This post introduces the Rambling tag into the blog tags. I'll be using it when I'm yapping about something. In essence, it's the inverse of the technical tags, or a lack of technical tags.
Today I ran into a post on Tildes about AI discourse, and how looking past the polarized discourse on the subject actually can reveal beauty in how ideas are embedded within language. I thought this was a beautiful way to look at it, and inspired some of my own introspection on the topic, and effectively also inspired this post.
Detractors of AI often like to boil down Large Language Models (LLMs) as being "unable to reason", or as "overhyped auto complete". I believe that the implications of LLM output run significantly deeper than that, however, I also wish to offer up what I believe is actually missing from AI output. Because in a sense, the detractors are correct... it's critically lacking something fundamental in the experience of being a consciousness.
But let's start with what AI is, not what it isn't.
This section will almost certainly be a rephrasing of the Tildes post above. I strongly recommend you check it out. Hopefully the link doesn't die any time soon.
LLMs do essentially one thing - it calculates a probable output from a context. The context window (a phrase literally used in the field) is a critical part of the LLMs view. It is a machine that can take some amount of text, as defined by the computing resources supplied to it, and predict what the most likely next set of words is. You can see why detractors use the aforementioned terms, it is essentially a word machine. However, I think this has more nuanced implications of what we are capable of extracting from predictions. In a way, we are using statistics to predict ideas.
Language, as a tool, has been honed over thousands of years to effectively encode ideas into something we can share. My words, your reponse to those words, even this very post, are a collection of words and phrases meant to contain ideas. We use words to record knowledge, wisdom, and everything in between. With this in mind, I would put forward that the output of LLMs isn't just text. LLMs are an invention that allow us to use statistics to predict ideas, using text as a vehicle. Phrased another way, in the same way we use statistics to predict social trends or scientific phenomenon, we are using it to predict the output of thought. LLMs allow us to use statistics to distill ideas out of words.
This means that LLMs exist in an interesting space - they are capable of reason. They reason all the time. In the same way a domino falls and knocks over the next one, an LLM goes from one word (token) to another, constructing ideas in the form of sentences. This is a form of reasoning, even if it's done mathematically using text. The human advancement that allowed this to exist is absolutely staggering. It's a combination of thousands of years of language development, mathematics, and computer processing. Despite our current world climate, it is without a doubt a beautiful achievement.
However, it is missing something. And I believe that thing has a lot to do with the experience of being human. It's something i'll boil down to...
Here, I'll use the term inspiration in a slightly specific way. I use it to describe something I believe LLMs are fundamentally incapable of experiencing. Inspiration, to me, is a phenomenon in the brain that introduces variance, and allows us to come up with unique ideas.
An LLM introduces variance into its output by rolling a random number, and using it to decide if it will vary from its statistical calculation. As humans, we do no such thing. In fact, the output of our thoughts is even more deterministic. It's the result of the lighting pathways within our minds. A mass of connected neurons firing together to form thoughts, ideas, memories. However, humans manage to achieve variant thoughts in a totally different way: experiences.
The core idea of inspiration, to me, is that human thought becomes unique when it's paired with the experiences of being a human. An LLM may reason in one line of thought, but a human's line of thinking is directly influenced by whatever else hits the brain at that time. It might be a stimuli (smell, sight, sound), it might be a memory (technically feasible with a big context window, but humans have context windows that AI could only dream of), it might even be literal damage (maybe you're concussed? idk). Regardless of what it is, the system of human thought is more than just context and linear reasoning. The chemistry in the brain is subject to the environment to such a degree that variance is inevitable. An LLM would not have an apple fall on its head like Newton, nor could it get an idea for a melody by walking down the street. Even writing this post, I am listening to relaxing music, which is without a doubt setting the tone for my writing.
To me, Inspiration is consciousness incarnate. It's whats missing in AI art, it's what's missing in AI writing, and it's what the greatest minds in history have ample of. To be more human, is to be more inspired.
I feel the need to write a closer here, though I think the last line of the previous section closes it out well. Who knows if AI advancement will get to the point where it manages to encode inspiration somehow. All I know is, there are humans that are capable of reason without any inner monologue at all. So until we find more ways to distill ideas out of math (like we have with statistics and language), I'm not confident that we'll ever hit something that replicates a human perfectly.