# A world no one has understood AI can produce verified results faster than people can understand them. What happens when we keep building without waiting for understanding? - Forfatter: Mikkel Freltoft Krogsholm - Type: Essay - Udgivet: 2026-10-09 - Opdateret: 2026-10-09 - Sprog: en - Emner: ai, knowledge, understanding, science, singularity - Kanonisk URL: https://mikkelkrogsholm.dk/en/articles/en-verden-ingen-har-forstaaet/ --- If I want a Raspberry Pi to send messages over the mobile network, I do not have to begin by learning to build a radio. A Raspberry Pi is a small computer you can use for anything from a homemade measuring device to controlling things around the house. I can buy an add-on circuit board with a mobile communications module, connect it to the computer, give it a SIM card and follow the instructions. My little project can then get online without a conventional internet connection. I need to know something about power, connections and the commands that make the module transmit. But I do not need to understand every calculation and electrical circuit that makes it possible to communicate with a mobile mast. If it does not work, I check the connection, the antenna, the power supply and the settings. I do not start by doing a PhD in radio waves. This is a perfectly ordinary way to build things. I can use something because other people have done the complicated work and packaged it into a part I can connect to other parts. Yet there is an assumption beneath my trust: the understanding exists somewhere. There are people who know how the module is designed. Others understand the rules of the mobile network. None of them needs to understand my entire project, and none needs to hold every layer in their head. But if we trace the connections far enough back, we arrive at human expertise. What happens if one day we trace the same path back and find no one? ## We already build without understanding everything Our society works because we do not all have to start from scratch. Someone building a house does not have to have developed the insulation themselves. Someone using a spreadsheet does not have to know how the computer represents numbers in its circuits. Even a skilled surgeon cannot necessarily explain every component of the device monitoring a patient during an operation. This is not a failing of civilisation. It is one of its greatest achievements. We have turned specialised expertise into something other people can use. Understanding is distributed across people, professions and generations. Some of it is in books and drawings. Some is in the experience of the technician who can hear that a machine sounds wrong. We lean on far more than any of us knows individually. So it is not in itself a rupture when AI produces something I do not understand. Almost anyone with a different speciality from mine can do that too. The difference becomes interesting if AI creates a result that no human has yet studied, and that result nevertheless becomes a useful building block. I can still connect it to my project. It can still pass the relevant tests. But the understanding I used to assume existed in other people has not yet come into being. We are used to delegating understanding. We may be moving towards doing without it in parts of what we build as well. ## From reading the code to testing it When AI writes a program, you can read the instructions line by line. You can also run the program and examine what it does. Does the right price appear? Does the right user get access? What happens if the internet connection drops in the middle of a payment? These two forms of checking reveal different things. A program passing a test does not guarantee that it is correct in every situation. But that does not mean a human must read every instruction to use it responsibly. We already work this way with large amounts of software written by other people. AI makes the distinction more visible. When the machine can write more code than we have time to read, it becomes tempting to start with behaviour and only look into the details when something requires it. In [*When code is no longer the human language*](https://mikkelkrogsholm.dk/da/articles/naar-kode-ikke-laengere-er-menneskets-sprog/), I explored how the machine's work can be translated back into something people can understand. Code needs a human way in, even if we do not read all of it. But even a good explanation takes time to understand. If ten changes become ten thousand, the machine may be excellent at explaining every one of them without people being able to keep up. That limitation is not about the code being impossible to read. It is about there being too much of it. It is worth taking that thought out of software and into science. ## 719 manuscripts and a queue of people On 6 October, OpenAI published a [repository of mathematical manuscripts](https://github.com/openai/math) — a public collection of documents and accompanying material. The catalogue lists 719 manuscripts across 372 groups of related results, produced by an internal AI model that has not yet been released. That is not the same as 719 fully verified breakthroughs. The results are at different stages of verification. OpenAI has already [withdrawn three manuscripts](https://github.com/openai/math/blob/main/history.md) because a sign error invalidated an argument on which other results depended. The company also reports that 300 of the 719 main results have been formalised in Lean. This is a tool in which mathematical statements and proofs are written so precisely that a computer can check whether the conclusion follows from the definitions and assumptions. That is different from trying a program and seeing whether it works in a few selected situations. A formal proof can establish a statement for all the cases it covers. Mathematicians still have to examine whether the precise statement corresponds to the problem claimed to have been solved, and what the result means. Formalisation is therefore a genuine gain. We should