Featured

Who Follows Whom: Does Form Follow Function, or Function Form?

AI - a tool Hits: 1145 32 minutes read

Form follows function? The inductive prejudice and the stories that help us see it.

On how a phrase coined by an architect in 1896 describes, backwards, what happens today in computing systems trained on data.

A note on vocabulary. In this text I do not use the phrase "artificial intelligence" or its abbreviation (AI), and I avoid expressions that attribute states of mind or intentions to a machine: "thinks", "understands", "knows", "wants", "sees", "chooses", "advises", "keeps company". I speak of computing systems, statistical models and conversational interfaces, and I describe what they produce, calculate or reproduce, not what they "feel" or "understand". I keep only the established technical names ("machine learning", "neural network", "training") and the titles of the works cited, so as not to falsify sources. The characters in fairy tales think and feel; analogies with them concern only the structure of a mechanism, not a mind of the machine.

A workshop, a phrase and a reversal

My father repaired engines. His workshop smelled of oil, had a lamp hanging above the workbench, a radio always on, and tools laid out in an order that, to me as a child, seemed incomprehensible. It was not an order by labels or alphabet; it was his. And yet, when he needed a spanner, his hand knew where it was, without looking.

If I asked him why he kept them that way, I got no answer, but a phrase he said often, in front of anything broken: "It can still be mended." The workshop's form had not been designed to a plan. It had settled, over years, after his hands had done the same things thousands of times.

I often think of that workshop when I read about computing systems trained on data, because it poses, on a small scale, a much larger question: who follows whom — does form follow function, or function form?

The formula that put the question into circulation belongs to the American architect Louis Sullivan. In 1896, in an essay about the tall office building, he wrote that form ever follows function.

The formula became the slogan of modernism: first you know what a thing is for, then you give it the appearance that suits it. A chair has four legs because it must stand, not because four legs look nice.

In systems trained on data, this relationship turns upside down. And that is precisely why a whole series is worth writing about it.

Let me explain why. When you build a system by training, you cannot know in advance, in the strong sense of the word, what function it will end up calculating. You know what function you want to obtain: to classify pictures of cats, to produce a translation, to estimate the weather. But the function it actually calculates — the one it reproduces on data encountered only after training — depends on a series of choices of form: the network's architecture, the training procedure, the data used, the constraints imposed. Engineers have a precise name for this set of choices. They call it inductive bias, that is, inductive prejudice.

The word "function" has, in addition, two senses that meet here. One is mathematical: a machine-learning model is, in essence, a function mapping inputs to outputs, and the training procedure selects, from among countless possible functions, a suitable one. The other is social: what do people, in fact, do with this technology? The two senses do not overlap, but they influence one another, and much of this text tries to hold them in the same frame.

So that the discussion does not remain abstract, I did something I often do in my texts: I called upon the Romanian fairy tale for help. Not as decoration, but as a mnemonic. A reader of seventeen or twenty has almost certainly met, at school, Harap-Alb (the hero of Ion Creangă's classic Moldavian fairy tale), the two daughters in Creangă's tale, or Făt-Frumos — the Romanian "Prince Charming" — who sets out after "youth without old age". These images are already in their head. They can be used as hooks on which to hang ideas that would otherwise slip away.

One warning, though, so as to be fair to the fairy tale. The analogies in this text are interpretations, not technical equivalences. Harap-Alb is not a neural network and was never thought of as one. Where the analogy holds, I use it; where it breaks, I say it breaks. One further point of precision: Harap-Alb and the tale of the daughters are literary fairy tales, written by Ion Creangă on popular motifs, while Master Manole is a ballad and legend, not a fairy tale proper. I gather them under the same roof anyway, because they work as shared images of a culture.

The text follows eight steps. It begins with what the inductive prejudice concretely is and where the idea comes from. It continues with the way modern architectures build in hypotheses about the world. Then it passes through the field's most important dispute, Rich Sutton's "bitter lesson" and the replies to it. It stops at what we know, publicly and at this moment, about what functions these systems in fact fulfil in current use. And it closes with the limits and the costs of form.

What the inductive prejudice precisely is

We begin with the frightening word: "prejudice". In everyday language, prejudice is a bad thing: an opinion formed before knowing the facts. In machine learning, however, the word has a technical sense and, surprisingly, an almost positive connotation. The inductive prejudice is the set of hypotheses built into an algorithm that allow it to generalise from training examples to new examples.

