What lies inside a stone, no eye can see. A stone does not wander or dance. It rests idle through seasons, indifferent to the world around it. On mountain ridges, there are boulders our grandparents looked upon, our parents knew, and we find today, unmoved in the same spot, shifted only by an earthquake. Even then, barely.
In the earliest times, we knew nothing of what lay inside a rock. Today we see minerals, chemical compounds, atoms, protons, neutrons, and quarks, even as the search for what lies deeper continues.
We have never understood any natural object in its absolute totality. Each one holds its unknown layers, and none is entirely predictable. But that incomplete knowledge never stopped us from building. From the stone axe to the space shuttle, our approach has remained consistent. We study materials to the utmost perimeter of our understanding and combine them into sophisticated structures. Upon that foundation we raise another tier, and upon that tier another, erecting ever more ambitious architectures.
Artificial intelligence has hidden layers, and it is far from infallible. It appears to be the latest crowning achievement in that same ancient lineage of building things. But treating it as merely another powerful artefact of engineering is a profound mistake.
To understand why requires re-examining our fundamental ideas. Ludwig Wittgenstein famously said that ‘a picture held us captive’, warning that the mental pictures we form of everyday concepts quietly dictate the boundaries of what we can anticipate. Before we can analyse the silicon intelligence we now manufacture, we must disassemble the foundational philosophy we all carry about how we construct our world. We must ask what it truly means to understand an object, and how that gave us the confidence to build with it.
Unlike people, a thing is defined by its past. Its history dictates its future. This is why we look for intention and character in humans, but demand rigid consistency in objects. If you test a thing today and it works well, you can be confident of its tomorrow. Our ancestors struck a rock against rock and learned to distinguish unyielding stone from dried mud that would dissolve under autumn rains. Only this hard knowledge allowed them to build their dwellings and put their children to sleep, trusting that the roof would hold and the sky would not fall.
A natural thing like stone can be completely unknown at first, until we uncover some of its layers. But our quest did not stop at knowing nature as we found it. We crushed limestone into mortar and melted metals into bronze, revealing their own hidden properties. Still, the rule never changed. We only relied on them when we understood them as thoroughly as we could. But the uncovered layers could sometimes be stunning mysteries themselves, defying our expectations.
Consider salt. We know how it looks, the way it seasons food, and how it affects our health. This is the surface layer. We also know that salt consists of sodium and chlorine atoms, which forms the core. Still, place solid rock salt under extreme pressure, and it begins to flow like toothpaste, as one can see in the salt domes of the Zagros Mountains.
A single grain of salt is brittle and gives no sign of acting like paste, but when billions of them are pressed together on a geological timescale, they do. We are still untangling exactly why this happens. But in the meantime, we can continue to analyse the paste and the single grain separately, as though they were disconnected. In our effort to understand the world, recognising this separation is vital. The very same thing, when grouped into vast numbers, exhibits entirely new behaviours not present in the individual unit. This is what we call emergence.
An emergent behaviour is not always calm and scenic. It occurs in chaotic, colossal forms as well. Think about a rainstorm, the roaring climax of a conversation between the ocean and the sun. A rainstorm is born of a vast array of components, from gas molecules to aerosol particles, encountering each other at immense scale. At the core are the same simple interactions, repeated billions of times. Out of this turbulent choreography emerge powerful winds, thunderous lightning and heavy downpours.
For the vast majority of our ancestral lineage, the forces that brewed such a storm were an absolute mystery, something to fear or appease. Today we can predict the shifts in the weather days in advance with reasonable certainty, though never with absolute precision. We need the rainstorm to fill our wells and rivers and nourish our harvests, but we must ensure our shelters do not dissolve beneath it. Throughout history, we learned to build structures and guardrails to make the best use of a rainstorm.
Stone, salt, and rainstorms are all natural things that we try to understand iteratively, layer by layer. That is science. But artificial intelligence is human-made. Anything understood well enough by science is systematically put to work to build larger structures. That is engineering. Before we can evaluate the nature of artificial intelligence, we must dismantle the foundational assumptions around engineering as well.
There is a key difference when we move from natural objects to human creations. Human-built things, however colossal they may be, can be understood right up to the very boundary of our scientific knowledge. In engineering, there are no skipped layers or logical leaps. Even the theoretical physicist sipping coffee in a rooftop bar would agree that the pillars beneath must respect this rule. They do not want a brick in the foundation to suddenly defy its nature and slip out for a run when no one is looking.
Take the Eiffel Tower, for example. Every force, every weight and every anticipated gust of wind was calculated before placing a single iron beam. A team of fifty engineers and three hundred workers spent two years mapping every millimetre of the structure, accounting for every one of the two and a half million rivets required to hold it together. It is an architectural marvel that translated the best scientific knowledge into physical form, leaving no unknown layers.
Engineering is all about this kind of controlled orchestration, building ambitious structures with utmost transparency of their inner workings while anchoring them firmly in science. Still, even with this rigorous discipline, we run into formidable challenges when we step into uncharted territory.
If you travel south from Seattle to Tacoma and look beneath the water, you will find broken steel structures nearly nine decades old. These are the remains of the Tacoma Narrows Bridge, built in 1940 and designed by Leon Moisseiff, the renowned engineer who was also involved in the creation of the Golden Gate Bridge. But within months of opening, this massive steel structure twisted like a ribbon and crumbled into the water. The subsequent analysis was handed off to scientists, who later explained that the collapse was due to aeroelastic flutter, a violent, rhythmic vibration where the wind and the steel began to feed off each other’s energy in a destructive dance.
But engineering failures are not always about a lack of scientific knowledge. We humans are obviously imperfect, even with all the science at hand. We make errors. Sometimes, we are collectively cruel.
