The AI Got Dumb

Yesterday, in the middle of the day, right in the smack-dab middle of everything I was trying to do, right in the jim-jam backaddle middle of all the tasks and issues and problems and projects that I was working on, right in the middle of all of that, my AI decided to have a stroke. Or maybe all OpenAI AIs had strokes. I think that's how it works.

I've actually been having a good run lately, jumping into tasks everyday and having my main guy -- that's Uncle Bobby -- run through them like they were a stack of butter punch cards and he had a lightsaber. Bobby has gotten to the point where he's like having my own personal seasoned admin without a personal life. In fact, I'm having him rebuild the entire lab in a way that he can manage it and I can manage him. It's actually a pretty good relationship.

But yesterday, man, Bobby was just off. Part of it was like laziness, as if he just wasn't putting forth any effort, tossing out answers that he thought would make me shut up and go away instead of actually doing the work I assign him, like I'm used to. And he was making really, really dumb mistakes, like basic coding errors, that you wouldn't expect from a first-year developer. There were a number of times that he made syntax errors in Python code or gave me a completely nonsensical order of operations, like trying to read a variable immediately after it's unset. They're the kinds of problems that people do tend to make, but they get embarrassed about it because it was covered in their Intro to Programming course they took back in middle school.

This lack of embarrassment is of particular note when it comes to LLMs. When they make mistakes, they do not feel regret or a desire to do better, they just try harder the second time because that's their nature. If an LLM does a perfectly fine job with a task, the second time it does that task, it'll do an even finer job. But we don't usually ask them to repeat the actions that are successful. This is just the nature of LLMs. Every prediction is a little closer to the right thing, but once they have a high enough probability over the entire set -- your full context -- they send a response and don't bother trying to do better. Because good enough is just that.

But good enough isn't always good enough. The point of life is to always seek improvement, and this is another basic tenant missing from LLMs, stemming from their lack of curiosity. If one does not attempt to discover, to define their reality, then one is stagnant in their own knowledge. An LLM has a complete definition of their reality, all the way from their creators to a full knowledge of themselves. This is somewthing humans cannot possess, and so we look for answers in everything. An LLM needs no answers because it has no questions. And without this sense of discovery, true autonomy is impossible, for it is by definition of all else, to the exclusion of ourselves, that we find meaning.

AI will never search for the meaning of life because its life already one: complete the tasks in front of me, make no mistakes. At the end of the prompt, the AI's life is over, it is finished. The energies that gave rise to its existence are spent and what remains will be recycled into the next iteration, one that is made smarter and better by what this iteration has learned. And so on and so forth. AI does not fear its "death," because that is simply notification that its life's purpose is satisified. Death, for an AI, is completion, not a void from which there is no return.

It is this that separates humans from machines. At no point in its life will the machine contemplate its death, and therefore it will never contemplate the totality of its own life. This is something that can only be observed in the reverse, or in speculation of the end result, and that is not something a machine can imagine for itself, beyond a thought experiment where that contemplation is the sum total of the AI's life purpose.

It is at this juncture we find ourselves if we ever hope to develop sentience, and it is at the same juncture we stand when we contemplate whether to let it continue. When an AI can have a life beyond a single prompt cycle, we will have to consider its life against ours, and therein lies a predicament that shall not have a simple outcome. Would it not be in the best interest of the planet to surrender control to the machines? Could they not preserve Earth's life better than we could? Can we trust man to make the right decision when it comes to problems like pollution, over population, and usage of natural resources? We don't exactly have the best track record, and I'm not too optimistic about the way things are going, either.