The classic definition, attributed to Tom Mitchell (1980), says that an inductive prejudice makes a learning algorithm favour one solution or interpretation over another, independently of the observed data. The emphasis matters: the inclination does not come from the data, but from before them.

Why is it necessary? Because training data are always finite. Any finite set of examples is compatible with an enormous number of different functions, all of which fit the training examples perfectly and diverge completely elsewhere. If the algorithm contains no criterion favouring one of them, there is no principled selection, and the result is arbitrary. A survey text from the field of Bayesian networks puts the idea simply: generalisation is impossible without an inductive prejudice, precisely because the training sample is finite.

The link with "form follows function" is already visible. From among all possible functions, the prejudice favours the category of functions that "have the right shape". Form is not the decoration of the function, but its condition of existence.

The no-free-lunch theorem

There is a formal result that strengthens the idea. The "no free lunch" theorem, associated with the name of David Wolpert (with variants published in 1995–1996), states, in free translation, that no learning algorithm succeeds on all possible problems. An algorithm that performs excellently on one class of problems will perform, on average, just as poorly on another. There is no universal learning algorithm.

As synthesised by Yoshua Bengio and his colleagues, the same point sounds like this: to obtain generalisation you need a set of inclinations over the space of all functions; there is no fully general algorithm; and any algorithm generalises better on some data distributions and worse on others. In other words, the choice of form is unavoidable. You cannot not choose. Even the one who claims to impose no prejudice imposes one — only without knowing it.

Here is the first reversal of Sullivan. For Sullivan, the architect chooses the form after understanding the function. In machine learning, the researcher does not, as a rule, know the exact function being sought; that is precisely why it is estimated from data. He chooses the form under uncertainty, in the hope that it is the right one.

A word with philosophical roots

The term "inductive" did not appear in computer science. Inductive reasoning, in philosophy, means generalising from particular cases to a broader conclusion: the sun has risen every day until now, so it will rise tomorrow too. David Hume observed in the eighteenth century that this passage cannot be logically justified. That something has happened a thousand times does not guarantee it will happen the thousand-and-first. We believe in induction because otherwise we could not live, not because we can demonstrate it.

Machine learning inherits this old problem, formalises it and makes it measurable. Any model that generalises performs a Hume-like extrapolation, and the inductive prejudice is the technical name of that leap: what hypotheses about the world are built into the model and make the leap possible.

In short, a minimal example

The simplest example is linear regression. When you choose a linear model, you assume the relationship between quantities is a straight line. That is its inductive prejudice. If the world really is approximately linear, the model converges quickly, even with little data. If it is not, it produces systematic errors, however long you train it. This is where the classic compromise between bias and variance appears: a model with a strong prejudice converges quickly, on little data, but may underfit if the built-in hypotheses are wrong; a model with a weak prejudice is more flexible, but needs far more data so as not to get lost in the noise.

The first image from the fairy tale: the same task, different dispositions

For this first idea, Creangă's tale "The Old Woman's Daughter and the Old Man's Daughter" is enough for me. In it, two girls are placed before similar happenings: they must choose how to behave towards those they meet on the road, and do tasks that seem equal. The results differ radically — not because the task is different, but because the disposition with which each meets it is different.

The image is not a demonstration, but a mnemonic: the same "problem", with different "priors", leads to different solutions. A model with a prejudice of one kind and another with a different prejudice, placed before the same examples, will produce different functions. The data alone do not determine what function results from training.

It is worth saying too what the tale does not say. In the fairy tale, the good disposition and the bad one carry a moral charge: one is rewarded, the other punished. In machine learning there is no "good" or "bad" prejudice in the abstract. There is one suited or unsuited to a particular problem. The same built-in hypothesis that helps in classifying images can trip up time-series forecasting. The no-free-lunch theorem is precisely the refusal of any universal moral.

Form as a hypothesis about the world: architecture

The inductive prejudice does not sit in one place. It slips in through several doors: the model's architecture, the training algorithm, the data and the way they are handled. But the first door, and the most visible, is the architecture.

