8.64 seconds: the winner we are laughing at

On Wednesday, in Beijing's speed skating oval, a robot ran the hundred metres in 8.64 seconds. Usain Bolt needed 9.58 in Berlin in 2009. A few strides past the finish line the machine went into a blue crash mat at full speed, sparks came out of its torso, and stewards ran up with fire extinguishers and hosed down the robot where it lay. It was then carried out on a stretcher, as was almost every finalist in that race.

The second half of that scene has been circulating on social media ever since. The first half, those 8.64 seconds, has stayed largely unknown there, although it sits in the same clip and belongs to the same machine. The robot everyone is laughing at won the race.

I spent this weekend reading Rüdiger Safranski's “Die Vierte Kränkung”, a title that translates as the fourth blow to human self-regard. It appeared four days before these games opened. Safranski's subject is the humiliation we have to absorb after Copernicus, after Charles Darwin and after Sigmund Freud, and this one strikes the last ground on which we still held ourselves unique: our own mind.

There is less Schadenfreude in all that audible laughter at a toppling robot than there is quiet relief.

The fall became a genre

What ran through social media in the days that followed amounted to a genre of its own. Gizmodo assembled the thirteen finest malfunctions, from the weightlifter who tips into the judges' table, through the high jumper whose head comes off, to the dancing robot the commentators diagnosed as drunk. An Australian broadcaster cut a segment from it called “Best AI Robot Fails”. The Mirror collected “hilarious robot fails”, Kotaku gave a headline to a machine's burning backside, and the Times of India had a robot die at the barrier, in quotation marks at least. Instagram files the material under “Bot bloopers”.

For the 8.64 seconds, a wire report on Tagesschau was considered enough.

The fall was cut, sorted, ranked, captioned and scored. The time was copied out. Human care did go into this material, and it went into the seconds in which the machine lies on the ground.

That can fairly be put down to a healthy reflex, an immune response to the steady bombardment of superlatives from Silicon Valley and Shenzhen. Anyone who has been told for three years that every new generation of models changes everything is entitled to a certain pleasure at the sight of a humanoid staggering into a foam wall. Scepticism towards staged technology is standard equipment for an educated audience.

But scepticism is a stance that asks questions. What was on display here asked none. Nobody wanted to know why the machine does not brake. The answer was available and would have spoiled the joke. Lightning, the runner-up, built by Honor, still ran the Beijing half marathon in April on legs ninety-five centimetres long. For the sprint it was given legs of one metre five, ten centimetres longer, as Chinese state television reported. That improves acceleration and degrades control under braking. Somebody weighed the two against each other and chose acceleration, in the full knowledge that a mat would have to be waiting at the end of the track.

A prototype is what a finding costs

Anyone who mocks the mat holds a particular view of what product development is supposed to look like: you build a device until it is finished, and then you show it. On that view, a robot that runs into a wall without braking is an unfinished device shown too early.

Beijing worked to the other idea. What stood on that track was a hypothesis rather than a product: how fast can a two-legged body with this actuation get before the control loop loses balance. Answering that question needs no braking mechanism. It needs a hundred metres of track, a timing system and the willingness to take the result even when the machine ends up on a stretcher. What looks like failure at the finish is a function left out from the start, and therefore a textbook Minimum Viable Product (MVP).

The cost of that attitude is worth discussing, and it turns out to be smaller than the pictures suggest. Bank of America puts the bill of materials for a typical humanoid at around $35,000, with a good half of that in the actuators. Unitree lists its G1 at $13,500 and the H2 at $29,900. A wrecked prototype therefore costs about as much as an entry-level company car, which no German firm regards as a daring investment.

Lorenzo Masia of the Technical University of Munich has described what events like this are good for, writing about their close relative the robot half marathon: they force developers to deal with battery performance, the durability of motor drive units, real-time control and fault tolerance. Those are exactly the bottlenecks nobody finds in a laboratory, because nobody in a laboratory runs for forty minutes without stopping. A stress test yields knowledge in proportion to the material it destroys.

