Exploring the Illusion of Sentience in Large Language Models
aarushgupta.com
aarushgupta.com
In reference that LLMs "don't understand anything and are stochastic parrots over their training dataset" refer to in context learning, llms can be taught new concepts not present in their training data.
I wonder what humanity could accomplish if some non-trivial subset of it could learn to consistently think in ternary logic.
Sentence, self-awareness, can only exist when you receive yourself as an input in some form shape or way.
These llms just don't do that. Technically recursive networks that we were using before the llms could, but these new models have no method but the context to understand their own state of mind.
And my knowledge they aren't currently trained on their own output when it comes to context, they're just trained on a data set that exists already. As a result they have no way of differentiating their own output from the input they get from the world. I don't see them being able to be sentient in any way as a result.
Which is not me saying they can't be, we just have to get to the point that we are training that into them somehow.
A fine argument could be made, but that argument rests upon the question of the degree to which you yourself are a LLM. You may not have any option of believing other than you do.
> Sentence, self-awareness, can only exist when you receive yourself as an input in some form shape or way.
Perhaps (you may not be able to realize that you only believe this, you don't actually know it), but how does one know it isn't being simulated in some manner by consciousness, which is further distorted by culture (ie: the culture we've been raised in severely underweights the importance of epistemology, and consciousness' "interesting" implementation of it)?
I made a mistake here in in using the word sentient the way I did. To help you understand here, I use these words as follows.
Sentient - to be aware of yourself.
Sapient - to be humanlike in your ability to think and speak
Conscious - to have a self experience of any form.
What I am saying is that an LLM is definitely not aware of itself. It doesn't receive itself as an input, so it cannot have thoughts like "I'm happy right now". It's functionally incapable of doing so because it has to receive itself as an input in order to do that.
However, it's possible that such a machine still has a experience of some form in the sense that there is some ethereal consciousness created by the fact that exists and is thinking, but we can't observe that aspect of the world so it's not really something we can discuss in a functional way beyond speculation.
That's the point - in fact, I would even speculate that you are may not be [currently] able to.
> I was fairly specific in speaking about the ways in which an LLM is different and the ways in which it lacks the functional capacity to be aware of itself.
Right, but don't forget that this is what you think, and what one thinks is typically what one "sees".
> What I am saying is that an LLM is definitely not aware of itself.
I tend to lean strongly in that direction as well, but I am a ~~strict epistemologist~~ pedant (in ~colloquial reality).
User: I've invented, with my friend, a new concept that is symmetrical to that of "universal": we've called it the "adverse." We're now exploring adjacent notions such as universe, universal, universial, and god. ...
Assistant: ...
4. God is often conceptualized as the ultimate universal, the unity that transcends and governs all realities. In this view, God would be the final organizing principle of the universe, the principle that gives meaning to everything. Conversely, in the adverse, there might be a concept like the Adieu (a pun on "adieu" and "God"), which would be the ultimate possibility, the principle that gives birth to an infinity of distinct and unique realities. The Adieu would not be an organizer, but a facilitator, allowing each reality to follow its own unique laws and to realize its own unique possibilities.
You routinely see arguments like "AI LLM models must be sentient because look at my cherry picked, very minor example X of one saying Y to me, just like a thinking intelligence!"
Or " Well if you claim LLMs are just Markov chains and sophisticated stochastic parrots, how can you know that human's aren't exactly the same thing?" Well, no it doesn't work that way. The one making the fantastic claim needs to justify it, not the person claiming have unique traits based on thousands of years of evidence.
And if that person can't even notice any difference between human self-directed behavior/reasoning and how something like GPT acts after being given ungodly amounts of text content to algorithmically filter through for a text response (and still often failing at answering even very simple questions), then that person is unlikely to use good reasoning for their notions of LLM sentience.