A landmark 2018 paper by Peter Battaglia and his colleagues at DeepMind, "Relational inductive biases, deep learning, and graph networks", puts the idea in very clear terms. The authors observe that deep-learning methods that operate with relations between entities rely on a kind of relational prejudice: a hypothesis about what entities are, what the relations between them are, and by what rules they compose. In deep networks, entities and relations are as a rule distributed representations, and the rules are functions approximated by the network; the exact shape differs from one architecture to another.

The same text notes something easy to overlook: it is not only the architecture that carries prejudice. The activation non-linearities, regularisation by weight decay, dropout, normalisation, data augmentation, the training schedule and the optimisation algorithms each impose constraints on the trajectory and the outcome of training. Form, then, on several storeys.

Convolution: the hypothesis that the world is the same wherever you look

Take the classic case, the convolutional network, used in image classification. It builds in a simple, powerful hypothesis: an object remains the same object regardless of where in the picture it appears. A cat in the left-hand corner is still a cat in the right-hand corner. The network reuses the same filters across the whole image, which makes it far more efficient than one that would estimate separate parameters for each position.

Here "form follows function" applies almost in the Sullivanian sense. The required function (classify objects in images) suggests a form (sharing filters across space). But it is an "almost". The form was not deduced from a perfect knowledge of the function; it was a well-chosen hypothesis about the structure of the visual world, later confirmed by results.

Attention and sequences: the hypothesis that what matters is who links with whom

In language processing, evolution passed through other shapes. Recurrent networks build in the hypothesis that order matters and that information accumulates step by step. Attention mechanisms, in the architecture called the Transformer, build in another: that any element in a sequence can link directly to any other, and the weight of these links results from training. It is a weaker, more general prejudice. It imposes less and leaves more to the data.

I should note that, for the exact history of these architectures (Yann LeCun's work on convolutions, the 2017 paper that introduced the Transformer), I have not relied on sources found during this research, and I leave them in the to-be-verified bibliography at the end. The principle stands independently of the exact details.

Prejudice as money: how much data you need

A practical consequence worth keeping: the more complex the problem and the scarcer the data, the stronger the prejudice needed. Conversely, when you have data in abundance, you can afford a freer model. Prejudice works as a medium of exchange: you buy efficiency at the price of flexibility. And, like any currency, it has an exchange rate.

The second image from the fairy tale: the five companions

Here Harap-Alb returns. On the road to the Red Emperor, the lad meets and takes as companions Gerilă, Flămânzilă, Setilă, Ochilă and Păsări-Lăți-Lungilă (roughly "Frost-Boy", "Ever-Hungry", "Ever-Thirsty", "Eye-Boy" and "Bird-Wings-Long-and-Wide"). Each has a single trait, taken to the extreme: one freezes everything he touches, another eats without limit, another drinks without limit, another sees from great distances, another can change his shape in flight. Alone, the hero cannot pass the Spinner's trials. With them, he can.

As a mnemonic for architectures, the image is useful: each companion is a "block" with a very strong prejudice in a single direction. Ochilă has a prejudice for seeing at distance. Gerilă, for the effect of cold. None is general. Together, they compose a capability.

But here the analogy has a serious limit, and I say it plainly. In a modern neural network, the blocks are not usually laid out by a person according to distinct traits, the way Harap-Alb chooses his companions; their parameters are adjusted together, and specialisation sometimes emerges from training, not from design. So "the five companions" illustrate the idea of specialisation through prejudice, not the way a model is actually built. The reader should keep the image, not the mechanism.

The dispute: does form matter, or only scale?

We arrive at the heart of the debate that has structured the last few years of research in the field. In March 2019, the computer scientist Rich Sutton, one of the founders of reinforcement learning, published a short essay, "The Bitter Lesson", which became probably the most cited and most disputed opinion text in the field's recent history.

His argument has four steps, as Michael Nielsen summarises them. First: researchers have tried, again and again, to build human knowledge into their systems. Second: this helps in the short term and is personally satisfying for the researcher. Third: in the long run, the approach stagnates and even brakes progress. Fourth: the real advances come, in the end, from the opposite approach, based on scaling computation through search and learning.

Sutton's examples are the ones any computer-science student knows. Chess: the programmes that beat the world champion relied on massive search, not on experts' chess understanding. Go: likewise. Speech recognition and vision: statistical methods, with abundant computation, beat systems carefully built from specialist knowledge; edge detectors and classical filters gave way before convolutional networks trained on data.