Tiangong Ultra ran 9.39 and then 8.86 seconds between its first and second outings, and 8.86 then 8.64 between the second and the final. Three measurements in five days, in public, with witnesses. Nothing has been published about the timing method – more on that shortly – and the first of those figures was set at the opening ceremony. All of it was broadcast anyway, every fall included. In Europe this chain of findings would have happened on a test rig, behind closed doors, with a final report the following quarter. The numbers would have been the same. Nobody would have laughed, and nobody would have seen it either.

What a year changes

At the first edition of these games in August 2025, Tiangong won the 100 metres in 21.50 seconds. Twelve months later the same competition ends on 8.64. That is a factor of 2.5, over the same distance, in the same format, from the same developer. The Beijing robot half marathon, which had its second running in April, traced a similar curve: in 2025 the winner needed two hours and forty minutes and followed a human who ran ahead of it with a signalling device strapped to his back. In 2026 an autonomously running machine won in fifty minutes and twenty-six seconds. The fastest time in the field was lower still, forty-eight minutes and nineteen seconds, but it belonged to a remotely operated machine and took a penalty under the rules.

The surrounding industry grew at the same rate. 280 teams became 666, some 500 robots became 2,056, 26 disciplines became 51. How quickly the order changes here can be seen in the results of Unitree, the best known Chinese manufacturer. In 2025 the company took four gold medals on the track, among them the 1,500 metres in 6:34.40. In 2026 its robot got no further than 12.41 seconds in the heats and had won no track medal at all by 23 August. On the sprint distances the pace is now set by X-Humanoid and Honor, two firms that played no part there a year earlier.

There is an objection to this reading, and the Neue Zürcher Zeitung put it well in a comment piece on 23 August: the robot records in Beijing are above all a good show, and genuine innovation takes a different form. The objection carries weight. The 9.39 seconds on the opening day were set at an opening ceremony rather than in a race. Honor's 9.32 seconds come from a test run reported by Chinese state television alone, while the same machine ran 9.47 in the official final. No federation has certified these times, and nothing has been published about the starting procedure, the timing or the wind measurement. Robert Katzschmann and Thomas Buchner of ETH Zurich, in the same paper, pointed out how far off a machine that can be put to general use still is.

Whoever stages a show, though, decides what there is to see. What there was to see in Beijing was failure, from several cameras, in slow motion, for five days. A director interested in an image of superiority would have cut the burning winner. Anyone who holds the series to be staged has to explain why the staging needs a lower number every year than the year before, and why it lets itself be filmed failing.

Looking at these games from Germany does not mean starting from scratch. Neura Robotics of Metzingen raised up to $1.4 billion in June, the largest single financing round for a full-stack robotics company in Europe, at a valuation of around seven billion. The round was led by Tether, the largest issuer of stablecoins, in a consortium with Nvidia, Qualcomm, Amazon, Bosch, Schaeffler and the European Investment Bank. That line-up says as much about the state of European capital formation as it does about robotics. The company brought its manufacturing back from China to Germany in 2024 and intends to build 6,000 humanoid robots this year. On robot density in manufacturing, Germany ranks third in the world with 449 units per 10,000 employees. China appears in the same statistic at 166, and that figure deserves a footnote: in last year's report China stood at 470, ahead of Germany. Not a single robot was removed in the meantime. China's statistics office had simply recounted how many people in the country are in work, and the measure sets robots in proportion to precisely that number. Counted in units, China operates the largest installed base in the world at around two million industrial robots, and in 2024 it absorbed 54 per cent of the global market with 295,000 installations. That is more than the total installed in Germany.

