to help them think better.
Outward and upward: The training cycle, including any unreplaced human components. The model lineage.
forgot to close "LLMs aren't conscious"