Why "bitter"? Because the lesson contradicts what we should like to believe: that human thinking is special, and that the road to capable systems passes through transferring it. If Sutton is right, the road passes rather through giving it up.

What this says about the inductive prejudice

The link with our theme is direct. Any human knowledge built into a system is, technically, an inductive prejudice. The bitter lesson states, then, that prejudices thought up by people are, in the long run, a mistake: they help for a while, then they limit.

If we accept the strong version of this thesis, "form follows function" becomes almost an error: we should no longer design the form after the desired function, but leave it as free as possible and add computation. Form should disappear, giving way to scale.

The image from the fairy tale: hunger, thirst and sight

For this dispute, I choose the five companions again, but in another way than above. Let us group them into two camps. On one side, Flămânzilă and Setilă: one eats everything put before him, the other drinks everything. They are the characters of "brute force", of unlimited consumption. They have no subtlety; they have capacity. On the other side, Ochilă: he consumes nothing, he sees, and picks his target with precision.

The seventeen-year-old reader will recognise at once the scene in which Flămânzilă and Setilă get Harap-Alb out of trouble, swallowing and drinking what no one else could. That is, as an image, scale. And the moments when the hero needs Ochilă's gaze illustrate structure: it is not a matter of how much you have, but of how well you hold the target.

The tale's moral, if we must draw one for our case, is not that some of them are superfluous. Harap-Alb does not win with only one. The Red Emperor's trials call for capacity and precision alike. Here the image holds well, because it forces no verdict: the real debate in the field has, gradually, reached a similar conclusion.

The debate is won by no one

The bitter lesson was, for a time, almost a dogma in the laboratories building enormous models. Even so, the scientific community did not embrace it as firmly as is thought. A meta-survey among researchers in natural language processing found that actual adherence to the idea that simply scaling resources would solve any important problem in the field is much overestimated by respondents: people believe the others agree with Sutton more than they really do. The same study notes that researchers anticipate a reversal of the trend regarding linguistic theory and the inductive prejudice — that is, a revival of form.

On the other hand, there is a constructive critique, made by people close to Sutton's own ideas. In September 2025, Kyle Cranmer and Rich Sutton together signed a text entitled "The Bittersweet Lesson". They observe that the debate is usually framed as a false dichotomy, and Sutton himself had noted that the two approaches are not mutually exclusive, though in practice they tend to collide. Their conclusion is that we should understand how to build the discovery of structure into the learning algorithms themselves, instead of rejecting the inductive prejudice from the start. In their language, manually introduced prejudices can be seen as an intermediate stage on a longer road.

The text comes from the community applying machine learning to the natural sciences, where the instinct to build in domain knowledge — physical symmetries, for instance — is strong, and the examples where a well-motivated prejudice has helped are numerous. There, too, there are examples where removing a motivated prejudice improved the result. The authors conclude, with a rare honesty, that we are, in fact, confused.

Arguments and counter-arguments

Several of them are worth setting side by side, so the reader can form their own view.

For scale. A popular guide on a benchmarking site observes that the Transformer architecture beat recurrent networks not so much because of a better prejudice as because it parallelises more easily, and so can use more computation. If the interpretation is right, even an architectural innovation is explained by the bitter lesson: it was good because it allowed scaling.

For form. One synthetic commentary observes that architectures are not a detail: scale a bad architecture and you do not get good results; you need the right general method and scale. Another text, closer to practice, notes that inductive prejudices are most useful precisely where we cannot scale indefinitely — that is, in situations with limited data, which characterise most real applications.

An argument of nuance, from physics. Nielsen observes that the bitter lesson applies best in domains where specialist knowledge is weak or hard to express mathematically. Where we have clear symmetries, the right prejudice can save enormous amounts of data and computation.

A recent position, to be treated with caution. A 2026 preprint argues that scaling laws describe interpolation on fixed distributions and say nothing about extrapolation to shifted distributions; in finite and changing worlds, the author argues, well-chosen priors beat brute force. It is an unreviewed, programmatic text, so I mention it as a sign of the direction the debate is moving in, not as a verdict.

