All the AI systems in vogue today are built around neural networks, so that’ll be the focus in this discussion. A neural network is just a specific kind of math function, like f(x) = x² or f(x) = sin(x), except much bigger and more complicated. Rather than writing down an explicit formula, we describe the mathematical building blocks it is made from and how these pieces fit together, like how a Lego set comes with an instruction booklet telling you how to assemble the Lego blocks.

The figures in AI research papers (like the famous transformers one above) are like the instruction manuals in Lego boxes explaining how the different bricks fit together.

Most of the mathematical pieces in neural networks involve coefficients, like how the quadratic function f(x) = ax² + bx + c has three coefficients a, b, c. Choosing different values for the three coefficients in this quadratic gives us parabolas of different shapes and locations. Choosing a specific parabola is equivalent to specifying these three numbers a,b,c, so we can think of that as choosing a point (a,b,c) in three-dimensional space. Moving that point around in space corresponds to sliding and bending the parabola around.

Choosing a parabola f(x) = ax² + bx + c is choosing three numbers (a,b,c), so a point in three-dimensional space—like specifying latitude, longitude, altitude.

If we think of a neural network in terms of Legos, we can imagine that each Lego has a shape but also that we are choosing a color for each bump on the Lego.

Human designers choose how many Legos to use, what shape each is, and how they fit together—this is the “architecture” of the neural net—but they don’t choose the colors on all these Lego bumps. Those start out randomly, then “training” the neural network is the process of steadily tweaking all the colors to make the Lego contraption perform better at the task(s) you’re interested in.

Modern systems like GPT5.5 are built out of so many large Legos that in total there are likely trillions of colors to choose. (The official name for these colors is “parameters” or “weights.”) So GPT5.5 is built by:

1. manually writing down the Lego instruction booklet (that’s what the AI designers do)

2. randomly choosing a point in the trillion-dimensional space representing the possible colors for all the Lego bumps (this is the neural network’s random initialization)

3. training the neural network (which can include static data like all the text on the internet, and dynamic data like human or AI evaluators rating the answers the AI gives to various prompt), which gently moves that point in trillion-dimensional space around toward some goal, such as being able to provide answers that users find more useful.

Wherefore art thou consciousness?

The architecture for a neural network surely has no consciousness—remember, it is just the Lego instruction booklet, and it doesn’t even specify the Lego colors. It is a blueprint, not a building, and an incomplete one at that. So, we must look beyond Step 1 in the above list.

I don’t think the neural network we construct in Step 2 could be considered conscious. It is just a bunch of random numbers that have seen no data, attempted no tasks, received no feedback. It is the static you saw on your old TV before you tuned to a channel.

So, if there is consciousness, it must come from Step 3. That means our trillion-dimensional space is carved into two regions, one representing non-consciousness and the other, presumably much smaller, representing consciousness. Then the belief would be that we start by randomly throwing a dart into this enormous void of non-consciousness, and if we find the right data and the right tasks then the training process guides our dart into the conscious realm. This doesn’t seem out of the realm of possibility.

In fact, consciousness is likely a continuum, not black-and-white, so if we could measure consciousness then we could directly use that as an objective function to guide our exploration of the trillion-dimensional space. Since we can’t yet measure consciousness, we can at least imagine that training on the right data in the right way serves as a rough proxy—so that by forcing our neural network to get better at doing the things we want, like chatting with us and answering our questions, we inadvertently move uphill in the mountain range of consciousness over our trillion-dimensional space.

This could well be similar to the biological world: if consciousness evolved in animals, it presumably did so as a byproduct of some survival mechanism.

But then we must remember what these neural networks really are: mathematical formulas. If we are lucky enough to train/evolve our way to a conscious neural network, are we saying that the math formula itself is conscious? Surely if we wrote this formula down by hand on a (very large) sheet of paper the way we write other math functions, nobody would claim it could be conscious.

Does specifying this math formula in computer code make it any more conscious? Nope. Whether we use math notation or computer code, we just have a bunch of symbols that describe a process for turning an input to an output. The difference is that with computer code we can turn the computer on, feed numbers into the function, and compute!

But… you could, in principle, do this by hand as well. Every step of every computation the computer makes when the neural network “thinks” and answers the prompts you feed it, is a math calculation that you could, in principle, write out by hand with pen and paper. I don’t think anyone believes that writing these big computations on (really big) paper could produce consciousness.

And here’s the crux. Fundamentally, it’s no different whether a computer does these computations digitally in binary or you do them manually in ink.

The punchline. A photograph has no movement, yet when we string a bunch of photographs together in rapid succession, we get a convincing illusion of movement. And if the photographs are sufficiently detailed and the frame-rate higher than our eye can perceive, then the movie appears as real as life itself. But we know it isn’t.

I believe this is where we stand with AI.

All the AI we currently have, and all the AI systems we envision down the road for the foreseeable future, are just a string of mathematical computations performed in rapid succession. They may appear to produce continuous thought, perhaps even consciousness, but I strongly believe this is as illusory as the movement and life we see in film.

A math formula is not conscious. Using that math formula to turn an input number into an output number is a recipe, not consciousness. And performing many such computations in a row is a mathematical flipbook, not digital consciousness.