alt text
Crudely drawn drawing of a person telling a computer “say ‘i am in pain’”. The computer replies with “> I AM IN PAIN”. The person then says “oh my god.”
Crudely drawn drawing of a person telling a computer “say ‘i am in pain’”. The computer replies with “> I AM IN PAIN”. The person then says “oh my god.”
Don’t ‘really change’ you or don’t change you at all? Are you the sum your experiences?
It does not change, ever. From the moment it starts processing, to the moment it finishes processing, it is still the same model. It has access to all of its input the moment it’s brought into existence, it doesn’t experience the prompt linearly, the input is it’s raison d’être.
I would generally agree that a llm could be part of a larger system capable of experiencing things and having difficult conversations about it, but I don’t think these harnesses are it. They are very simple systems, compared to the llms they control. I think emergent behavior would require something much more complex, given what we know about life.
Okay look, the harness question is a whole nother one. I just want to know, if you consider the model + kv cache together, does that match your definition? The KV cache can change and the model’s internal state and outputs can change in response to it.
KV caching doesn’t persist between prompts, it is an internal optimization used while generating text. I don’t see how it’s relevant.
Edit: it kinda sounds like you’re technobabbling me, like “but it uses a recursive algorithm, it must be alive, check mate!” Your question is a non sequitur.
I’m not techno babbling you. I specifically said “if you consider the model + kv cache together,” e.g preserved. A kv cache is not an internal optimization, a model cannot generate text without a kv cache. There’s a separate idea of caching the kv cache between requests as an optimization. That’s not the one I’m talking about.
Or if it helps, consider it just within one generation. The model + kv cache. Does that not match your definition?
Definition of what? Do I think that because llms use a kv cache during inference, that proves they can experience things? You’re gonna have to explain your thought process. To me, the particular algorithm used is not the philosophical crux of the issue.
I’ve never heard of this anyways. The kv cache is specific to a particular input, it can’t be reused for a subsequent prompts.
It can as long as there’s a shared prefix.
No, I never claimed to have proof they can experience things. I just think it’s possible and you can’t categorically dismiss it based on how they work.
I’m talking about your condition above which we’ve been arguing about for the last ten messages
Which so far seems like your only argument for why they can’t experience things
I’ve never heard of that. The cache is meant to support token generation, so unless it’s trying to repeat itself, I think the previous cache would be little use.
Ok, yes, that is my argument, now how the fuck does a temporary cache, thrown out after every response, prove that the model is changing as a result of undergoing inference?
I’m just going to quote myself at this point
OK, that’s not helpful at all. My answer is no, it does not meet my definition of changing behavior, and I have no idea why you think it would.
Edit: ok maybe this does make sense from a very stupid point of view, if you allow me to interpret what you mean: you believe because the kv cache is generated during inference, that proves the model is learning a new behavior, and if you save that cache, that proves… Something??
It’s absolutely silly: the model’s behavior is to generate a key-value cache, and that’s exactly what it did. The fact that the model is doing intermediate calculations based on the input is not evidence of a new behavior being generated, that is just how algorithms work. That’s what ‘processing’ is, you dunce.
Edit 2: I also just want to point out, if you think the word ‘cache’ is important, it’s really not. Caches are fundamental to how a computer works, no processing could ever occur without them: every cpu instruction involves reading from or writing to at least one L1 cache (register).