not make humans' slow reading the only acceptable route to assurance if we can build stronger checks. But we can check the validity of an argument without thereby having the understanding that enables us to explain it, see its possibilities or find a simpler route to the same result. A proof can be ready. The work of turning it into human insight may still be waiting. In [*AI finds the connections we overlook*](https://mikkelkrogsholm.dk/da/articles/ai-finder-de-forbindelser-vi-overser/), I described a fruitful cycle: AI finds a new route, people understand and improve it, and AI continues from that new starting point. But that cycle assumes people can understand before the work moves on. ## Even if all of it were right It is easy to make AI's production of knowledge a question of errors. The machine writes a great deal, some of it is wrong, and people are kept busy sorting it out. That is an important problem. But imagine for a moment that we become much better at checking. That many of the results really do hold. That machines can check more of their own and one another's arguments, and experiments confirm new predictions. Human understanding would still be a possible bottleneck. A researcher may be perfectly capable of understanding a new proof and spend three months doing so. If a thousand new results appear in her field during those same three months, her ability to understand the individual result is not enough to keep pace with developments. More researchers can share the work. AI can help with explanations, examples and shorter proofs. All of that can make an enormous difference. There is simply no law of nature saying that human understanding will grow as fast as machines' production and verification. In [*Knowledge explosion*](https://mikkelkrogsholm.dk/da/articles/videnseksplosion/), I wrote about the difference between generating hypotheses and having them tested. But there is another distinction: a result can have been verified without people having made it part of their own understanding. If AI can also use the result as the starting point for the next discovery, a chain of research can emerge in which the human work of understanding is always catching up. We do not merely get more papers to read. We get new things to read that already build on other new things we have not yet had time to understand either. That is a different kind of distance from ordinary specialisation. Specialisation divides the work among people. Here, some of the work can move on before any human has taken on the understanding. ## Understanding is more than approval There is good reason to welcome that possibility. We could get materials that make batteries better, computational methods that make research cheaper, or treatments that help patients. A missing link in human understanding does not in itself make a result worthless. If a treatment has been thoroughly tested and helps, it is not the patient's job to wait for a complete explanation of all its mechanisms. But a mobile communications module must do more than send a single message on my desk. It must also be usable under particular conditions. How much power does it need? Which networks can it communicate with? What might cause the connection to fail? Understanding helps us choose the right tests and notice when we are moving beyond the area in which we have reason to trust the result. It also helps when we want to change something. A circuit board that works is one thing. Being able to repair it or build a better version requires more. AI may be able to help us with all those tasks. It may also become the only thing that can follow the connections between them. We could then gain more capacity to act and less independence at the same time: more things become possible, but we become more dependent on the machine to know what we can change. The independent [advisory group AGMAI](https://agmai.org/general-sep29/) makes precisely this distinction between AI results that people understand and results that no one has yet understood. The group recommends that laboratories support the work of developing human understanding, including financially, and that this work be led by scholarly communities rather than the laboratories themselves. OpenAI has [announced support for workshops, conferences and programmes](https://openai.com/index/sharing-ai-progress-in-mathematics/) intended to help people understand the results. People still need to gather around the arguments, then. Verification and understanding are related tasks, but they are not the same. ## The singularity on a little circuit board The singularity is often imagined as a dramatic moment: the machine becomes so intelligent that humans can no longer follow its thoughts, and the world beyond that point becomes unpredictable to us. I think there may also be a more everyday route there. It does not require the individual result to become incomprehensible. We can still take a proof out of the collection and study it. Just as we can open a program and read some of the code. But meanwhile, work continues elsewhere, and understanding plays an ever smaller role as the entry ticket to the next step. We test, connect and build on it. It works. We need it. More arrives. The mathematics repository does not show that we have already crossed such a boundary. But it makes it possible to ask a more precise question than when AI will become smarter than us: when does the world begin to change through results that people have not yet understood, and that we do not wait to understand? If I connect a mobile communications module to my Raspberry Pi, I do not need to be able to explain the radio engineering myself. That is a freedom I have gained through other people's work. One day I may hold a part just as small, connect it and see the connection come up. It has been tested. It does what I need. I can build on it. But if I trace its design back, I may arrive at a result that no human has yet understood. And the next version of the module may already be on its way.