AI Began With a Question About the Human Brain
Where did today's artificial intelligence really begin?
Not with ChatGPT.
Not with Google.
And not even with Geoffrey Hinton.
The deeper story begins with scientists trying to understand the human brain.
In the late nineteenth and early twentieth centuries, neuroscientists such as Santiago Ramón y Cajal established that the nervous system was composed of individual cells called neurons.
Those neurons received signals, processed them and communicated with other neurons through enormous networks of connections.
This biological discovery eventually inspired one of the most important ideas in computing:
Could a simplified version of a neuron be represented mathematically?
From Biological Neurons to Mathematical Neurons
In 1943, Warren McCulloch and Walter Pitts proposed one of the first mathematical models of an artificial neuron.
Their model was primitive compared with today's neural networks, but the principle was revolutionary.
A computational unit could receive several inputs, apply different weights to them and generate an output.
The biological brain had begun its migration into mathematics.
In 1958, Frank Rosenblatt introduced the perceptron.
Unlike conventional programs, the perceptron could change its internal weights after seeing examples.
The machine did not need every answer to be explicitly programmed.
It could learn.
That single idea remains at the center of modern AI.
The Long Road to Deep Learning
The early promise of neural networks was followed by decades of difficulty.
Computers were slow.
Memory was expensive.
Datasets were small.
And training multilayer neural networks was extremely difficult.
But researchers continued.
During the 1980s, John Hopfield, Geoffrey Hinton and other scientists developed new approaches to neural computation.
A particularly important development was backpropagation, popularized in multilayer neural networks by David Rumelhart, Geoffrey Hinton and Ronald Williams.
Backpropagation allowed a network to calculate how wrong its answer was and send that error backward through its layers.
The network could then adjust millions of internal weights to improve the next answer.
This became one of the foundations of deep learning.
Geoffrey Hinton therefore did not invent artificial intelligence.
His importance is that he helped preserve and develop neural-network research during periods when many researchers had lost interest in it.
When large datasets, powerful GPUs and faster computers finally became available, those decades of research suddenly became extraordinarily valuable.
A Century of AI Development
| Era | Breakthrough | What Changed |
|---|---|---|
| 1890s–1930s | Neuroscience | Neurons and neural connections are identified |
| 1943 | Artificial neuron | Biological neurons become mathematical models |
| 1958 | Perceptron | Machines begin adjusting weights from examples |
| 1980s | Backpropagation | Multilayer neural networks become trainable |
| 2000s | Deep learning | Larger networks learn complex representations |
| 2012 | GPU acceleration | Massive parallel computing transforms AI training |
| 2017 | Transformer | Attention enables highly scalable language models |
| 2020s | Foundation models | AI learns language, images, code and multiple modalities |
| 2020s | Reasoning and agents | More computation moves from training into inference |
| Next | AGI | General-purpose machine intelligence may emerge |
2012: When GPUs Changed AI
Another major turning point came in 2012.
Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton demonstrated AlexNet, a deep neural network trained with GPUs for image recognition.
The improvement was dramatic.
This showed that neural-network ideas developed over previous decades could become extraordinarily powerful when combined with enough data and computing power.
AI development was no longer constrained primarily by theory.
Compute infrastructure itself became part of intelligence.
This is one reason today's AI revolution is directly connected with GPUs, high-bandwidth memory, advanced semiconductor packaging, AI data centers, power infrastructure and liquid cooling.
Then Came the Transformer
In 2017, Google researchers published the landmark paper "Attention Is All You Need."
Its authors included Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan Gomez, Łukasz Kaiser and Illia Polosukhin.
They introduced the Transformer architecture.
The Transformer did not replace neural networks.
It reorganized them around a powerful mechanism called attention.
Instead of processing information only in sequence, the model could determine which pieces of information were most relevant to one another.
Just as importantly, Transformer training could be massively parallelized.
That made it possible to train increasingly large models on enormous amounts of text, code, images and other data.
