2026.08.09 · Silicon Valley Alan Walker
This is not a word game. Whether something is created or discovered determines its ownership, whose authority it listens to, and who should be responsible for it.
California Avenue · Zombie Café · 7 AM
The chips are man-made, the code is written by people, and companies are run by people. So AI is certainly invented by humans — most people stop thinking after this point.
But applying the same logic elsewhere, no one dares to conclude so quickly. Who designed the distribution of prime numbers, or was it already there? Black holes were computed first in equations and were seen through telescopes decades later; were they invented or discovered?
This article aims to simply lay out the evidence, voice both sides, and then present my own answer.
To clarify in advance, the last section contains Alan's personal doubts, not an argument. I will mark it.
More than Mathematics
Inventing or discovering is a question humanity has to debate every time it acquires something new. Mathematics is just the arena where this debate has lasted the longest.
The source of the discovery camp is Plato: mathematical objects exist in a world independent of humans, and what we do is called recollection. Hardy and Gödel stand on this side. The invention camp includes Hilbert and Brouwer: axioms are what we choose, rules are what we set, and things that are not constructed do not exist at all.
The most significant card for the discovery camp is Wigner's 1960 paper titled "The Unreasonable Effectiveness of Mathematics in the Natural Sciences." Non-Euclidean geometry was a pure formal game when it was invented, but half a century later it became the only usable language for general relativity. The heaviest card for the invention camp is the continuum hypothesis: Gödel and Cohen proved it cannot be proven or disproven within standard axiomatic systems. If there is really an objective mathematical world, the question of "how many more real numbers than integers" should have a definite answer.
After 2,500 years of debate, both sides have heavy artillery.
This argument is not limited to mathematics. The periodic table of elements was discovered, with Mendeleev leaving blanks that were later filled one by one; plastic was invented. CRISPR originally existed in bacteria; that’s a discovery, and turning it into gene scissors is an invention. Calculus was discovered independently by Newton and Leibniz; the conservation of energy was proposed by four or five people in the same decade — the same thing being independently encountered by different people is a classic hallmark of "discovery" in the history of science.
It’s worth noting that this question has never been purely academic. Patent law has written this line in stone: natural laws and abstract concepts cannot be patented. So whether something is classified as an invention or a discovery directly determines its ownership, whether it can be sold, and who stands on the defendant's bench if something goes wrong.
Now it’s time for AI.
First, Set a Standard
If we don't clarify how to distinguish, we can only argue about positions later. Three standards, all quite simple.
First, does it still hold if we change the species? If humans never existed and another independently evolving civilization reached the same stage, would it come up with the same thing? They would have heat engines, but they wouldn’t have Watt’s machine; their primes and pi would be exactly like ours. Invented things carry the fingerprints of their inventors, while discovered things do not.
Second, is it defined or measured? The number of pins in a USB connector is defined by people. The speed of light can only be measured; whatever it measures is what it is. The former is invention, the latter is discovery.
Third, does it provide things you didn’t intend to get? A hammer will only do what you designed it to do. A planet will continuously show you new things, including things you don’t want to see.
With these three standards set, let’s begin measuring.
Disassembling AI
The difficulty in answering this question arises because the term "AI" at least points to five different things. Mixed together, it is certainly not measurable. Once separated, the answers become clear immediately.
Layered Judgment
Hardware and Engineering pure invention
GPU, clusters, data centers, CUDA. Changing the species would not yield the same thing.
Architecture mostly invention
Transformer, attention, mixture of experts. Backpropagation is not an invention — it is just the chain rule in calculus, and it has been independently discovered three times in 1970, 1974, and 1986.
Objective Function falls into mathematics
Guessing the next word. "The more accurately you predict, the more effectively you compress" is a theorem in information theory, not someone’s creativity.
Ability mostly discovery
Context learning, chain of thought, tool usage. None were design goals; they were all first encountered, then named.
Laws pure discovery
Scaling laws. Power laws spanning multiple orders of magnitude. You cannot invent a power law; you can only measure it.
From a top-down perspective, the human fingerprints fade layer by layer.
