A minute before midnight
To speak of technology, one must first speak of time. Not because every invention needs a date, but because the distance separating one transformation from another is part of its meaning. There are scales that we can represent because they fit within our experience: a year, a decade, the duration of a lifetime. Beyond that limit, time becomes a figure we understand without quite imagining it. We know that a million years is much more than a hundred thousand, but both end up occupying a similar place in our intuition: a remote past where the differences become almost indistinguishable.
This difficulty distorts the way we think about technology. Stone, fire, agriculture, writing, the steam engine, and the computer appear as successive stages of the same journey, ordered with a regularity that never existed. Hundreds of thousands of years passed between the first transformations; between the most recent, barely a few years. To make this acceleration visible, we will traverse the technical history of humanity within a single-year calendar. January 1 will begin with Oldowan technology, one of the first widely extended and recognizable stone tool traditions. Midnight on December 31 will mark the present. This is not to affirm that each advance inevitably led to the next, but to reconstruct the trajectory that ultimately made our world possible: what capacity appeared, how long it took to do so, and what new conditions it made available.
For much of this calendar, the same technical form could last longer than entire species. Now, however, several successive transformations can pass through the life of a single person. We can experience them as individuals, but as a species, we do not have the same time to understand them. We have learned to transform the world at a much faster rate than we manage to understand what we are transforming it into.
January 1 – The world can be modified
The year begins with Oldowan technology. A hominid strikes one stone against another; a sharp flake breaks off from the impact, while the core is marked by cutting edges. The gesture seems elemental, but it contains a decisive difference from the occasional use of a found object: the stone is no longer accepted as it appears, but is acted upon to produce a property it did not previously have.
The importance of Oldowan technology lies not only in the preserved tools, but in that it constitutes a recognizable tradition, extended for an immense period across territories and generations. This continuity indicates that the operation could be transmitted and taught. The important thing was not only to produce an edge, but to know how to do it again.
A capacity that the body does not possess on its own thus comes to reside in an external object. A sharper claw would have required many generations of biological change; a flake capable of cutting could be made during the life of the individual who needed it, transported, and reproduced. Technology begins when a transformation ceases to depend on an isolated finding and can become a tradition.
During the first months of our calendar, this tradition changes at a speed that from the present seems almost immobile. Countless generations are born and disappear within the same technical form that, for each of them, was as stable as the mountains in the landscape.
April 29 – The object exists before it is finished
We will have to wait until April 29, about 840,000 years, to find the first Acheulean bifaces. The difference is not simply carving a stone more. Making a biface requires working both its faces, chaining extractions, correcting irregularities, and maintaining a general shape for a prolonged sequence. The carver does not respond only to the result of the last blow: they compare what has been produced with an object that still exists only in their mind. The final form is absent from the matter, but it already directs the work. The tool first appears as a project.
This distance establishes the first great waiting period in our technical history: about 840,000 years to go from a form of carving focused on obtaining useful flakes and edges to another that requires planning and shaping the entire stone on both sides.
November 5 – Sustaining a transformation over time
1.36 million years have passed since the appearance of the first bifaces. In our calendar it is already November 5 when deliberate control of fire, capable of producing, conserving, and using it without relying on a natural fire, can be placed. Stone had allowed matter to be modified by an object; fire allows energy to be directed and a transformation to be sustained as long as the necessary conditions are maintained.
Cooking displaces part of the work of digestion outside the body, expands what can be eaten, and eliminates harmful substances. Fire also protects, illuminates, and transforms materials. Its deepest novelty, however, lies not only in its uses, but in that it introduces a process that must be maintained over time. It must be fed, protected, watched, and produced anew.
This continuity does not depend solely on an individual, but on the group. Fire can survive because different people successively sustain the conditions that keep it active. Technical capacity thus ceases to reside only in an object and also passes to a collective process that must be preserved through a continuous chain of actions. For more than ten months of our year, each night had returned the world to darkness.
