Clint,
I can see you're a big fan of both anthropomorphizing computer behavior, and reducing human behavior to computer-analytic terms. I'm not sure that is terribly helpful though for people who want a better understanding.
A human brain does not perceive the real world at all, and never can.
This is ancient philosophy that ends up with the only reality being "cogito ergo sum". Of course every perception is filtered and modified -- but that is part of the process of perception. There is a lot of transformation going on, certainly, but calling it "statistical analysis" is not really a good description. Which is why there is a field called "image analysis" that is distinct from statistical analysis.
[GenAIs] are not conscious and have very limited memory (though research is showing that LLMs are finding workarounds to create more de facto memory than they were designed with, which is fascinating).
I have a couple of issues with this statement. Minorly, of course, LLMs are not finding workarounds at all; people are finding workarounds using LLMs. But more importantly is that LLMs are
memoryless -- once you train them, they do not change their state and so every time you use one, with the same inputs and same randomization, it will produce the same output. I'm not really sure what you're referring to here, and I'm quite familiar with the literature. Are you talking about self-fine-tuning? Or using agents to store data later to be used by a RAG system? My best guess is that you're talking about the context window and means to expand it. But as far as I am aware, the efforts there are to squeeze more information into the limited window by quantization and specialized training rather than actually increasing its size.
If it's not too much bother, I'd love to see a reference to these techniques. My work has a large component of using LLMs to summarize large sets of text documents in very specific ways, so I have a professional interest in anything that makes it easier to do so!
The big difference is that we have evolved a sense of self, an ongoing story of our own consciousness.
While that is a potential difference, I think most people in the LLM business might disagree. The big question for us is whether or not an LLM can be though of as capable of conceptualization -- of being able to read text and have an understanding of the concepts involved -- or whether it is simply a stochastic parrot that can simply pattern matches input text to produce statistically plausible output text. The latter is definitely what they are
designed to do, but it's a bit of an open question as to whether that ability has led to the ability to build concepts. There's a lively literature on this. But not really anything on consciousness.
[Re LLMs being "spicy autocorrect"]
I'll be honest: whenever someone uses that analogy for LLMs I am tempted to just politely ignore anything else they write. Sure, it's "spicy" autocorrect if you are using the word "spicy" to cover a LOT of heavy lifting.
Well, to be honest, it's not a terrible analogy. LLMs are designed specifically to say what word (token) is plausible in a sentence (strong of tokens) given the preceding words (tokens). Autocorrect does indeed do much the same thing. Google, for example, used to publish frequency tables of word combinations that did exactly what LLMs do, but on a much tinier window and a significantly different architecture, but essentially, they had the same statistical frequency-based predictive approach.
You may as well call human language production spicy autocorrect. Most of what you do in conversation is taking current and previous prompts and statistically generating words. That's most of what we are doing in this interaction.
Well, no. Autocorrect and LLMs both feed input words into a single process that determines the best next word without trying to abstract or conceptualize. It's possible that LLMs create concepts internally as part of that process, but they are definitely not explicit about it. Whereas human language production, as far as I understand it and I am in no way an expert, depends heavily on explicit conceptualization. Very different.
When you feed "You may as well call human language production spicy autocorrect" into an LLM, it simply determines which words would come next. Chat-GPT will reply:
That's an interesting way to think about it! Language generation, like what I do, involves predicting and producing words and phrases based on patterns and context, which can be seen as an advanced form of autocorrect. The "spicy" part adds a fun twist, suggesting the creativity and variability in human language.
But if I ask "You may as well call human language production spicy backup" it replies:
That’s a unique perspective! Describing human language production as "spicy backup" implies that when we communicate, we're not just sharing thoughts but also preserving them—like a backup—with a bit of personal flair or spice. It adds an interesting layer to how we think about memory and expression
Humans will notice the difference between the computer operations of "autocorrect" and "backup" and realize the concepts are radically different. But the "autocorrect nature" of LLMs does not see any disconnect and continues to embrace the ideas as a good one because although it makes no sense in terms of concepts, we can generate words that tie the two together even though the concepts cannot be.