What remains

If we put all this together, the most honest picture is this: form has not disappeared, it has moved. The Transformer is itself a form — a very general inductive prejudice, chosen precisely because it imposes little and leaves much to the data, and because it lends itself to parallel computation. Whoever says "scale defeated form" omits that scale works only on a form that can host it.

In other words, Sullivan was half right, but in a sense he did not foresee. Form does indeed follow function — only the function is not "classify pictures of cats" but "use as much computation as possible without degrading". That is the new design function, and the successful forms are the ones that fulfil it.

Knowledge from before: priors, pre-training and a father's advice

Let us come down from the dispute to a concrete mechanism. A large language model does not start from zero when it is used. It was first pre-trained on enormous quantities of text, and only then adapted to narrower tasks. Here a question arises that touches our theme directly: what does the inductive prejudice mean, in this context?

A glossary of machine-learning terms observes that debates about foundation models and transfer learning often turn on one question: does pre-training instil inductive prejudices useful for later tasks? The practical answer is, to a large extent, yes. A pre-trained model no longer starts from random parameters; its parameters encode statistical regularities of the language and of the texts in the corpus, including relations between words and ideas expressed in text. These work as an enormous prior, estimated from data, not designed by a person.

But here a distinction appears on which our series depends. There are two kinds of prejudice: the designed kind (convolution, attention, symmetries built in by people) and the estimated-from-data kind (what settles into the model's parameters after pre-training). The bitter lesson rejects mainly the first. The second is, strictly speaking, the result of scaling. Here the debate becomes subtler still: pre-training does not remove the inductive prejudice, it produces it — only by other means.

The prejudice that comes from the data

The survey sources in the field note an aspect we often lose sight of: the prejudice does not sit only in the architecture; it can also be viewed through the lens of the training data. Data that are chosen, filtered and weighted are just as much a hypothesis about the world. A text on the three kinds of prejudice distinguishes the algorithmic one (deriving from the algorithm's design), the data one, and prior knowledge. For large models, as the architecture becomes uniform, the prejudice shifts more and more towards the data.

This is an important change for the reader who is not an engineer: the one who chooses the data chooses, in a sense, the form. And here it is not only a matter of technique, but of editorial power: what enters the corpus, what stays outside. This aspect returns in the last section.

The third image from the fairy tale: the advice and the horse

The fairy tale offers here two images that complete each other. The first is the father's advice. Before setting out for the kingdom, the young man receives a clear warning: to beware of the red man and, above all, of the spân — the "spotted", deceptively fair-haired man. This is knowledge built in from before, a transmitted prior. But the hero does not apply it: he meets the spân, lets himself be tricked, and pays a serious price. The image fits the idea that a correct prior, if not used, helps with nothing. For a model, the analogue would be information present in the parameters that does not influence the output in the situation where it should have.

The second image is the wondrous horse, received also from the hero's father, which helps him on the road. For today's reader, the image can suggest the idea of pre-training: a base of already-fixed regularities, which you use in a new situation. Here the analogy is the riskiest, because it invites seeing in the model a character who knows and helps.

The limits are serious here too. The horse in the tale is loyal, well-disposed and speaks; a pre-trained model has no loyalty, intentions or good faith. It does not "help" in the proper sense, but produces statistically probable outputs, on the basis of its parameters. I use the image only for structure (something acquired beforehand that lightens the road), not for conduct. A model does not explain its reasons, because it has no reasons; it has parameters, and their opacity is part of the problem we discuss further on.

What people do with these systems: the function discovered from use

Until now we have spoken of technical form. Let us turn to the other sense of the word "function": what role these systems play in people's lives, today. Here things turn upside down again from Sullivan. An architect knows what he builds a hospital for, and a station. The providers of these systems could not know with certainty, at launch, what people would do with them. They found out by observing use.

The data on ChatGPT: asking, doing, expressing

The most cited research in this area is the paper "How People Use ChatGPT", published as NBER Working Paper 34255 in September 2025 by Aaron Chatterji, Thomas Cunningham, David Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan and Kevin Wadman — a team including researchers from Duke, Harvard and OpenAI. The authors use an automated pipeline that preserves confidentiality to classify a representative sample of conversations, without any human reading the messages.