The path to today's large language models had opened.
The evolution can therefore be summarized as:
Brain Science → Artificial Neuron → Neural Network → Backpropagation → Deep Learning → GPU Computing → Transformer → Foundation Models → Reasoning → AGI
Training Is No Longer the Whole Story
For most of the deep-learning era, AI progress was associated primarily with training.
More data.
More GPUs.
More parameters.
More training time.
But another major shift is now underway.
Modern AI systems increasingly use additional computation during inference.
Instead of immediately producing an answer, an AI system can spend more computational effort analyzing a problem, comparing possibilities, using tools, checking results and refining its response.
This means future intelligence may depend on two enormous computational processes.
The first is learning before deployment.
The second is thinking after deployment.
Training creates the model.
Inference increasingly determines how deeply it can reason.
That change could become one of the most important steps toward more general intelligence.
When Will AGI Arrive?
No one knows.
There is no universally accepted definition of Artificial General Intelligence, and there may never be a single day when the world officially announces that AGI has arrived.
The transition could happen gradually.
AI may exceed average human capability in coding first.
Then mathematics.
Then engineering.
Then scientific research.
Then business analysis.
Then robotics and long-term autonomous operation.
Eventually enough capabilities may converge that society realizes something fundamental has changed.
DATAAD View
The late 2020s through the early-to-mid 2030s may become the most important period to watch for practical forms of AGI.
This should not be interpreted as a precise prediction.
Important limitations remain in reliability, memory, physical-world understanding, energy consumption, computing infrastructure and autonomous operation.
But AGI is increasingly becoming an engineering question rather than simply a philosophical idea.
And AGI may not even be the most important destination.
What comes afterward could be much larger.
Human Imagination Meets AI Knowledge
Humans have one extraordinary capability.
We imagine things that do not yet exist.
A novelist can imagine a civilization thousands of years into the future.
An engineer can visualize a machine before the first component has been manufactured.
A scientist can imagine a theory before an experiment proves it.
A child can imagine cities on Mars.
AI has a different advantage.
It can potentially connect enormous amounts of mathematics, physics, chemistry, biology, engineering, medicine, history and accumulated human knowledge.
No individual human can possess all of this knowledge.
But perhaps the future human will not need to.
The new equation may be:
Human Imagination × AI Knowledge × Machine Execution
A human proposes an idea.
AI analyzes millions of possibilities.
Simulation tests them.
Robotic systems manufacture prototypes.
Sensors collect the results.
AI analyzes the results again.
The human changes the idea.
The process repeats.
The distance between imagination and reality begins to shrink.
From AI Data Centers to Mars
This combination may eventually extend beyond Earth.
Human beings require oxygen, food, water, temperature control and protection from radiation.
Machines have very different limitations.
AI-controlled robots could explore dangerous environments for years.
They could construct solar farms on the Moon.
They could mine materials from asteroids.
They could build infrastructure on Mars before large numbers of humans arrive.
Autonomous factories could manufacture structures using local materials.
AI systems could operate scientific instruments millions of kilometers from Earth.
The first intelligence to permanently expand throughout the solar system may therefore not be purely biological and may not be purely artificial.
It could be human imagination extended through AI and machines.
The Bigger Meaning of Artificial Intelligence
The history of AI is therefore not simply the history of computers becoming smarter.
It is the history of humanity trying to understand its own intelligence.
Neuroscientists discovered neurons.
Mathematicians turned neurons into computational models.
Computer scientists created learning networks.
Deep learning allowed machines to discover complex representations.
GPUs allowed those networks to scale.
Transformers allowed knowledge to scale.
Reasoning models are beginning to make inference itself computationally intensive.
AGI may be the next milestone.
But the real transformation may come when billions of human imaginations gain access to intelligence and knowledge far beyond the capacity of any single human brain.
For thousands of years, humans looked at Mars and other worlds and imagined reaching them.
AI may not replace that imagination.
It may finally give humanity the tools to turn it into reality.