Mikkel Freltoft Krogsholm
Menu

A mind, not a brand.

AI · mennesker · fremtidKøbenhavn · Danmark
Essay
8 min read

Schrödinger's AI

AI's capabilities make both fear and excitement reasonable. But the technology now helps tell the stories that will shape its future.

Mikkel Krogsholm står foran en lukket kasse, der kaster to uforenelige skygger.

On 1 August, OpenAI published ten new results in mathematics and theoretical computer science. Some resolved old conjectures. Others moved boundaries that had stood still for decades. The arguments were generated by an internal version of the company’s next major model, Astra, and later formalised in Lean so a computer could check whether the individual proof steps held together.

In the middle of the announcement was a number that is difficult to reconcile with our old ideas about intelligence. OpenAI estimated that the tokens the model used to find the solutions would cost roughly $2,000 at the company’s own API rates.

That is not the cost of the research as a whole. The number tells us nothing about the cost of developing the model, selecting the problems, investigating failed attempts, or having people review and communicate the results. Nor do we know how many problems Astra tried and failed to solve. Even so, something strange remains: a machine has supplied the material for advances across eight mathematical fields, and the bill for its own work can be compared to the price of a used car.

If we had imagined a computer with these capabilities in the past, I think we would also have imagined an economic explosion around it. A machine able to write, program, analyse, hold conversations, and find new mathematical paths would surely have seemed like an almost unimaginable engine of value.

The capabilities have arrived. The economic explosion is harder to spot. AI is already moving into work, education, and everyday life, but in many companies it still looks like a subscription, an experiment, or an extra field in a piece of software they were already using. There is a strange distance between what the models can do in their most impressive demonstrations and what we can currently see in companies and institutions.

Perhaps the capabilities are more fragile than the demonstrations make them appear. Perhaps society has not yet found the forms that can translate them into value. Both explanations may be true at the same time.

Fear and excitement come from the same place

Half jokingly, I have called this situation Schrödinger’s AI.

In the famous thought experiment, the cat remains in an unresolved state until the box is opened. I am using the image less as physics and more as history. We have a technology for which many radically different social futures still appear possible. AI could give more people access to knowledge and capabilities, accelerate research, and help us with problems we have not been able to solve ourselves. It could also concentrate power, replace work, manipulate people, and make control cheaper.

It is tempting to treat fear and excitement as two camps, one of which understands the technology while the other does not. But they are connected. No one lies awake at night worrying about a calculator. It is useful, but the range of things it can do is small and transparent. AI makes us both hopeful and afraid precisely because it is capable. As its capabilities grow, so does what we might gain from them — and what we might lose.

It therefore does not help to dismiss fear as ignorance or excitement as naivety. Sometimes both are reasonable. The Astra results could become the beginning of an engine for scientific discovery. At the same time, they move machines into work we have associated with the highest forms of human intellectual achievement. We are looking at the same result, but the story around it helps determine what we do next.

The ape that builds with stories

Yuval Noah Harari has often described humans as storytelling animals. We can cooperate in enormous groups because we are able to believe in things that do not exist like stones and trees, yet become real through our shared belief and actions. Money, nations, companies, laws, and human rights do not live in nature in the same way as an oak tree. Even so, they can organise the lives of millions of people.

A story does not have to be false to work this way. It gives direction to something that has not yet been assembled in physical form. A company begins as an idea that certain people, contracts, and accounts belong together. A law begins as words about a world that ought to be different tomorrow from the way it was yesterday. When enough people act according to the story, it acquires offices, courts, budgets, and consequences.

The same thing happens around technology. We do not only tell stories about it once it is finished. The stories help decide what gets built.

If AI is primarily described as an assistant, companies build products in which it suggests and a human approves. If it is described as digital labour, it is given roles, system access, and responsibility for tasks. If it is described as an existential threat, boundaries, control, and international competition become decisive. Each story pulls investment, regulation, talent, and attention in its direction.

None of them has to win outright. We can easily build the assistant, the employee, and the weapon at the same time. That is exactly why the Schrödinger image keeps fitting. AI does not have a single social consequence hidden inside the model, waiting for us to discover it. Its capabilities open a field of possibilities, and our stories help organise our movement through it.

Stories cannot perform magic

There is a temptation in this idea that I want to avoid. If stories shape society, it can sound as though we are free to choose a pleasant story and thereby get a pleasant technology. That is not how it works.

You cannot tell a calculator into becoming a threat to humanity. Nor can you tell an unreliable model into becoming a researcher simply because the story is inspiring. A technology’s actual capabilities set limits on which stories can acquire lasting force. If fear of AI looms larger than fear of other software, it is because the models already do things that seemed remote only a few years ago.

Conversely, capabilities do not arrive with a finished instruction manual for society. A mathematical discovery does not tell us who should own the model, which problems it ought to work on, or how its gains should be distributed. A robot capable of caring for a person does not decide whether it should free up a care worker’s time or replace the care worker. The capability makes both stories more realistic. It does not choose between them.

Stories, then, are neither decoration nor magic dust. They operate within the space opened by the technology. But inside that space, they help decide which applications we try to make normal, which we prohibit, and which never occur to us at all.

When the machine began telling stories too

Harari points to a difference between AI and the media that have previously transformed our culture. The printing press could reproduce a book, but it could not write a new one by itself. Radio and television could broadcast stories into millions of homes, but people still had to create them. AI can distribute, rewrite, and generate.

That does not mean the model holds political convictions of its own. It was trained on human-created material, developed by a company, guided through instructions, and set in motion by a user. There is no reason to assume that fluent text conceals an inner desire for one future or another. Yet it still participates when it explains to a student what AI will mean for her education, writes a company’s automation strategy, or helps a civil servant draft the first version of the rules governing it. The same model can formulate the advertisement that promises abundance and the catastrophe scenario that demands control. It can create a hopeful, cautious, or frightening future and adapt it to the person asking.

This is a different kind of cultural power from that of mass media. Television could show millions of people the same story. AI can tell millions of people a version shaped around their own questions, language, and concerns. The shared story may become more vivid, but also less shared.

This essay itself emerged inside that strange circuit. I began with a thought about how our fears and hopes shape our view of AI. I threw it into a conversation with an AI, received it back in another form, corrected it when it became too polished, and followed the connections onwards to Harari, Astra, and Schrödinger’s cat.

The AI does not decide what I believe, and responsibility for the essay is mine. But it would be artificial to claim that it has only been a passive typewriter. It has participated in developing a story about what AI might become. As I write about the technology as a co-narrator, the co-narrator is already sitting on the other side of the conversation.

Inside the box

We would like to know which AI future is the right one before we act. Whether we should be excited or afraid. Whether the machines will become tools, colleagues, competitors, rulers, or something for which we do not yet have a word.

We do not know. Astra’s mathematical results do not settle it. Neither does the economic explosion that has not arrived. They only tell us that capabilities and social impact do not move at the same pace, and that several outcomes may still be alive.

Schrödinger’s box is therefore not a sealed container standing in front of us, waiting for the day when we can open it and look inside. We are already in it. Every investment, rule, product decision, and everyday use changes a little of what is taking shape.

And now the technology is in here with us, helping to tell the stories.

Sources and further reading