‘I will be ready to get back to teaching. Kids have a wonderful way of bringing you back to reality,’ said the schoolteacher when she was asked what her plans would be after returning from outer space. Still, when the entire world held its breath to watch her rise, a single O-ring compromised by human negligence turned Christa McAuliffe and all that bright hope into glowing smoke against a clear blue sky.
We fail, sometimes hauntingly. In our unending ambition to walk on water and touch the sky, we leave a loud silence at some of our dinner tables.
But if you are still diving beneath the Tacoma Narrows, you should come to the surface and look up. There, you will see a grand steel structure, the new Tacoma Narrows Bridge, built in 1950. It still stands tall and strong, waltzing with the winds. In a way, the skeletons of the crumbled bridge beneath the water hold together this new one, and every suspension bridge we have built since. The life of the fallen bridge lingers in all our physics textbooks.
Every successful thing we have ever engineered, from stone axes to space shuttles, rests upon the core principle that it is understood right up to the scientific boundary, before we put it to use. The rainstorm is a natural mystery that exists in the world, while the Eiffel Tower is a human-built colossus that we engineered. Where, then, does the artificial intelligence of today fit?
The neural network sits at the core of modern artificial intelligence. Its simplest building block is a single node that performs basic maths. It receives inputs, weights each one according to its importance, and transmits an output. This basic setup is repeated billions of times digitally and arranged into intricate hierarchies to create the giant network. What happens within that node and how those hierarchies are designed have evolved across decades of research to tackle specific applications.
But to understand how a neural network as a whole functions, it helps to imagine a simple clay road. If a town redirects its traffic onto a clay road for a week, deep impressions will form in the mud along the most common paths of the passing cars. We can then use those physical impressions to pave a permanent shortcut of the right width.
A neural network functions, in a sense, like that clay road. While the clay records only the path of a tyre, the digital surface of a neural network deforms under the pressing of every tiny piece of information, permanently recording the common shapes, patterns, and subtle features of the millions of digital artefacts we drive across it. This grand deformation is stored across billions of weights across the network.
If a neural network is shaped by the data of a million chess matches, it gradually hardens into an unbeatable chess engine. Through similar iterations of structural shaping, we arrived at a system for the board game Go, another for the folding of proteins, and finally, the colossal language models that take the form chatbots or autonomous agents.
Building the bedrock of a neural network demands sophisticated engineering. Scaffolding it with interfaces and guardrails requires engineering as well. But shaping that core network with vast amounts of data is not engineering. It is an act of forging.
In the Eiffel Tower, iron beams were forged into specific shapes before being placed into a calculated, predictable position. With an AI, however, we engineer a sophisticated yet shapeless tower first, only to forge its final form by hammering it with a million diverse objects until it settles into an unimaginably complex, useful structure. This reverses the foundational logic of building.
The artificial intelligence of today is not a product of pure engineering. It is a result of crude forging.
Because of this forging process, the resulting output is often not even understood by its own builders. This is not a natural mystery. Rather, the sheer scale of the system’s complexity makes the chain of causal reasoning vanish into billions of digital weights.
In a language-generating AI, the forging process gives rise to stunning emergent behaviours that mimic human reasoning. Its performance sometimes far exceeds our own and sometimes falls laughably short, a phenomenon popularly termed the jagged frontier. Still, there is a pervasive tendency to hide this untraceable complexity behind a human mask. This anthropomorphism asks us to view artificial intelligence as a single, cohesive mind. At best, that view presents a fractured, erratic and fundamentally unpredictable intelligence.
The alternative is to recognise the splintered capabilities of these systems as an emergent property of human language itself, arising when trillions of textual fragments collide across billions of neural nodes, enabled by the digital realm. In any case, it is an emergent behaviour that demands further analysis, rather than fear or appeasement.
Even then, is artificial intelligence simply a piece of material technology that we should learn to trust in the same way we trusted space shuttles? The reality is that unlike engineering, which is liable to human error or a lack of scientific knowledge, the forging of artificial intelligence has error baked into the process itself. This is the underlying reason why these models are released for public use with sweeping disclaimers warning that they may produce errors. This is not merely pragmatic compromise under specific conditions, but across all conditions.
An AI can generate extremely constructive patterns as output while generating fatally flawed patterns with same mechanical ease. It is only reasonable to expect this erratic behaviour from something forged from vast, complex data of any kind, whether those patterns represent protein structures or human-written text.
Taking a tape to measure every single car in a town is unimaginably laborious when compared to simply redirecting daily traffic to shape the clay. The contrast widens dramatically when we scale this approach to millions of digital artefacts. For the foreseeable future, building artificial intelligence will remain a process of forging.
In traditional engineering, the unknown layers are natural mysteries, but in the forging of artificial intelligence, the unknowns are human-made complexities. We tame natural mysteries through science, and we navigate human unpredictability through character and collaboration, but the complexities inside artificial intelligence are of different nature and scale. Despite their immense benefits and our powerful guardrails, AIs remain no less volatile than the weather.
No matter how capable artificial intelligence becomes, these fundamental distinctions will remain unchanged. While traditional engineered technologies were meticulously tested piece by piece, we deployed artificial intelligence with provisional guardrails before we could understand it reliably. Unlike natural phenomena that we must simply endure, where and how to deploy it remains a human choice.
Artificial intelligence is a rainstorm that we have forged inside tiny silicon cubes, only to end up living alongside it. Ultimately, to use this colossus wisely, nothing can equal the accountability of the flawed human who wants to ensure the roof holds and the sky does not fall while we sleep.
— sAb
(RECORD 008)
I want to thank the people listed below for reading my drafts, offering critiques, and keeping me motivated.
Alphabetical list:
Kerstin Haack
Ron Watson
Sreenivaasan Gajapathy
Tamara Koskinen
Vinod Balachandran