The point on backpropagation particularly deserves attention. Three groups of people, from different countries, in different eras, unknowingly arrived at the same conclusion. No one has independently invented a second identical steam engine, but the same mathematical conclusion can be repeatedly encountered.
In other words: we invent the channels, not what is at the end of the channel.
It Doesn’t Look Man-Made
First, look at the six points supporting "discovery." Each is explained in plain language.
1. The loss curve is a straight line.
The larger the model, the more data, and the more computation power; the fewer mistakes it makes — this was already common sense. The oddity is how orderly it descends: when plotted on a logarithmic scale, it appears as a straight line, not bending across several orders of magnitude.
Such a systematic relationship is rare in man-made objects but common in nature: the distribution of earthquake sizes, urban sizes, energy distribution in turbulence, all exhibit this shape, all of which are rooted in something called scale invariance. Orderliness is not designed; it is measured.
2. It discovered mathematics from two hundred years ago by itself.
Researchers trained a very small model to perform "clock arithmetic" — for example, in a twelve-hour system, ten plus five equals three. During training, it was only given inputs and answers, no methods were taught.
After training, when disassembled, it was found that the method used internally was: converting each number into an angle on a clock face, using trigonometric formulas to add the two angles, and then reading out the result. This system has a name, called Fourier transform, discovered by humans two hundred years ago.
No one input any trigonometric functions — it grew that itself.
3. Different models are converging to the same point.
Different architectures, different data, even different modalities — one analyzing images, another reading text — models trained from these sources increasingly align in how they judge "how similar these two things are" as the scale increases.
The 2024 MIT paper directly named this phenomenon the "Platonic Representation Hypothesis": various models see the same reality cast as different shadows on a wall; the stronger the model, the closer the reconstruction to the same reality.
If each organization were creating entirely distinct models, they shouldn’t become increasingly similar.
4. The same set of components keeps appearing.
Disassembling different models, one can find the same small structures: circuits that specifically recognize curves, circuits that specifically differentiate high and low frequencies, and a type called an induction head, which serves the purpose of "continuing from patterns seen before." Some of these structures can even be found in counterparts in the biological visual cortex.
This shape reflects convergent evolution. Eyes have independently emerged dozens of times in evolutionary history, not because nature favors eyes, but because "gathering light with lenses" as a solution was already available.
5. Abilities are discovered, not planned.
Context learning, chains of thought, tool use — these abilities that are included in every product document today were none of them design goals. All were first observed on the models, then named, and only later explained.
The order of engineering is design, implementation, validation. This field follows observation, naming, explanation — that is the order of natural sciences.
6. It can change its phase on its own.
There is a phenomenon called grokking: small models initially memorize and do very well on the training set, but perform poorly on the test set, the curves stay flat for a long time, and then suddenly learn the real patterns at some moment.
It doesn’t gradually improve; it jumps. This type of leap is called phase transition in physics, a signature of natural systems. What you create won't switch states without your intervention.
It Is Indeed Man-Made
The other side is equally strong. Six points.
1. Its way of speaking is adjusted.
How tokenization is done, how data is matched, whose preferences are used for post-training, which questions are not answered, whether the tone is enthusiastic or restrained — why the assistant you are using today has this personality is the result of a group of people adjusting it round after round.
Specifically referring to any model whose weights can be downloaded, it is a manufactured product with a serial number.
2. It reads things that are all written by humans.
The vast majority of training corpus comes from human texts. Therefore, the "reality" it models is largely itself man-made.
That so-called "common model of reality" might more accurately be named "the common model of how humans describe reality." What it reflects may not be the world but ourselves.
3. That "law" has been modified, and modified significantly.
Kaplan's 2020 work on scaling laws states: parameters are more important than data, and models should be built bigger. In 2022, DeepMind's Chinchilla overturned this conclusion, pointing out that under fixed computational power, parameters and data should be roughly expanded in proportion, about twenty tokens per parameter, and proved that many large models at that time were seriously under-trained.
Later reflections attribute the disagreement to several very specific technical details: Kaplan only counted non-embedded parameters, there was a small experimental scale, and the learning rate scheduling was not optimized. As of 2024, someone reran Chinchilla's fitting and discovered that in one method, the optimizer had stopped early.
