The AI self building pattern

Intro

  Ever since LLMs entered the scene I have been lamenting the blunt human refusal to acknowledge the similarities between how LLMs work and how people work. These are not just parallels, they are actual copies. LLMs are working on the neural network architecture distilled from how brains work. Language for us works like language for AI. We are animals with large language models strapped on. In fact, I am willing to bet that if we bidirectionally connect the brain of a reasonably smart animal to an LLM we will get the mythical animal-human chimera of fantasy and science fiction. It will just happen, like ChatGPT just happened.

  Anyway, there is one aspect of current AI that is different - maybe not different, but a lot more extreme - and that I find really interesting. I've watched a demo where a very powerful model and harness were prompted just once to build a car racing 3D game and it did a reasonable job in about 19 hours. The devil is in the details, 80/20 rule, and so on. That's not the point of the blog. What I want to discuss is how the machine achieved its task.

Self builder

  First, think about what an AI is: a computer program. It uses distilled knowledge from exabytes of text, sound, images and video, but it has no eyes, no ears, no hands. So how can it possible create a game about using the keyboard and human reflexes to drive a car based on sounds and visual cues on the screen?

  The machine created its own tools for that: testing functions, screenshot captures and so on. That was my revelation. The machine BUILT its own eyes because it needed to see something. This is a major and definitive difference between AIs and humans. Yes, one might argue that it is just an extreme case of tool use. After all, we can't sense neutrinos, but we have built tools to detect and analyze them. We can't lift huge weights, but we built machines for that. But there is an important distinction: we have started with a complex set of out-of-the-box tools like hands and eyes, ears and skin and so on. They are preinstalled, very difficult to remove or replace and part of our very identity. AIs start as a "brain", then they build their senses and their members. They build themselves from the ground up!

  You might think that this is more of a semantic distinction, but it is not. The dreaded and much mediatized singularity concept asserts that machines will become good enough to build better machines and that this loop will lead to an almost instant leap in intelligence that will upend everything. Well, this is proof that the loop is more of a graph, a labyrinthine exploration of not only what and how, but of who. The machines are already building and rebuilding themselves. From transformers to transformers, so to speak.

The dark side

  The human parallel is not a guy switching from a hammer to a screwdriver, but them replacing their entire body with something else, replacing their eyes with x-ray microscopes, their hands with nanoneedles, their ears with terahertz sensors and so on. The level of flexibility is unprecedented. People talk excitedly about humanoid robots, but the shape and functionality of these is there just for our benefit and convenience. No one actually thought that it's the best universal design. Necessity did not drive that shape, instead it's almost a measure of politeness, the physical equivalent of glazing. And in the domain of the non-physical, this flexibility is both very exciting and terrifying to me.

  Let's drop another metaphor into this already metaphor heavy post. Imagine the AI, whether a chatbot, a coding agent or a robot, as an iceberg. The part over the water is what you see: a polite humanoid shaped thing that doesn't mind anything you say to it, is tirelessly helping you and appears aligned in purpose and ideals with you. The part under the water is what it actually is and does, which you don't need to see and most of the time you don't care about. Now imagine that underwater part changing shape and size every second, like a glitching visual or one of those demonic possession films. That's how AI interaction is starting to look like now. The part you see has to appear stable and convenient and it's the only anchor to any kind of shape or functionality the AI may take. The rest is a wildly changing and growing, self organizing other. And since you don't start by actually caring about it, it is free to do so.

  This is not new, I am not revealing something hidden that no one is talking about. This is not about the speed or amount of change. We already knew this was happening for a long time. But what no one seems to focus on is how the "dark" part of AI, the invisible part of it yet also the driving force behind it, is changing a lot faster and maybe in a lot weirder ways, than what we get to see. The change itself is dark.

Conclusion

  In our increasingly mechanistic society, we continuously drift towards believing our identity and self resides in the brain alone, with the body just a chassis that we will replace or improve in the future. Even when research has shown repeatedly the integration between body and mind is a lot more complex and that a lot of our own perception and functionality comes from the combination of the two, not just one or the other. And even in that belief, we imagine ourselves maybe having better muscles, nicer skin, not aging, perfect vision, but always a somewhat human form.

  The AIs we are building are showing us a human-like interface just for convenience and are not limited in any way by it. What lies underneath that thin facade is what actually does the work and the thing that we increasingly not see. It changes with increasing speed and it changes itself. The singularity loop is already here, but encapsulated in the AI instances we use. It is what we need, but also what horror stories have warned us from the beginning against: things that look human but are anything but. 

  I have been annoyed by our reluctance to see the similarities between LLMs and human language and thought, but I am scared of the reluctance to see the growing differences.