> "like an LLM" means no feedback loop, no refinement, just input to output, like an LLM.
But that's not what an LLM is, an LLM is a large language model. The distinguishing characteristics between an LLM and other AIs are not whether or not an LLM has a feedback loop. Lots of AI categories (arguably the majority of AIs) generate output from a single set of inputs as a single "operation" (ignoring the fact that most neural networks including LLMs internally have multiple layers and do in fact have multiple transformation steps, but whatever, we can treat that as one step for the purposes of conversation).
If anything, LLMs are the exception here in that they very often do have a feedback loop during normal usage; they are most commonly used in a conversational context where they generate the next "chunk" of a conversation after being fed back their previous answers alongside the followup responses of the person working with them. Arguably the feedback loop of an LLM is that as a conversation progresses, its future output is based on its previous output, which very literally becomes its new input after being marked up and extended by a human being. That is notably more feedback than many other AIs get.
Doubly so if you're trying to argue that an LLM can encompass a multi-modal setup, because suddenly refinement and feedback between multiple models is a core part of the final product.
So I'm not sure I agree that an LLM fits into that category in the first place, but even assuming it does, if your definition of an LLM is is just that it takes an input and turns it into an output as a single step without help... that is just such a broad category, I would guess the majority of AIs fall into that category. It's not what makes LLMs special; most predictive neural networks take a single set of inputs and generate a single set of outputs without intermediary human input. What makes LLMs interesting as a category is the training and structure and quirks of how they work, what makes them interesting is the differences between LLMs and other AI techniques.
To jump back to the Markov chain again, Markov chains do not have a feedback loop or refinement step: they take an input and map it to an output as a single step. Are Markov chains LLMs? Is any data transformer that operates in a single step an LLM? That is a really broad definition to use.
----
And we run into the same problems, because if you're arguing that a human brain is like an LLM and what you're really saying is, "it's kind of like an AI in general" -- well, that doesn't really say anything about whether a human brain is specifically similar to a system like GPT. There are lots of different ways to build a neural network, LLMs are one strategy.
When you say this:
> The "here's your answer" intuition that I experience is the "like an LLM thing".
What this sentence is actually saying is that you experience conclusions and knowledge where you don't know the source or where the process of your brain generating a decision or information happens unconsciously. The sentence is just saying that there are unconscious parts of your brain where you are not an active participant in the thinking process and it feels like the a spontaneous transformation from input to output.
But does that really strike you as a strong indicator that your brain is like an LLM, or does that sound more like a general description of almost any black-box system where the mechanisms are hidden and you can only examine the inputs and outputs? To say that there are parts of our thinking process that we're not able to consciously observe or examine is really just saying that parts of our thinking process resemble a black-box oracle. It's not making a strong claim about GPT.
But that's not what an LLM is, an LLM is a large language model. The distinguishing characteristics between an LLM and other AIs are not whether or not an LLM has a feedback loop. Lots of AI categories (arguably the majority of AIs) generate output from a single set of inputs as a single "operation" (ignoring the fact that most neural networks including LLMs internally have multiple layers and do in fact have multiple transformation steps, but whatever, we can treat that as one step for the purposes of conversation).
If anything, LLMs are the exception here in that they very often do have a feedback loop during normal usage; they are most commonly used in a conversational context where they generate the next "chunk" of a conversation after being fed back their previous answers alongside the followup responses of the person working with them. Arguably the feedback loop of an LLM is that as a conversation progresses, its future output is based on its previous output, which very literally becomes its new input after being marked up and extended by a human being. That is notably more feedback than many other AIs get.
Doubly so if you're trying to argue that an LLM can encompass a multi-modal setup, because suddenly refinement and feedback between multiple models is a core part of the final product.
So I'm not sure I agree that an LLM fits into that category in the first place, but even assuming it does, if your definition of an LLM is is just that it takes an input and turns it into an output as a single step without help... that is just such a broad category, I would guess the majority of AIs fall into that category. It's not what makes LLMs special; most predictive neural networks take a single set of inputs and generate a single set of outputs without intermediary human input. What makes LLMs interesting as a category is the training and structure and quirks of how they work, what makes them interesting is the differences between LLMs and other AI techniques.
To jump back to the Markov chain again, Markov chains do not have a feedback loop or refinement step: they take an input and map it to an output as a single step. Are Markov chains LLMs? Is any data transformer that operates in a single step an LLM? That is a really broad definition to use.
----
And we run into the same problems, because if you're arguing that a human brain is like an LLM and what you're really saying is, "it's kind of like an AI in general" -- well, that doesn't really say anything about whether a human brain is specifically similar to a system like GPT. There are lots of different ways to build a neural network, LLMs are one strategy.
When you say this:
> The "here's your answer" intuition that I experience is the "like an LLM thing".
What this sentence is actually saying is that you experience conclusions and knowledge where you don't know the source or where the process of your brain generating a decision or information happens unconsciously. The sentence is just saying that there are unconscious parts of your brain where you are not an active participant in the thinking process and it feels like the a spontaneous transformation from input to output.
But does that really strike you as a strong indicator that your brain is like an LLM, or does that sound more like a general description of almost any black-box system where the mechanisms are hidden and you can only examine the inputs and outputs? To say that there are parts of our thinking process that we're not able to consciously observe or examine is really just saying that parts of our thinking process resemble a black-box oracle. It's not making a strong claim about GPT.