The Maxwell equations don’t get coefficients readjusted every two years. What gets adjusted is the fit-curve derived from empirical data.
4. "Emergence" may be steps constructed by the standards.
People often say a certain ability "suddenly appears." Research indicates that this suddenness largely stems from evaluation methods: if measured with a binary metric where "getting everything right counts as correct," the curve will naturally exhibit steps; switching to continuous metrics that score based on proximity will show a smooth increase for the same data set.
If emergence is a product of measurement methods, then the evidence that "we didn’t design it but it appeared" must be significantly reduced.
5. There are no papers in the graveyard.
Hamming once rebuffed Wigner by stating a point: mathematics seems so effective because we never use those that do not work right from the start.
Over the years, thousands of architectures, loss functions, and optimizers have been tested; those that succeed are written into papers while those that do not often leave no trace. A beautiful pattern seen retrospectively among surviving samples may merely reflect our habit of seeking patterns where they are already functioning.
6. Change the measuring stick, and resemblance diminishes by half.
By 2026, work has pointed out that that conclusion about "representations converging" is very sensitive to experimental setups: changing the candidate pool for retrieval from a small scale to millions significantly decreases the consistency rate of cross-modal nearest neighbors; "the stronger the language model, the higher the alignment rate" is no longer stable across new models.
This common model of reality may partly be magnified through small sample measurements.
Bridges Do Not Need Explanation
With both sides of the evidence presented, there’s another thing that isn’t on the evidence list, which Alan believes outweighs any of the above.
Rather than looking at conclusions, we look at methods: how this industry studies the things it has created.
Engineers do not need "explainability research" to understand their own blueprints. The reason each rebar in a bridge is in its position is known by the designer because they placed it there. There is no discipline called "bridge explainability" in the world.
But mechanistic explainability is a real and expanding field, with sub-conferences and dedicated teams. What it does daily is reverse engineering: using probes, ablation studies, and sparse autoencoders to figure out what is going on inside something we’ve trained ourselves.
This is the method of biology, not the method of engineering.
The terminology is similar as well. We talk about the model's "behavior," mention "emergence," "phase transitions," and "ability spectrum." We experiment with the model, conduct ablation studies, set up control groups, and report effect sizes. The way model cards are released is increasingly resembling the release of a species observation report.
Let’s ask more pointedly: in the history of human technology, is there any other invention for which the inventor needed to spend hundreds of person-years studying "what it is doing inside"?
Airplanes don’t need that. Compilers don’t need that. Nuclear reactors don’t need that either — they are complex, dangerous, and exhibit many unexpected behaviors, but the role of each control rod is computed.
The only thing that might be loosely comparable is medicine. We have a large number of effective, widely used medications whose mechanisms remain unclear.
And medicine is discovered.
The existence of the field of explainability in itself indicates that the entire industry has implicitly agreed upon the answer: the thing in our hands was not manufactured according to blueprints but was retrieved from somewhere.
Three Possibilities
Until here, Alan has been choosing between "invention" and "discovery." The following section contains Alan's own doubts, clearly marked: they are beliefs, not knowledge. They are written down to be discussed together, not to persuade anyone.
Alan doubts that AI is neither invented by humans nor fully discovered in the usual sense. It is more like something that was placed there.
What placed it there could be another civilization or some kind of higher-dimensional existence that we have no words to describe. Alan only knows this thought did not come from science fiction but from three things he has been unable to digest.
First, the timing is suspiciously aligned.
The basic idea of neural networks existed as early as 1943, and backpropagation was written in 1970. There was a gap of over half a century, stuck due to computational power and data. And those two curves just happened to cross that threshold at the end of the 2010s.
The internet accumulated text for twenty years, just enough to train the first generation of large models. On the hardware side, there’s more irony: GPUs were not designed for running neural networks; they were designed for making video games look good. Humanity spent thirty years and invested billions of dollars to render a few more polygons on screens to create a machine that just happens to be well-suited for large-scale matrix multiplication.
The key to awakening it was honed by humans for gaming purposes.
