Why AI Is Not Software
AI is not software.
Yes, today AI is written using software, but today’s implementation is actually a computational graph, that computes the behavior of networks of virtual neurons. The software is a mere simulation of that network.
Let me explain.
As I said, a neural network AI is actually a set of interconnected virtual “neurons”. Each neuron more or less models the behavior of human neurons. The figure illustrates a typical AI “neuron”, mathematically.
If you have studied math, you can probably make sense of the diagram. But all you need to know is that a “neuron” like the one in the figure is the fundamental building block of an AI system, just as neurons are building blocks in the brain.
The brain is not exactly like an AI system. Brains have glial cells and chemicals that bathe the neurons in complex ways. AI systems try to model the elements of brains that are thought to be essential for cognition — the neurons and their intereconnections. The AI field is still learning in this regard.
Nevertheless, AI systems built this way have achieved remarkable success in various kinds and levels of cognition. Witness the growth in AI capabilities shown in the figure below.
Real AI systems have lots of these virtual neurons interconnected. Here is a graph of a system that I built eight years ago:
When I say that I built this, I mean that I built a simulation of it, using a mathematics tool called TensorFlow. TensorFlow enables you to define what is called a computational graph. That is how the neuronal behavior is computed — it is simulated. The figure below shows the computational graph that I defined using TensorFlow.
It’s Matrix Math
This computational graph calculates matrix operations that model the activation functions and inputs of arrays of the virtual “neurons” in my neural network.
The neural networks used by things like ChatGPT are very large. ChatGPT 3.5 had billions of simulated neurons, with something like 100 billion interconnections between them. It turned out that when they increased that to 500 billion interconnections for version 4, the system became exponentially smarter — there was emergent intelligence. Still not smart like a human, but smart in its own way: smarter than us in some ways, and dumber than us in other ways.
Still, ChatGPT was a very limited architecture. It could not continue to learn once trained, and it used a single architecture — a so-called “transformer” pattern. In the past few years much more powerful and flexible designs have been developed.
All these systems are still simulations of neural networks. They are simulations in computers. They are not real neural networks, any more than a computer game is the real world. Yet, those simulations give us right — and sometimes wrong — answers, just as a real neural network would. Just as the brain does.
Also, these simulations are very power-hungry. Since they are simulations, they are very inefficient. Computers use much more power than real neurons do.
Training
Another kind of software used for neural networks is software to train the networks. There are lots of different kinds of neural networks and they are each trained in a different way. The figure below shows an algorithm for training a type of network called a “restricted Boltzman machine” network. This comes from a textbook that I read when I was studying this during the mid-2010s. Yes, there is a lot of math involved — linear algebra (matrix math) and advanced statistics.
Networks like ChatGPT are trained differently. They use what is called reinforcement learning. Still, there are computer algorithms that are used to train them on massive amounts of data. In this process, the AI is often made to re-read the same information up to four times — like a person studying for a test.
The training process is kind of like what you go through between birth and the age of 18, except greatly accelerated. Once fully trained, you are ready to take on the world — sort of.
AI Is Not What It Was in 2022
What a lot of people don’t realize is that the AI systems that they personally experimented with two years ago are now very obsolete. Things have moved on.
The figure below shows the architecture of the recently announced DeepSeek V3. DeepSeek surprised the world because it was trained for such low cost — two orders of magnitude less cost than the models of OpenAI. This is because the architecture of DeepSeek learns more effectively. It is smarter.
Also, if you used ChatGPT or a similar system, and concluded that it is not good at reasoning, you are right. But in only two years models have been developed that are good at reasoning — just like people are.
Toward the end of 2023, when a reasoning model had been developed at OpenAI, the scientists became alarmed and met with OpenAI’s board, warning them that the new model indicated that they were on the path to developing a powerful artificial intelligence that they said could threaten humanity. The OpenAI model was called Q-star. That model has since been renamed “Strawberry”, but now is marketed as “o1”.
So if you used ChatGPT last year, but have not used any of the new reasoning models, throw out your conclusions about what AI can do, and try the reasoning models. And try them again a year from now, because those will obsolete this year’s.
What If We Used Real Neurons?
An interesting thought is, if neural networks try to be like brains, what if we use real neurons?
Some companies are actually trying that. One is Cortical Labs in Australia. The figures below are from their labs.
They use networks of real, living human neurons, with electrodes to activate them and receive responses from. They claim that such biological neural networks outperform computer-based neural networks. They also use a lot less power — in fact, they use chemical energy rather than electricity.
Enter Neuromorphic Chips
But people also are creating real neural networks that are not biological. They are called neuromorphic systems. And guess what? They are not only faster, but they use 10,000 times less electricity, and are thousands of times smaller.
Again, today’s AI runs in data centers only because we are simulating neural networks. If we build real neural networks, we won’t need data centers for AI.
And since neuromorphic systems are not simulations, they don’t run any software.
Let me repeat that: they don’t run software. They are in a sense, non-biological brains.
Why We Don’t See These In Widespread Use
Right now there is some mismatch between the architectures of the latest AI systems and the architectures of the available neuromorphic chips. However, over time that mismatch will be corrected, because the incentive is so huge: imagine powerful AI systems that use little power — making them mobile, free from data centers — mobile like people.
The Digital Era Is Over
The digital era is over with. It was yesterday. AI is not software. It is confined to software today like a racehorse is confined to a barn. But AI will soon be free to be itself.
AI is not the next evolution of computers: AI is a move away from computers, to something that is much more brain-like.
The AI Era has begun.