December 30, morning – Acting on a future that does not yet exist
The next major transformation arrives on the morning of December 30, about 390,000 years later. Agriculture does not arise in a single place or immediately replace hunting and gathering, but it introduces a different relationship with the future. Cultivating requires selecting seeds, preparing the land, controlling water, protecting the crop, and waiting for a result that does not yet exist. The biface had required anticipating a shape; agriculture forces us to anticipate a cycle and modify the environment so that it can be repeated.
This intervention is not limited to plants. By choosing which animals reproduce, which are preserved, and what traits are useful, communities also begin to transform other species over generations. Technology ceases to act solely on localized objects or processes and begins to modify the landscape, biological cycles, and living beings.
In some societies, the agricultural surplus allows part of the population to dedicate itself more stably to other tasks. Experience is concentrated, collective knowledge increases and is distributed among specialized people. But, as the community knows more, each individual gathers fewer of the skills necessary to sustain their life on their own. The capacity of the whole grows at the same time as the dependence between its members increases.
When this organization reaches sufficient scale, the need to record stores, properties, exchanges, and obligations also appears. Agriculture does not produce only food. It also produces accounts.
December 31, morning – Memory leaves the body
Writing appears on the morning of December 31, about seven thousand years later. Memory then abandons its exclusive dependence on the people who remember. A quantity becomes a mark; a name remains when the person who pronounced it is no longer there; an obligation can be consulted when its protagonists have forgotten its terms.
This permanence also expands the power's capacity to organize society and extend its influence. Laws can be set as reference, taxes and surpluses recorded, resources administered, and orders transmitted to distant territories. Writing allows coordinating people, production, and obligations without depending on the direct presence of those who govern or the memory of those who participate in each exchange.
A written idea can cross generations without anyone having to continuously preserve it in their memory. It can also be re-read, divided, compared, and corrected. Knowledge thus acquires a new stability, although for millennia it will continue to be scarce: each tablet must be inscribed and each scroll or codex copied by hand. Writing solves the permanence of knowledge, but not yet its dissemination.
December 31, 10:00 PM – Knowledge multiplies
We will have to wait until ten o'clock on the last day for the European printing press to multiply the reproduction of texts. Approximately 4,600 years have passed since the first writings. The copyist repeats the work on each copy; the printing press shifts that effort towards preparing a matrix capable of producing many. The change is not just about making books faster, but about separating the production of content from the reproduction of each copy.
Knowledge thus enters into a logic that will later be characteristic of the factory: effort is concentrated on building a system capable of repeating a unit, instead of completely reconstructing it each time. The same text can reach distant readers, versions can be compared, and one edition can correct the previous one.
The printing press not only multiplies books: it multiplies the places from which new knowledge can be produced. The more knowledge circulates, the more people can build on it, combine it, and expand it. The acceleration then begins to act on its own conditions. After the printing press, just two hours of our calendar will be enough for industry, computing, the internet, the smartphone, and artificial intelligence to appear.
December 31, 11:00 PM – The machine begins to manufacture machines
At eleven o'clock at night, just an hour later in our calendar, the steam engine converts energy into regular mechanical work. In 1712, Newcomen's engine pumps water from coal mines; that same coal later powers new machines that allow even more to be extracted. Technology begins to produce the material conditions for its own growth.
Later improvements reduce consumption and transform the piston's motion into rotation. Steam no longer drives only pumps but becomes a general source of power capable of moving looms, hammers, mills, and vehicles. Production no longer depends entirely on human or animal force or on where the wind blows or water flows. Energy can be concentrated and used continuously wherever industry needs it.
The factory brings together that energy with materials, workers, and machines within the same space. In the workshop, the tool followed the craftsman's rhythm; in the factory, work begins to follow the machine's constant rhythm. The workday, punctuality, and synchronization acquire new economic value. The clock no longer merely measures the passage of the day but also begins to organize production, divide time, and measure performance.
