Post cover image
Image generated by Midjourney using my own prompt

October 20, 2026

From 8 neurons to Terminator: an AI’s story

Humanity has been playing with thinking machines for at least 200 years. Now, perhaps, this toy is slipping from our grasp

By 'Wine' Roland Mucciarelli

5 min read

Models: the Right One?

When chatGPT became available on our devices in November 2022, we all thought it was just a fun game, just a different way to search the web. And instead, it was a revolution.

After the initial shock, we realized the power of LLMs, Large Language Models, and many of us began to think, with a bit of fear, about the possibility that thinking machines had become a reality. Could Skynet, the Artificial Intelligence that governed Terminator, become our future? Or perhaps the Matrix would be our destiny? Would we need an android hunter to fight the replicants? Had we built a Frankenstein's monster, and not one as cute as Mel Brooks's?

Probably, if humanity's future is truly so apocalyptic, it's too late to stop it now. The speed with which various AIs like Gemini, Claude, and GPT have expanded presents us with profound questions: for or against, there seems to be no middle ground.

AI had appeared in our lives, and we were both mesmerized and terrified.

And, above all, we thought it was something radically new. And here, too, most of us were wrong. The New York Times was publishing headlines about AI seventy years ago. But the history of thinking machines began much earlier.

The Century of Wooden Brains

Electricity hadn't yet been discovered, but humans were already obsessed with mechanized thinking. In 1623, the German scientist Wilhelm Schickard built the first mechanical calculator capable of performing six-digit operations. Blaise Pascal improved it by introducing a mechanism for addition with carry, and Gottfried Wilhelm Leibniz later mastered multiplication. But the true visionary was Charles Babbage, who between 1834 and 1837 began designing his Analytical Engine, a steam-powered machine with the architecture of modern computers, featuring a central processing unit and external devices to manage number input and calculation responses. Babbage never completed his project, but his legacy passed to Ada Lovelace, daughter of the poet Lord Byron.

None

Ada Lovelace wrote an algorithm to calculate Bernoulli numbers. Without going into detail, the important thing is that this algorithm is recognized as the first computer program in human history. But the most important thing about Ada Lovelace's work is her belief that if a machine could manipulate numbers, and numbers can represent symbols, sounds, and ideas, then the machine could compose music, create graphics, and manipulate thought. To underscore the importance of Ada Lovelace, in the 1970s the US Department of Defense named one of its programming languages ​​Ada.

The "Blade Runner" Test

Another fundamental breakthrough came in 1937, when Claude Shannon demonstrated that electrical circuits and telephone switching networks could be used to perform Boolean algebra calculations (invented by George Boole and based on the binary code, i.e., the numbers 0 and 1). Thus was born information theory.

Meanwhile, a British mathematician named Alan Turing was asking an even bolder question. In his 1936 paper, "On Computable Numbers", Turing laid the foundation for modern computers, and then posed a fundamental question: how can we determine whether a machine is intelligent?

None

His answer was the Turing Test: if a human interviewer interacts with a machine and cannot reliably distinguish whether it is conversing with a person or a device, the machine must be considered intelligent.

If this sounds eerily similar to the Voight-Kampff test that Deckard administers to Rachael in the movie Blade Runner, you're not imagining it.

Then came 1943. Neurophysiologist Warren McCulloch and mathematician Walter Pitts demonstrated that the neurons in our brain could be modeled as simple binary on/off switches. Logic could be represented with neurons and a nervous system modeled with circuits. That discovery sparked a full-blown ideological battle among scientists of that technology.

Rules or Hippie Culture?

When the term "Artificial Intelligence" was officially coined by the father of AI John McCarthy at the historic Dartmouth Conference in 1956, researchers had already divided into two starkly opposing camps:

The first was called the symbolic (top-down) approach, which consisted of teaching the machine the rulebook. For example, if you want a computer to compose a symphony, you need to teach it the strict rules of harmony, counterpoint, and rhythm. To get it to solve mathematical problems, you need to provide axioms and theorems.

This approach dominated the early days of computer science. It was the logic that allowed IBM's Deep Blue to defeat world chess champion Garry Kasparov in 1996: immense computing power processed through rules created by hand.

The second was the evolutionary (bottom-up) approach, based on the idea of ​​letting the machine learn on its own. Instead of teaching the rules of melody, you have it listen to 100,000 songs, read the scores, and access online music databases. At that point, it will have found its own rules for creating music and will be able to compose pieces according to its own logic. Like students on a 1960s campus.

In 1957, the American psychologist Frank Rosenblatt took the second approach and proposed his idea to the U.S. Office of Naval Research and the Air Force Development Center. He called his creation the Perceptron.

The 8-Ton Monster

The Perceptron was the world's first physical neural network, an eight-ton behemoth composed of wires, relays, potentiometers, and noisy motors that turned gears. It could simulate a neural network of eight neurons.

Yet, when Rosenblatt presented it in 1960, the media hype was deafening. The New York Times and other newspapers ran sensational headlines, claiming that the military had invented the embryo of an electronic brain that would soon be able to walk, talk, see, reproduce, and become aware of its own existence.

As has been the case since 2022, humanity then also felt real panic and excitement at the highest levels.

The reality, however, was much more banal, almost disastrous: the Perceptron could barely distinguish an object placed to its left from one placed to its right.

Researchers realized that it was impossible to build a thinking brain with the computing power of the 1960s, and funding collapsed. AI therefore took the more reassuring, less chaotic path of the symbolic, rule-based approach.

The right approach (maybe)

If we look at current AI systems like Gemini, chatGPT, and Claude, however, we realize that they are not based on dictionaries or grammatical rules. Modern AI is based on the second approach: you provide a huge amount of documents and data and let the system learn on its own. Compared to the 1960s, we have the enormous mass of information that is the Web. The AI ​​we know today is Frank Rosenblatt's dream, born from the bottom up.

This is the meaning of the "P" in GPT: Pre-training.

Instead of teaching systems the rules of human thought, we provided books, conversations, images, and videos, giving AI access to the web. This information fed into neural networks infinitely more complex than the first Perceptron network, but based on those intuitions. We scaled the architecture from Rosenblatt's eight neurons to hundreds of billions of parameters. From analyzing these texts, the machine developed relationships between words, images, sounds, formulas, and videos based on statistics.

The AI ​​learned the difference between a dog and a cat by observing millions of photographs and drawings; it learned to write well and not be rude.

Of course, it all depends on the data we initially provided.

If we take an LLM and download it to our computer, and provide the AI ​​only with religious texts, we might have built an Artificial Preacher. And if we provide the AI ​​only with texts that speak of violence, we might have built the Artificial Dictator.

And perhaps that's exactly what we've done.

Sources