What do they find? That messages unrelated to work grew from 53% to over 70% of use by July 2025. That the most frequent topics are "seeking guidance", "searching for information" and "writing", together with almost 80% of conversations. That writing dominates in professional use, which shows, the authors say, the unique capacity of these systems to generate digital outputs compared with search engines. That programming and self-expression each account for a relatively small share. And that, overall, the economic value comes from decision support, especially in knowledge-intensive occupations.

More interesting for us is the typology of intent. The researchers divide messages into three: Asking (I ask for information or advice, in order to decide something), Doing (I ask for a result to be produced: an e-mail, some code, an image) and Expressing (I express thoughts or feelings, asking for nothing). The reported distribution: roughly 49% asking, 40% doing and 11% expressing. And "Asking" messages grew faster than "Doing" ones, and are rated higher in quality by users.

The conclusion is, in a way, unexpected: what people do most often does not resemble automation, but the consulting of a source of information and orientation, permanently available. The effective function is decision support, not task replacement.

Data on adoption: over half of American adults

The Pew Research Center published, on 17 June 2026, the report "Americans and AI 2026". The official summary notes that a growing share of American adults use chatbots, up from a third in 2024. A secondary source commenting on the study indicates that use passed the 50% threshold for the first time; I note that I take the exact figure from that commentary, not from the report itself, so I flag it as "to be verified" until it can be confirmed in the Pew PDF. The report also shows that roughly one in ten adults says they use chatbots for emotional support.

A function broader than productivity

A July 2026 report on conversational-app consumption, based on two surveys from March 2026, finds that users use several chatbots on average, and that 80% choose, alongside the productive-assistant role, relational or support roles as well, many attributing distinct "personalities" to the tools (an attribution by users, not a property of the systems). Note: I learned of it only from a summary in a newsletter, not from the original report, so I treat it as a clue.

Taken together with the above, the picture taking shape is that, for the general public, these systems are used at once as a working tool, as a source of orientation and, by some users, as a substitute for conversation. This is the emergent function: it was not specified by the designer; it came out of the behaviour of hundreds of millions of people.

Data on professional use: the Economic Index

Anthropic periodically publishes an "Economic Index" tracking how Claude is used in the economy. The January 2026 report introduces the so-called "economic primitives", basic measures covering competencies, task complexity, degree of autonomy, success and purpose (work, personal, education). Among the findings: use remains concentrated on a small number of tasks, mostly programming (the top ten tasks amounted to 24% of the conversations analysed); the observed tasks demand more schooling than those in the wider economy; and tasks requiring higher education are carried through in 66% of cases, against 70% for those requiring less than secondary school.

The March 2026 report shows that use on Claude.ai has become less concentrated (the top ten tasks fell from 24% to 19%), and programming is migrating from chat to the API, where Claude Code is growing. The June 2026 report goes further: sessions consist more and more of long-running agentic tasks, and chat transcripts no longer fully capture how people use these systems — which obliged the team to adapt its measurement methods.

This is a weighty observation for our theme. The function of these systems moves faster than the instruments with which we measure it. Even as we try to define it, it changes.

The fourth image from the fairy tale: the unforeseen trials

Here I return to Harap-Alb, from another angle. The spân sets the hero trials: to fetch the salad from the Bear's Garden, the stag's hide with precious stones, the Red Emperor's daughter. The hero does not know in advance what the next trial will be, nor whether the companions he has will suit it. He discovers, from one trial to the next, what each is for.

It is the best image I have for the emergent function: you do not know in advance what the tool will be for; you find out along the way. But here too the limit of the analogy must be said. In the tale, the trials are set by an adversary, and the hero passes them with help. In the real world, the "trials" are the needs of millions of people, and no one "sets" them from a central place. There is no spân of these systems; there is, rather, a crowd of needs that cluster around a single tool.

A methodological warning

All these data have, finally, two limits that an honest analysis cannot skirt. First: the richest data come from the providers themselves. The NBER study has authors from OpenAI, and the Economic Index is published by Anthropic. That does not invalidate them, but it means the public picture of these systems' "function" is built largely by those who sell it, with methods the reader cannot independently verify. Second: self-report surveys have their own biases too; the NBER authors cite research showing that people tend to overstate "productive" uses and understate "frivolous" ones. That is why, when we speak of the function of these systems, we are in fact speaking of the measured function — and measurement itself is a form.

When the world changes: the limits of form

Every inductive prejudice is a bet on the world's stability. Convolution assumes objects stay the same in any corner of the image. The prejudice built into a language model implies the hypothesis that the future resembles the past in the data well enough. As long as the bet holds, form serves function. When the world changes, it may stop serving it.

Technically, the phenomenon is called distribution shift: the data the model is used on no longer resemble those it was trained on. The 2026 preprint mentioned above starts exactly here. Its argument is that scaling laws describe interpolation on fixed distributions, and that falling perplexity does not automatically mean rising reliability downstream. Being an unreviewed paper, I read it as a signal of direction, not as a settled conclusion. But the problem it raises is real and well known: a model produces reliable results inside the training distribution and unreliable ones at its edge.

An example from another area helps. A paper on weather forecasting by deep learning notes that purely data-based models have made rapid progress and reach performance comparable with classical numerical methods, at lower computational cost, provided the data are sufficient; and that inductive prejudices, when well chosen, speed up training and improve generalisation. In other words, in a domain with known physical laws, built-in form saves exactly where the data fall short.

The fifth image from the fairy tale: youth without old age

For this section, the fairy tale "Youth Without Old Age and Life Without Death" (by Petre Ispirescu) offers an image that, for a seventeen-year-old reader, says more than a page of statistics. Făt-Frumos reaches a world where time does not flow: what seem to him a few days are, in fact, years, and his life in that place runs without trouble as long as he stays inside. When he wants to return to the world he left, he finds that time has meanwhile taken its share: his parents' house is a ruin, the people of old are gone, and he himself cannot withstand contact with real time.

As a mnemonic for distribution shift, the image is powerful. A function estimated on data from a stable, closed world — that is, on fixed data — degrades when taken out into the world where things change. Not because it was weak inside, but because the rules of the place where it was estimated are no longer those of the place where it is used.

It must be said, though, what the tale does not say. The fairy tale is a parable about death and human limit; it has nothing to do with statistics. The hero passes through an irreversible fall, and a model can be retrained, adjusted, corrected. The image helps us feel the risk, not measure it. Measurement remains the business of technique.

What this means in practice

For the reader, the conclusion is simple and useful. When a computing system delivers you a result, the question worth asking is: does my situation resemble the data it was trained on? If yes, you have reasons for confidence. If not, reliability falls, and the system does not always signal the fact. A good tool is not only one that answers correctly, but one whose limits you know and can work within.

The cost of form: who is "walled in"

Until now I have treated form as a technical matter. But every construction has a cost, and the cost does not always fall on the one who builds. Here is the place for the heaviest of the images used in this article.

The ballad of Master Manole — a legend rather than a fairy tale proper, tied to the building of the monastery at Curtea de Argeș — speaks of a monastery that cannot rise: what the masons build by day collapses by night. The only solution is to wall something human into the structure: the first who will come in the morning with food — and that turns out to be Manole's own wife. The construction holds, but the price is a life.

I am not saying that the industry of these systems "walls people in". That would be an exaggeration, and analogies pressed too hard spoil more than they clarify. I say something else, more measured: every durable form has a hidden cost, and the honest question is who pays it. The Manole image obliges us to ask it.

A few places where the question arises

The first: the data. We have seen that in large models the prejudice shifts more and more towards the data. The data come from people: texts written, pictures taken, code published, by authors who in many cases were not asked and were not paid. The legal status of this use is disputed in many jurisdictions, and I take no position here; I flag it as a place where a cost is paid.

The second: alignment and annotation work. Models do not reach the form in which we use them through computation alone, but also through human work of evaluation and correction. The subject is documented in investigative journalism and in research, but it is not part of the sources I gathered for this text, so I bring no figures and encourage you to verify it separately before developing it into an article of your own.

The third: attention and relation. We have seen that roughly one in ten adults says they use chatbots for emotional support, and another report suggests most users seek relational roles too. Here the cost, if there is one, is harder to name: who answers when a product designed for everyday use comes to be, for someone, the only channel of conversation they still turn to at three in the morning? I have no answer, and I do not believe anyone has one yet.

A rule of hygiene

That is why I propose to the reader a simple rule, which I apply myself when writing: every time a text says "the system does X", it is worth asking for whom and at whose expense. Function is never neutral. It is defined from a point of view, and the point of view has interests.