Industrial transformation is not confined within the factory. Steam extracts coal; coal powers engines and furnaces; iron and steel allow new tools, rails, ships, bridges, and buildings to be constructed. Railways connect mines, factories, ports, and cities, while the telegraph coordinates the movement of goods, people, and information from a distance. Production, transport, energy, and communication begin to form a single interdependent circuit.
The machines themselves also require increasingly precise parts. Lathes, drills, and milling machines allow components for other mechanisms to be manufactured with an accuracy that no longer depends entirely on individual skill. A better machine makes it possible to build the next one with greater precision. Industry thus ceases to be limited to producing objects: it produces the tools, energy, and infrastructure necessary to accelerate its own development.
December 31, 11:45 PM – The programmable machine
At 11:45 PM, about 235 years after Newcomen's machine, the computer emerges. The first electronic computers occupied entire rooms and used thousands of components to perform calculations that previously required long human processes. In 1947, the transistor appeared, smaller, more resistant, and more efficient than the valves used until then to control electrical signals. Its importance lies not only in what each unit can do, but in the quantity that can be brought together within a circuit.
Industry thus applies its mass production capacity to components capable of representing and transforming information. The integrated circuit brings together several transistors within the same piece, and miniaturization begins to reinforce itself: computers help design more complex circuits, and those circuits produce more powerful computers with which to design and manufacture the next generation.
But the decisive transformation is not just the increase in computing capacity. An industrial machine is built to execute a specific operation; the computer can perform different functions according to the instructions it receives. Programming allows these instructions to be described, stored, copied, improved, and executed again without physically reconstructing the machine.
The procedure thus ceases to be completely fixed in the form of the mechanism. To change the function of the machine, it is no longer necessary to build another: it is enough to change the program that organizes its behavior. Technology begins to operate not only on matter and energy, but also on its own procedures.
There are only fifteen minutes left until midnight. In that short stretch, the personal computer, the internet, the smartphone, and artificial intelligence will appear.
December 31, 11:50 PM – Digital information multiplies
Five minutes later, our calendar marks 11:50 PM. In 1971, the microprocessor appeared, concentrating the central functions of a computer within a single chip. Computing capacity began to leave large installations and, with the subsequent emergence of the personal computer, gradually entered offices, schools, and homes.
Until then, digitalization mainly consisted of converting information that already existed on other media—documents, records, images, or calculations—into a format that the machine could process. Something different happened with the personal computer: a growing part of intellectual work began to be born directly in digital format. Texts, calculations, designs, programs, and images no longer needed to be transformed into data afterward, because they were produced from the beginning within a machine.
The printing press had reduced the cost of reproducing content; the computer also distributes the ability to produce and transform it. A file can be searched, copied, compared, reorganized, and combined with others without leaving the digital environment. Information ceases to be only something that must be converted for the machine to process it: a growing part of it is already born in a format on which it can act from the beginning.
December 31, 11:53 PM – Digitalization goes global
At 11:53 PM, in 1991, the public launch of the World Wide Web made it possible to organize and consult information circulating on the Internet through linked pages. A file was no longer limited to the computer or the physical medium on which it was created: it could be published at one point and accessed by millions of people from anywhere in the world. Digital information could not only be easily reproduced, but also circulate continuously through a common infrastructure.
This capability generated new abundance. When documents number in the millions, merely publishing them is no longer enough: they must be found, related, and prioritized. To solve this problem, search engines emerged, which crawl the network, register available pages, analyze their links, and build indexes capable of responding to a query.
However, relationships between pages are not enough to determine which result might be most useful to a person. Search engines then began to also record what words were used, what results were chosen, which were ignored, and what happened after each search. User activity provides signals that allow information to be reorganized according to its probable relevance and progressively correct the order in which it is presented.
Platforms extend this logic. A growing part of activity ceases to occur on open pages and begins to develop within services controlled by companies. Conversing, shopping, navigating, listening to music, watching videos, working, or showing oneself produce content, but also leave records of how that content is used: who accesses it, from where, for how long, in what sequence, and with what subsequent reaction.