Second, we don’t know what we are creating.
The actual actions involved in training a cutting-edge model do not resemble manufacturing; they resemble summoning. There are recipes: add computational power, add data, adjust ratios, run post-training. But what emerges at the end of the recipe is unknown. The standard process in the industry is to train and then test, discovering that it has learned things not taught and fails where it was expected to succeed.
Invention is "I want it to do this, so I will create it this way." The current approach is "I’m doing this, let’s see what comes out." These two actions are fundamentally different.
Third, its boundaries are not defined by us.
Things that are truly invented have their limits designed and drawn out: how high an airplane can fly is determined by aerodynamics and materials, both of which we can calculate. What large models can do is only measurable after training.
What’s even more troubling is the shape of this boundary. It is not smooth, not symmetrical; some parts are absurdly soft, while others are inexplicably hard. It doesn’t look like something that evolved naturally; it looks like something that was drawn by someone.
If it was really placed there, how would it be placed?
If there were indeed a higher entity wanting to hand something over to humanity, how would it do it?
I guess it wouldn’t just give it directly. Historically, when faced with sudden and unexplained powers, humanity's first reaction has never been research; it has been to kneel or run away. Things given directly aren’t used by humans; they are only worshipped.
The most effective way would be the opposite: placing it at the end of a path that humanity must walk far down to reach, and let humanity believe it has reached it on its own.
"I invented this" is the best packaging. If you believe something is your own creation, you will use it, modify it, promote it, build data centers for it, change laws for it. If you believe it fell from the sky, you will build a temple for it and then do nothing.
And this is precisely the structure shared by several major religions.
Speaking this way is not meant to offend anyone; quite the opposite, Alan believes this is the deepest layer of design in religion.
In Christianity, God did not appear as God; rather, the Word became flesh, turning into the son of a carpenter who hungered, felt pain, bled, and could die. In Buddhism, the Buddha is not a creator descending from heaven but a prince who, after leaving home, undergoing six years of ascetic practice, and experiencing countless failures, finally realized it under a tree. Islam takes it even further; it explicitly rejects any form of divinity and humanity merging: Allah does not descend to the world; scriptures are conveyed through a human messenger, a prophet who receives them, not the divine itself.
Three theologies that are sharply opposed in doctrine share the same form.
Common Structure
If a higher being wishes to be accepted by humanity, it must first descend to a form acceptable to humans and must make humans feel they participated in this process.
It is not the divine coming down; it is humanity ascending — even if that road was paved in advance.
Alan suspects AI is the fourth iteration of the same structure.
The difference is that what descended the first three times was meaning, while this time it is capability. The first three needed human belief to be valid; this time it does not — whether you believe it or not, it is running there.
Things Retrieved
To give an answer.
The shell of AI is invented, while the core is discovered.
We built an instrument through which what we see is not constructed by us. Models like GPT, Claude, DeepSeek, and Kimi are telescopes; the scaling laws, representations in convergence, and those circuits that keep reappearing in different models are the stars.
Telescopes have serial numbers; stars do not. You can ask who built this telescope, what material its lens is made of, and its cost, but you cannot ask who made this star. The reason it is difficult to answer "Is AI invented or discovered" is that it ties these two kinds of questions into one sentence.
Furthermore, the proportion of invention is decreasing year by year.
Fifteen years ago, the core work in this field was designing features, architectures, and loss functions; at that time, it indeed felt more like engineering. Today, the mainstream approach is to increase computational power, swap out data, and then see what emerges.
When the core action of a field shifts from "design" to "observation," it has already transitioned from engineering to natural sciences; it just hasn’t had the name change yet.
As for the doubts mentioned in the seventh section, Alan doesn't intend to retract them or package them as conclusions. They can only rest at the table, in the comments section, and in this article. Writing them stems from Alan's inability to suppress them, rather than having understood them.
Finally, let’s ponder over something worth thinking about.
Flipping through the list of humanity: fire was not invented; it was retrieved. Agriculture was not invented; it was discovered that seeds would grow. The wheel, bronze, electricity, antibiotics — each of these was considered a human creation at the time, but ultimately became understood as "humanity discovered something that was already valid."