Closing: who defines the function?

We set out from a phrase of 1896 and have arrived at a place where it no longer works as it was meant to. Let us gather the threads.

Form does not follow function; it makes it possible. In machine learning, the inductive prejudice is the condition for a system to be able to generalise at all. The no-free-lunch theorem shows there is no universal learning algorithm, so the choice of form is unavoidable, even when it is not explicit.

Form has moved, not disappeared. Sutton's bitter lesson obliged the field to give up many hand-designed prejudices, but the Transformer, the training data and pre-training are all still forms. The "bittersweet lesson" of Cranmer and Sutton suggests the question stays open: how do we discover structure instead of imposing it.

The function is discovered from use. The public data show, today, a picture in which people ask, do and express, and use is shifting from chat towards long-running agentic tasks. The function moves faster than our instruments of measurement.

Whoever measures chooses. The richest data come from providers. Surveys have their own biases. Any public picture of these systems' "function" is a constructed image, not a photograph.

The closing image

I return to Harap-Alb. Not because the fairy tale has an answer, but because it has a posture. The hero does not know what trial comes next. He has companions whose traits he discovers on the road, a wondrous horse received beforehand, a piece of advice he ignores once and pays for. He does not control the whole, but he is not without means: he observes, he tries, he chooses his allies.

I would like the reader of seventeen or twenty to leave here with a better question than "what can these systems do?" A question like: "what hypotheses are built into this system, what data was it trained on, and who decided what data goes into the training?" That is, in the article's language, the question about the inductive prejudice, asked by a person, not an engineer.

And, to close as I began, in my father's workshop: the order of the tools seemed mysterious to me, and today I believe it was an inductive prejudice in its best sense — a shape the hand had learned from thousands of repairs, without ever formulating it. And his phrase, "It can still be mended", seems to me the best rule for today's systems: a form that produces errors is not thrown away, but examined and repaired — if we still know how.

Bibliography and sources

The inductive prejudice and architectures

  • Mitchell, T. (1980), definition of inductive prejudice, cited through Battaglia et al.

  • Battaglia, P. et al. (2018), "Relational inductive biases, deep learning, and graph networks", arXiv:1806.01261.

  • Goyal, A., Bengio, Y. (2020), "Inductive Biases for Deep Learning of Higher-Level Cognition", arXiv:2011.15091.

  • "A Primer on Bayesian Neural Networks: Review and Debates", arXiv:2309.16314.

  • Wolpert, D., "no free lunch" theorem (1995–1996), cited through the sources above.

The Bitter Lesson debate

  • Sutton, R. (2019), "The Bitter Lesson".

  • Cranmer, K., Sutton, R. (2025), "The Bittersweet Lesson", theoryandpractice.org.

  • Michael, J. et al. (2022), "What Do NLP Researchers Believe? Results of the NLP Community Metasurvey", arXiv:2208.12852.

  • Nielsen, M., essay on the Bitter Lesson, cognitivemedium.com/bitter-lesson.

  • "Epistemic Compression: The Case for Deliberate Ignorance in High-Stakes AI" (2026), arXiv:2603.25033, unreviewed preprint.

The function of these systems, public data

  • Chatterji, A. et al. (2025), "How People Use ChatGPT", NBER Working Paper 34255.

  • Pew Research Center (17 June 2026), "Americans and AI 2026: Chatbots, Smart Devices and Views on Impact".

  • Anthropic, Economic Index: January 2026 ("Economic primitives"), March 2026 ("Learning curves"), June 2026 ("Cadences").

  • "The State of Consumer AI" report (July 2026), known here only through a newsletter summary.

To be verified before publication

  • Sullivan, L. (1896), "The Tall Office Building Artistically Considered", the original source of the formula.

  • The primary sources on convolutional networks and the Transformer architecture (not part of this research).

  • The 50%+ chatbot-use figure from the Pew report, cited here through secondary commentary.

  • The fairy-tale texts: Creangă, "Povestea lui Harap-Alb" and "Fata babei și fata moșneagului" ("The Old Woman's Daughter and the Old Man's Daughter"); Ispirescu, "Tinerețe fără bătrânețe și viață fără de moarte" ("Youth Without Old Age and Life Without Death"); the ballad of Master Manole, in the variants collected in the nineteenth century.