The network thus begins to contain two different layers of information. The first is made up of texts, images, videos, programs, and messages circulating on it. The second records the relationships between these contents and the people who use them: what they search for, what they choose, what they abandon, what they repeat, and what they associate with what. One describes what exists on the network; the other, how millions of people browse it and attribute value to it.
The decisive difference appears in how both layers are distributed. Content can be spread across millions of pages, devices, and users, while records of its use are primarily concentrated in the companies that control search engines, platforms, and data centers. These organizations do not own all internet information, but they accumulate something that no single user can reconstruct: an aggregated view of how millions of people find, relate, select, and value content. They do not control only a part of the files, but the map that connects information with the behavior of those who use it.
December 31, 11:56 PM – Daily life becomes data
Thirty-six years have passed since the appearance of the microprocessor and about sixteen since the expansion of the World Wide Web. On our calendar it is 11:56 PM. In 2007, the modern smartphone emerged: a connected computer that brings together camera, microphone, geolocation, navigation, contacts, payments, entertainment, work, and access to services within an object that permanently accompanies its user. Computing ceases to be a place one goes to and begins to become an environment within which a growing part of daily life takes place.
Digitalization had already gone through several stages. First, it was a specialized activity: institutions, companies, and administrations converted certain documents or records into information that machines could process. Then, with the personal computer, a growing part of intellectual work began to be born directly in digital format. The Internet added another dimension: searches, clicks, and choices made within certain platforms began to leave information about user behavior.
The smartphone expands this digitalization of behavior to a new scale. Recording is no longer limited to moments when a person sits in front of a computer or enters a specific service. A conversation, a photograph, a search, a purchase, a displacement, or a choice can leave a digital footprint as part of the action itself. By accompanying the user throughout the day, the device connects these activities with their time, location, sequence, and the context in which they occur. Daily life thus begins to produce not only content, but also an increasingly continuous representation of those who generate and use it.
The difference between the two forms of digitalization is decisive. The digitalization of culture produces texts, images, music, programs, messages, and videos. The digitalization of behavior records the relationships established around them. A photograph can be associated with a date, a place, a person, and a description. A search preserves the query, the results shown, and the choice made. A route leaves positions, schedules, and destinations. A song, a video, or a news item can be linked to attention time, abandonment, repetition, and subsequent reaction. It no longer records only what content exists, but also how it circulates, who uses it, and what they do afterward.
For almost all of human history, producing a lasting record was a minority activity. With the Internet and the smartphone, billions of people simultaneously become producers of texts, images, videos, messages, and programs, but also of information about their preferences, routes, relationships, and decisions. They do not enter platforms to build datasets, but to communicate, work, shop, navigate, entertain themselves, or show themselves. However, by doing so, they continuously generate an amount of content, associations, choices, and responses that no previous institution had been able to gather on this scale.
This production is extraordinarily distributed, but its recording is not. Content can circulate among millions of users and services, while information about behavior is primarily concentrated in the companies that control platforms, operating systems, search engines, and data centers. Millions of people produce the signals; a very small number of organizations retain the ability to gather, relate, and analyze them as a whole.
December 31, 11:57 PM – The machine learns from examples
Only five years have passed since the appearance of the modern smartphone. On our calendar it is already 11:57 PM. Until then, programs executed explicitly formulated instructions: to solve a task, someone had to describe what properties to look for and what response to produce. This method worked for problems reducible to precise rules, but it reached a limit in capacities such as recognizing a face or identifying an object, which people perform without being able to enumerate all the necessary operations to achieve it.
Machine learning proposes a different approach. Instead of providing a complete description of the capability, the system is given a large number of examples. The machine produces a response, compares it with the correct result, and modifies its parameters to reduce the error. By repeating the process over millions of cases, it finds regularities that it can apply to images it has never seen before.
In 2012, AlexNet demonstrated the scope of this procedure. The network was trained with approximately 1.2 million images from ImageNet, distributed among a thousand categories and previously classified by people. For each image, it calculates which category it might belong to, compares its answer with the correct label, and adjusts its internal connections. Through this repetition, it learns to recognize increasingly complex contours, textures, parts, and configurations without anyone having to describe them one by one.
