How Large Thinking Models Evolve into the Same “Mind” As They Reflect Reality Increasingly Better
I am one of the most interesting discoveries (and topics) in artificial intelligence, leaving aside the debate about whether this is intelligence at all or not.
We (at least!) think that if you train only one AI model, say, on images and one on text only, they will develop completely different ways of “thinking” – let's not get into the debate about what this means. Our opinion would be that they use completely different architectures and process completely different data, so they should, in every sense, have completely different “brains”, even if they are both good at their jobs with images and text.
But according to some interesting research from various groups, that's not the case at all!
Already in 2024, MIT has presented strong evidence that every major AI model is secretly evolving towards a common “mindset”. (or mind, whatever you want to call it). As these models develop and become more robust, they all reach the same conclusion about how the world is built. Perhaps this was not evident in the first models, because they were wrong in thinking; but it becomes more obvious as they get better. And it is said, I would say, the reason for that is that if they are all correct they SHOULD create the exact same representation of reality.
Cave figure (AI).
To understand why this happened, some researchers looked back 2,400 years to Plato's “Allegory of the Cave” — which resulted in some interesting articles in print containing ideas such as the “Platonic Representation Hypothesis”. Basically, Plato believed that we humans are like prisoners in a cave, watching the flickering shadows on the wall. We think that shadows (our perceptions) are “real”, but they are actually just projections of a deeper, hidden, complex reality that exists outside the cave.
One of the many papers I read to prepare for this (links at the end) says that AI models do exactly the same thing, and in doing so they combine the same model of how the world works to understand input shadows.
The billions of lines of text, the billions of pixels in images, the endless audio files used to train our modern AI models are their perception (“shadows”) of our world. These powerful models look at these different aspects of human data and, completely independently, they are finding the same basic structure of the universe in order to make sense of it.
Different eyes, same vision
Here's the part that makes sense, to me at least: A model that only “sees” images and a model that only “reads” text measures the distance between concepts in the same way. (if both are good enough).
If you ask a visual model to write the “distance” between a picture of “dog” and a “wolf”, then you ask a language model to write the distance between the word “dog” and the word “wolf”, the mathematical structures they build are more and more similar as they are better able to distinguish between the two animals.
In other words, it seems that as these models grow and improve, they cease to be a chaos of random interactions. Research shows that they tend to align, and especially as conceptual models and language models grow, the way they represent data becomes more and more similar. It's amazing, don't think!
Why scale changes everything
According to the available research, all this seems to be all that happens and happens with today's models from all companies and trained by different sources, as long as the model itself knows enough. In fact, as the model gets bigger, whatever it is, it gets a “phase change” in their internal thinking. Research seems to show that these models stop remembering their specific functions and instead start building a realistic mathematical model of themselves.
And apparently, this happens because of “selective pressure” acting on the models:
- General work: If you want AI to be good at everything, there's only one “best” way to represent the world in a way that's not overly wasteful but predictable. Since there is only ONE best way, all must reach it!
- Power: Larger models have “room” to find a better, simpler solution. But having enough room in terms of the structure of the number of parameters must be balanced with avoiding excessive immersion.
- Simple bias: Deep networks actually prefer simple solutions to complex ones, and especially if over-encapsulation is avoided.
One important thing is that different AI models may adapt to these pieces of selective pressure at different speeds (or with different levels of efficiency); but surely they are all moving towards the same final state of higher understanding which is achieved through the same internal representation of how the world works.
The most modern study of “methods of knowledge”
If I were 25 years younger and had to choose a career now, I would probably choose something like computer science mixed with psychology. Because for me, this is the most exciting part of the AI world. Learn why!
A recent survey on “knowledge pathways” in LLMs adds another layer to all of the above. It suggests that the information in these models is not just randomly scattered; rather, it moves from simple memorization to more complex “understanding and application.” Then there is some sort of “power intelligence” at play. The trend that information and power tend to converge in the same representational areas seems to occur across the entire artificial neural model group, regardless of the data, method, or purpose. Even information that we people do not understand well yet (or only experts in a certain field grasp, say, how to create a composer's music or why and how photons can stick to a physicist) are drawn by these models as they find patterns (following for example, say in music or quantum physics) that our biological process can quickly.
Why this is cool, and the analogy with how we humans learn
This is one of those rare times when math, computer science, and philosophy collide. It seems to me that the models form a a unified view of reality only in the way they can consume it: as words, pictures, and sounds. It's not that different from the way children learn, maybe even in a different sequence and adding more input that includes reality (hard, bumps, etc.), and of course combined with physical effects (crying, laughing, moving limbs, walking, …)
Inside our brains, after all, multimodality is all integrated and operates under a common global understanding (which, yes, can be manipulated by manipulation, but that's for another day!). In other words, our brain remaps everything into a single reality, which is our interpretation of how the world works. If I take a picture of an apple, if I write the word “apple”, if I record the sound of someone biting into an apple, those are three different “shadows” but they all come from the same, physical reality of the apple. Even if the apple is one color from another, or artificially painted, or if I write the word in other languages, “manzana” or “pomme” will show the same kind of shadow when I speak Spanish and French.
Virtual representation of the world + physical input + physical output -> robotics
All of the above, including comparisons of children and people, can also be greatly expanded. Or at least I like to think so.
Add physical input and output capabilities to a truly global, efficient AI model, and we have a robot that can learn to interpret and interact with the world. It is hard-wired into the “nature” of survival, and well… who knows where this could go. But we are not very different from that construction, don't you think?
I'm going to leave this here, so I don't get into that debate, but please share with me about it!
Let's put it here with clear, research-based evidence that we are no longer just building tools to help us write emails, summarize something, write code, or edit or create images. We create digital mirrors for the universe. And we watch, in real time, as silicon and code independently discover the inner workings of the world we live in.
References
To write this post I went in detail with these two very interesting works: research presented as a preprint in 2024, and a recent update on information methods in LLMs:
Plato's Representation Hypothesis:
Knowledge Methods in Major Language Models: A Survey and Perspective:
