Everyone knows the game Civilization. It has a technology tree — you research writing, mathematics unlocks, then astronomy, then navigation. Clean and logical.
But the real technological tree is more complex. Behind every node of the visible tree hide three invisible trees. And without them the main tree simply cannot grow.
Let us go through this with concrete examples.
Tree 1 — The Main Tree of Technologies
This is what everyone sees. A new idea or scientific discovery gives birth to a new technology.
Einstein in 1917 described the theory of stimulated emission of light. In 1960 Theodore Maiman built the first working laser. In 1903 the Wright brothers understood how to control the lift of a wing. In 1938 Otto Hahn split the uranium atom. These are all nodes of the main tree — ideas that change the world.
Tree 2 — The Tree of Instruments
Every new technology requires a new instrument for its realisation. This is the first hidden tree — the direct manifestation of the main tree.
The laser required a ruby crystal of special purity, mirrors accurate to the nanometre, a flash lamp of strictly defined power. The Wright brothers' aircraft required a light petrol engine — a steam engine was too heavy. The atomic bomb required centrifuges for separating uranium isotopes — without them uranium enrichment is physically impossible. The transistor required monocrystalline silicon of extraordinary purity — 99.9999999%.
It would seem that is all. There is an idea, there is an instrument for its realisation. But this is where things get truly interesting.
Tree 3 — The Shadow Tree of Instruments
The new instrument from Tree 2 cannot be created with old instruments. To produce it, entirely new auxiliary instruments are needed — not for the final product, but specifically for the creation of the main instrument.
To build the ASML lithographic machine for chip production — special machines for polishing lenses to atomic-level precision were needed. These machines did not exist before the task arose. To build the engine for the Wright brothers' aircraft — Charles Taylor, their mechanic, constructed a special lathe for turning aluminium cylinders. The existing machines of that time did not provide the necessary precision. To build the centrifuges for uranium enrichment in the Manhattan Project — special balancing machines had to be built, because rotor imbalance at 50,000 revolutions per minute destroyed the centrifuge within seconds. To grow monocrystalline silicon for transistors — Gordon Teal at Bell Labs in 1950 built a special Czochralski apparatus with temperature control accurate to fractions of a degree. Nothing like it existed in industry at that time.
Tree 4 — The Shadow Tree of Technologies
But that is still not the bottom. To create the auxiliary instruments from Tree 3 — new technologies are needed specifically for their production. This is the fourth tree, the most invisible of all.
To polish the lenses for ASML — a new technology of ion-beam etching of glass was required, which simply did not exist before. It was developed by the company Zeiss specifically for this task. Zeiss has existed since 1846 and today is the sole manufacturer of these lenses in the world. To balance the rotors of centrifuges in the Manhattan Project — a new technology of dynamic balancing of rotating bodies at ultra-high speeds was required. Before this, balancing had only been applied to steam turbines at speeds hundreds of times lower. To control the temperature in the monocrystalline silicon growing apparatus — new thermocouple and temperature regulator technology was required, with precision unachievable by industrial equipment of that era. To build Taylor's lathe for aluminium cylinders — a new technology of aluminium cutting was required. Aluminium was then a new metal and would "smear" ordinary cutting tools — special geometry and cutting speeds were needed.
From Electronics to AI — The Same Four Trees
Let us apply the same system to artificial intelligence — one of the most complex technological chains in the history of humanity.
Tree 1 — The Main Tree of AI Technologies
In 1943 mathematician Warren McCulloch and neurophysiologist Walter Pitts described a mathematical model of a neuron — a binary element that either fires or does not. This was pure theory, with no practical application.
In 1957 Frank Rosenblatt created the perceptron — the first learning neural network. But it could only solve linearly separable problems and hit a dead end.
In 1974 Paul Werbos described the backpropagation algorithm. The idea allowed a network to learn from its errors, adjusting weights at each layer. In 1986 Geoffrey Hinton, Rumelhart and Williams republished it and made it practical. This thawed AI research after years of "winter".
In 1997 LSTMs appeared — Long Short-Term Memory networks that could work with sequences and retain context. But they processed text strictly one word at a time — parallel processing was impossible.
In 2017 researchers at Google published the paper "Attention Is All You Need". The transformer architecture made it possible to process all text simultaneously rather than sequentially — each token is compared against all others through the attention mechanism. From this grew GPT, BERT, Claude and all of modern AI.
Tree 2 — The Tree of AI Instruments
The transformer architecture requires processing enormous matrices of numbers simultaneously — billions of matrix multiplication operations per second. A CPU is physically incapable of this — it is sequential. What was needed was a GPU — a graphics processing unit, architecturally designed for parallel computation.
Training modern large language models requires not just GPUs, but thousands of GPUs working simultaneously. By 2024 Meta had 600,000 H100 GPUs dedicated to AI research and development. Elon Musk's startup xAI built the Colossus supercomputer with 200,000 NVIDIA H100/H200 GPUs.
Beyond GPUs, special chips with high-speed HBM — High Bandwidth Memory — were needed, which ordinary GPUs did not have. And specialised server platforms with hundreds of gigabytes of RAM and high-speed inter-processor NVLink connections.
Tree 3 — The Shadow Tree of AI Instruments
GPUs for AI contain billions of transistors on an area the size of a fingernail. To manufacture them, lithographic machines with extreme ultraviolet radiation are needed — EUV lithography machines from ASML. These are machines the height of a double-decker bus, costing 150 to 200 million dollars each. Before the 2010s such machines simply did not exist — they had to be created specifically for this purpose.
To manufacture GPU chips, ISO Class 1 cleanrooms are required — spaces in which no more than 10 particles of size 0.1 micron are present per cubic metre of air. This is millions of times cleaner than ordinary air. Creating such spaces required new filtration systems, new construction materials, new working protocols.
To connect thousands of GPUs together into clusters for AI training, specialised high-speed InfiniBand and NVLink switches were needed — equipment that before the AI era existed only in narrowly specialised supercomputers and was not commercially available at the required scale.
Tree 4 — The Shadow Tree of AI Technologies
To build the ASML EUV lithography machine, a new technology for generating extreme ultraviolet light was required — a laser beam strikes a droplet of tin in a vacuum 50,000 times per second, creating a plasma at the temperature of the surface of the Sun, which emits EUV light. This technology was developed from scratch over more than 20 years through the joint efforts of ASML, Zeiss and dozens of other companies.
To create HBM memory for GPUs, a new technology of three-dimensional chip packaging was needed — TSV, Through-Silicon Vias. This is a method of vertically connecting silicon layers through microscopic holes. Before the AI era, such packaging density was needed by no one.
To train transformers on thousands of GPUs simultaneously, a new GPU programming technology was needed — CUDA. It was released by NVIDIA in 2007 and became central to NVIDIA's strategy of positioning the GPU as a universal tool for scientific applications. By 2015 CUDA development was increasingly focused on accelerating machine learning workloads. NVIDIA developed specialised libraries — cuDNN for deep learning and cuBLAS for linear algebra. CUDA is not simply a library. It is an entirely new technology for programming parallel computation, which did not exist before it.
The Conclusion — The Correct Order of Growth:
A step forward in the main tree of technologies actually looks like this:
First a new idea or technology emerges. To realise it, a new main instrument is needed. But that instrument cannot be created with old instruments — new auxiliary instruments are needed. And to create the auxiliary instruments, new auxiliary technologies are needed.
That is, the real direction of movement is reverse — first Tree 4 grows, then Tree 3, then Tree 2, and only then does Tree 1 become possible.
For AI this looked as follows: first in 2007 CUDA appeared — a new GPU programming technology (Tree 4). This made possible the creation of next-generation GPU clusters — H100 and A100 (Tree 3). Those in turn provided the instruments for practically training transformers at industrial scale (Tree 2). And only then did the theoretical idea from the 2017 paper "Attention Is All You Need" become GPT-4, Claude and all of modern AI (Tree 1).
Note: the transformer idea appeared in 2017. But the infrastructure for its realisation was being built from 2006 to 2020 — in parallel and independently. No one planned this as a unified programme. It simply happened that at the right moment all four trees were sufficiently grown for AI to become possible.
Why This Matters
This is precisely why genuine technological progress is so slow and expensive. Behind every visible step stand three invisible steps that had to be taken first.
And this is precisely why when countries want to stop someone else's progress — they block not the final products, but Trees 3 and 4. They do not ban rockets or aircraft. They ban ASML lithography machines, special machine tools, precision bearings, metal alloying technologies. Whoever owns Trees 3 and 4 controls Trees 1 and 2. Always.
Top comments (0)