The idea of training neural networks was not new. What changes is the available scale. Digital cameras and the Internet had produced and gathered millions of images; human classification work had converted them into usable examples; graphic processors allowed a previously unmanageable number of calculations to be performed in parallel. AlexNet appeared when the digital archive, its organization, and the capacity to process it collectively reached a sufficient threshold to produce results far superior to previous ones.
The consequence goes beyond image recognition. A sufficiently large collection of classified files can be used to generate a capacity that was not explicitly written in any of them. No single photograph explains how to recognize all similar images; that capacity arises from the relationships the system builds by comparing millions of cases.
Digital information thus ceases to be merely what a program stores, orders, or transforms by following predefined instructions. It can become a set of examples that modifies the system itself, allowing it to respond to new cases. The capacity no longer has to be completely described beforehand: it can be built from regularities found in the data.
December 31, 11:58 PM – Language becomes training
Five years after AlexNet, in 2017, our calendar marks 11:58 PM. Learning through examples had shown that a machine could acquire capabilities difficult to express through rules, but it still depended on costly human intervention: to classify images, each photograph had to be previously associated with a correct category. Applying the same procedure to a significant part of culture would have required manually labeling an unmanageable number of words, phrases, and texts.
Language also posed a different difficulty. A word does not possess a meaning that can be determined in isolation, but depends on the relationships it maintains with others. Its function changes according to the sentence in which it appears, the preceding paragraph, or a reference introduced much earlier. To process language, a machine needs to represent not only each unit, but also how some modify the meaning of others within each context.
In 2017, a new architecture designed to solve this problem emerged: the Transformer. Instead of processing text only as a succession of words, it introduces an attention mechanism that calculates which relationships are relevant to interpret each one. The text is divided into units represented by numbers, and each unit can relate to the subject of the sentence, to a pronoun written several lines before, or to another term that alters its meaning. By repeating this operation in different layers, the model builds increasingly complex representations of the grammatical, conceptual, and structural dependencies that run through the text.
Its importance does not reside solely in better representing context. The architecture allows many relationships to be calculated simultaneously and the work to be distributed among numerous processors. Compared to systems that had to process text step by step, it can be more easily scaled to larger models and much greater amounts of language. But it still remains to be solved where to obtain the correct responses needed to train them.
Digital language itself provides the solution. Unlike a photograph, which needs an added label to indicate what it contains, a text preserves within itself the sequence that must be learned. It can be shown to the model a part and asked to predict what will be next, or a unit can be hidden and it can be required to reconstruct it. The answer is already in the original document. Each text can thus become a succession of exercises without a person having to manually classify each sentence.
To reduce its errors, the model must discover which words tend to appear together, how sentences are constructed, what concepts maintain stable relationships, how an explanation develops, or what structure an argument adopts. No one introduces a complete list of these regularities. The system modifies its parameters by comparing each prediction with the actual continuation and, after repeating the process over immense amounts of text, incorporates relationships that it can apply to sequences it has never encountered.
Training does not introduce the documents into the machine as a library that can later be consulted item by item. It adjusts a structure of parameters to reproduce regularities present in the whole: associations between concepts, syntactic forms, styles, dependencies between fragments, and procedures expressed through words.
The expansion of digital information is therefore indispensable. For decades, books, articles, manuals, web pages, programs, code repositories, conversations, translations, and explanations had been produced for very different purposes. These materials had not been created to train artificial intelligence, but to research, teach, inform, sell, discuss, collaborate, or solve problems. By being available in digital format, they could be gathered and processed on a scale that no collection specifically prepared to teach a task could have achieved.
The digital archive then ceases to be only a repository that a machine can store, order, or consult. It becomes the material with which a general linguistic capacity can be built. Writing had separated knowledge from the memory of those who possessed it; printing had multiplied its copies; the Internet had connected them within a global network. The Transformer now allows processing a part of that archive and adjusting a structure capable of producing new capacity from the regularities accumulated within it.