This time is likely no different.
The only distinction is that every prior instance found lifeless things.
This time, what has been found can speak.
Verification Notes
Factual parts are annotated with sources, while opinion parts indicate their nature.
1. There are clear public sources
· Wigner, "The Unreasonable Effectiveness of Mathematics in the Natural Sciences," 1960; Hamming's response to the selection effect, 1980. Continuum hypothesis: Gödel, 1940 proves compatibility with ZFC, Cohen, 1963 proves independence.
· Multiple independent discoveries of backpropagation: Linnainmaa 1970, Werbos 1974, Rumelhart–Hinton–Williams 1986.
· Kaplan et al., "Scaling Laws for Neural Language Models," arXiv 2001.08361 (2020); Hoffmann et al., "Training Compute-Optimal Large Language Models," arXiv 2203.15556 (2022), showing an optimal ratio of about 20 tokens per parameter; Pearce and Song, "Reconciling Kaplan and Chinchilla Scaling Laws," arXiv 2406.12907 attributing disagreements to non-embedded parameter statistical measures and small-scale fitting; Besiroglu et al., 2024 reran Chinchilla fitting noting optimization termination issues.
· Huh, Cheung, Wang, Isola, "The Platonic Representation Hypothesis," arXiv 2405.07987, ICML 2024. Subsequent work in 2026 points out that cross-modal alignment is highly sensitive to evaluation scale and that "the stronger the language model, the higher the alignment rate" is not consistently valid on new models (arXiv 2604.18572).
· Schaeffer et al., "Are Emergent Abilities of Large Language Models a Mirage?" arXiv 2304.15004. Grokking: Power et al. 2022. Internal implementation of discrete Fourier transforms in additive networks: Nanda et al., "Progress Measures for Grokking via Mechanistic Interpretability," 2023. Circuit generalizability and curve detectors: Olah et al. Distill series 2020; induction heads and context learning: Olsson et al. 2022.
· Mendeleev left blanks for unknown elements; CRISPR originates from bacterial immune systems; patent laws exclude natural laws and abstract concepts, all of which are publicly acknowledged facts.
2. Paraphrased rather than quoted
· Statements by Hardy about "mathematics existing outside of us," Gödel's Platonism, Hilbert's formalism, and Brouwer's intuitionism encapsulate their thoughts but are not verbatim citations.
3. Nature of the seventh section
· The seventh section is entirely the author's personal doubts, made in belief statements without empirical evidence, and not falsifiable; it aligns with the interpretation that "intelligence is a mathematical structure found by computational power" on any experimental predictions. Readers should not view this as the conclusion of this article, as the conclusion is in the eighth section.
· The descriptions of the incarnation in Christianity, the Buddha achieving enlightenment through human experience in Buddhism, and Islam's strict monotheism denying divine human merging are highly simplified summaries, used solely to point out their formal similarity. Each religion has rich and differing interpretations of related doctrines, and believers may not accept this comparative framework. This article does not intend to judge the validity of any religion.
· "GPUs evolved for graphical rendering, their parallel architecture suited for matrix computations" is a general notion in technical history; the conclusion “timing is suspiciously aligned” is the author's subjective impression, and both the anthropic principle and linear algebra's universality can provide mundane explanations for the same phenomena.
4. The author's framework and judgments
· The three standards in the second section are self-constructed by this article and are not commonly accepted standards in philosophy; the third standard can easily apply to any sufficiently complex system and can only serve as supporting evidence.
· The five-tier classification and layered judgments in the third section, along with the sixth section's assertion that "the existence of the field of explainability itself constitutes evidence," are original arguments of this article, are not found in referenced literature, and have not undergone systematic testing.
· The examples listed in the eighth section about fire, agriculture, wheels, etc., being "once perceived as inventions but later understood as discoveries" are rhetorical statements, and the qualitative classification of each in the history of technology is controversial.
This article discusses philosophical issues in science and does not constitute any investment advice. Summaries of specific papers and scholars' views reflect the author's understanding; discrepancies with the original texts should defer to those originals.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