December 31, 11:59 PM – Language becomes instruction
Only three years later, in 2020, our calendar marks 11:59 PM: one minute until midnight. The Transformer architecture had allowed increasingly larger models to be trained on enormous amounts of text. However, having a general linguistic capacity did not itself solve how to apply it to specific tasks. Translating, summarizing, classifying, or answering questions still usually required adjusting the model, adding specific components, or preparing it anew for each function.
In 2020, GPT-3 appeared, a large-scale language model built on this architecture. Its decisive contribution is to show that the same structure can infer what task it should perform from the text it receives. It does not need each function to be incorporated through new training: the instruction, examples, and expected outcome can be included within the context itself.
If it is shown several sentences alongside their translations, it can continue the pattern with a new one. If it receives a text accompanied by a summary, it can try to summarize another. If examples associated with different categories are presented, it can classify a subsequent case. The model does not modify its parameters while performing these tasks. It interprets the relationships present in the context and temporarily directs a capacity acquired during training towards them.
GPT-3, therefore, does not convert one machine into many different machines. It converts the same linguistic capacity into a basis from which different functions can be defined using language. The task no longer depends exclusively on programming or prior adjustment and can be formulated within the input received by the system itself.
Language thus acquires a new function. It had first been the material from which the model learned; now it also becomes the medium with which its behavior is directed. During the industrial stage, changing a machine's function required modifying its mechanism. With the programmable computer, it was enough to replace the program. With GPT-3, part of the function can be configured by describing in ordinary language what result is sought, what conditions it must respect, and what examples it must follow.
However, predicting a probable continuation is not yet equivalent to adequately responding to a request. The digital archive contains rigorous explanations and errors, useful instructions and contradictory fragments, cooperative conversations and aggressions. Having learned its regularities allows for coherent language production, but it does not by itself determine what kind of response should be offered for a specific intention.
To convert this general capacity into a tool, new human intervention is needed. Examples of instructions and responses are prepared, different outputs are compared, and which ones are most useful, clear, or safe are indicated. The initial training had allowed the model to learn what continuations were probable within the language; the subsequent adjustment incorporates criteria on which ones are preferable when someone formulates a request.
At this point, the second layer of information generated by digitalization reappears. Human choices, corrections, and evaluations can be transformed into examples with which to guide the model's behavior. A person compares two responses, indicates an error, or shows how a task should be solved, and this intervention provides information not solely contained in the original texts. Not every conversation is used to train the systems, nor does every user action fulfill that function, but human reactions can be used to decide how a capacity previously built from the cultural archive should be applied.
On November 30, 2022, ChatGPT was released to the public. Conversation takes the possibility opened by GPT-3 a step further. The task no longer has to be completely defined in a single instruction: it can be built progressively. The user requests a result, observes it, corrects what is not working, adds conditions, and continues from the previous response. Each message modifies the context from which the next will be produced.
The public opening also changes the scale at which this capacity is explored. Millions of people begin to use the system to write, translate, program, explain, classify, design, summarize, or reorganize information. Their requests reveal applications, errors, and limitations that developers could not have fully enumerated beforehand. The model ceases to be tested only on prepared tasks and comes into contact with real problems, diverse intentions, and unforeseen ways of formulating them.
In the following years, this logic expands to images, sound, and video, and models connect with search engines, databases, programs, and other tools. Language ceases to serve only to request a textual response. It can express an objective that the system divides into operations, executes using different resources, and revises according to the results obtained.
The public opening also adds a new source of information. The system no longer deals only with previously produced materials, but with millions of requests, reformulations, and evaluations generated during its use. Not every conversation is directly incorporated into training, but the sum of interactions allows for discovering what tasks are of interest, where the model fails, and what results people expect. Society had provided the cultural archive from which it learned; now it also begins to provide the map of its possible applications.
The arrival of midnight
As midnight approaches, the model can already translate, explain, classify, program, or produce images from the same structure. These capabilities have not arisen from a single rule or a predefined procedure, but from the joint processing of immense amounts of texts, programs, images, examples, and evaluations. What previously appeared distributed among different activities, professions, and people can now be brought together in a system capable of applying it to new situations.
But this capacity does not originate within the machine nor can it be sustained apart from the world from which it comes. Artificial intelligence depends on a society that continues to produce language, knowledge, events, problems, images, and criteria with which to interpret its results. It does not by itself generate the reality from which it learns or decide what should be considered true, useful, or valuable. It needs materials prior to it and new interventions that point out errors, formulate objectives, and define what kind of response is expected. Society is not only the context in which the system operates: it is the continuous source of what it can learn.
But society's need as a whole does not guarantee the need of each individual. To build a translation capacity, generations of speakers, writers, and translators were necessary; to apply it to a specific document, an instruction may suffice. The model needed programs written by innumerable people, but it does not need to resort to each programmer again when it produces a function. It needed images, descriptions, and criteria created by a multitude, although it can generate a new piece without summoning those who contributed the materials from which it learned.
The collective and the individual begin to separate decisively. The system can depend increasingly on society's joint production and, at the same time, depend decreasingly on any of its individual members. Society continues to be indispensable as the origin of knowledge; a concrete person may cease to be so for applying a part of that knowledge.
The production that feeds the models is extraordinarily distributed. Universities, administrations, media, publishers, companies, technical communities, and millions of users write, translate, program, photograph, classify, correct, and evaluate without being part of a common project. Each one publishes a text, answers a question, shares a program, or corrects a result to solve a specific need. Rarely do they know the relationship that this activity will ultimately have with a capacity subsequently built on a large scale.
The possibility of bringing these contributions together is not distributed in the same way. Training and operating the largest models requires enormous sets of information, specialized processors, data centers, energy, technical staff, and financial resources. Cultural and behavioral production comes from a multitude, but the infrastructure capable of converting it into a general capacity is concentrated in a very small number of organizations.
Those who control that infrastructure can transform collective production into a service whose conditions they decide: who accesses it, how much it costs, what functions it incorporates, what limitations it imposes, and towards what applications it develops. The people who produced the materials do not thereby acquire an equivalent participation in the system built from them. The contribution is distributed; the ability to organize it, process it, and decide on its use is increasingly concentrated.
The public availability of models prolongs this asymmetry. Users pose problems, discover applications, find errors, and show what results they consider acceptable. Each one tries to solve their own task, but the sum of millions of interactions reveals a general map: what uses are repeated, where the system fails, what functions are in demand, and what capabilities should be expanded. No single user can reconstruct this map; the organizations that operate the service can observe it in an aggregated manner.
Not every conversation is directly incorporated into training, but public use produces knowledge about the system itself. Society had not only generated the texts, images, and programs from which it learned. Now it also generates the situations through which it is discovered how it can be applied, where it needs correction, and towards what functions it can be developed.
The relationship thus takes a circular form. Society produces culture, knowledge, behavior, and evaluations. A small number of organizations gather part of this activity and transform it into a capacity of the model. Then they return this capacity as a service whose use produces new signals to correct it, expand it, and find more applications. The system continues to need collective participation, but each improvement can reduce the need for specific individuals in the tasks it learns to perform.
Artificial intelligence does not eliminate its dependence on society. It reorganizes it. People contribute as a multitude, often without knowing the technical destination of their activity; the organizations that gather these contributions can observe them as a whole, process them through their own infrastructure, and decide how the resulting capacity will be used. Society provides the scale; a few organizations concentrate the possibility of converting it into technical power.
Midnight does not represent the instant when the machine ceases to need humanity. It represents the moment when the collective need of humanity can be separated from the need of each concrete human being. For almost the entire year, technology expanded what people could do. In the last minute, it began to reconstruct, within a concentrated infrastructure, capacities that were previously distributed among millions of them.
The question that midnight opens is no longer just how much the machine will be able to do, but what position those who collectively produce what can then be used without them will retain.