And those 449 are measuring a different discipline in any case. An industrial robot is a very precise arm in a cage that repeats the same movement a million times over and knows exactly where the workpiece lies. What ran down the track in Beijing has left the cage. In that second discipline, industry figures put roughly 97 per cent of all humanoid robots shipped worldwide in the first half of 2026 as coming from China. The American company Figure AI was valued at $39 billion in September 2025, a multiple of what the European suppliers are worth combined. Germany leads the world in a discipline that no longer sets the standard on its own.

The fourth blow

In 1917 Sigmund Freud published a short essay that tells the history of science as a sequence of insults. It is called “Eine Schwierigkeit der Psychoanalyse”, and he opens the argument like this: “Nach dieser Einleitung möchte ich ausführen, daß der allgemeine Narzißmus, die Eigenliebe der Menschheit, bis jetzt drei schwere Kränkungen von seiten der wissenschaftlichen Forschung erfahren hat.” (After this introduction I should like to show that the universal narcissism of mankind, its self-love, has so far suffered three severe blows at the hands of scientific research; my translation.)

The first was cosmological and bears the name of Nicolaus Copernicus. It took the Earth out of the centre of the universe and made it a lump of rock in orbit. The Church took until 1992 to rehabilitate Galileo Galilei, which gives some idea of how long a blow can go on hurting. The second was biological and belongs to Charles Darwin. It ended the special standing of man in creation and put him on a par with the rest of the animals. Here too the response was mockery rather than assent: contemporary cartoons showed Darwin with the body of an ape, which had the advantage that one could laugh at his thesis instead of testing it.

Freud attributed the third blow to himself, with a matter-of-factness one has to admire. It is psychological, and it consists in the demonstration “dass das Ich nicht Herr sei in seinem eigenen Haus” (that the ego is not master in its own house; my translation). Man had lost his place in the cosmos and his rank in nature, but at least he believed he still governed his own head. Freud took that from him as well.

What was left after that was the mind, the last thing he still owned. That is exactly where Rüdiger Safranski begins. His essay puts the position in a sentence his publisher printed on the cover: because he possesses a mind, man was until recently able to believe himself something special, and the fact that artificial intelligence solves many tasks better than we do is for humanity a promise and a blow at once. Safranski takes the blow seriously and puts it in its place at the same time. What the machine lacks, he argues, is what makes a human being: consciousness, sensation, self-observation, the possibility of deciding otherwise. Speaking to the Austrian broadcaster ORF, he put it like this:

Whether that consolation holds is contested, and the objections come from sympathetic quarters. Angela Gutzeit, writing for Deutschlandfunk, praised how firmly Safranski defends the uniqueness of the creative mind, and at the same time faulted him for leaving the follow-up question open: how good old humanism is supposed to handle the matter in practice. That is the sore point. A defensive line that does not organise the retreat is an explanation, not a strategy.

It should be added that the ordinal number four was already taken, several times over. Sascha Lobo proclaimed the digital blow, from the moment the internet turned from an instrument of freedom into one of surveillance. Johannes Rohbeck names the technological blow, Reiner Klingholz the ecological one, and Gerhard Vollmer, totting the whole account up, arrives at ten. Humanity must by now be so thoroughly insulted that it can no longer keep up with the numbering.

Anyone with that series in mind sees the videos from Beijing from a different angle. A blow produces no discussion, it produces defence, and defence looks for relief. The relief lay in the second half of the video, where the machine goes into the wall, burns and is carried away. What people were looking at there was the most reassuring image of the week: the thing that is faster than the fastest human cannot stop.

Envy is the most interesting feeling at this point, because it has nowhere to land. It finds no counterpart that would notice it. The machine takes no pleasure in its 8.64 seconds; it has no inside, to use Safranski's word, for someone else's feeling to arrive in. Envy therefore falls back on the person feeling it, and it is not aimed at the device. It is aimed at the future, of which we were allowed a glimpse, represented by the robots. And a future can be begrudged its fall without any fear that it will therefore fail to arrive.

The second half builds a body

To see why a sports event with falling machines amounts to more than a curiosity, you have to set it beside what has happened in knowledge work over the past three years. There the first half is largely over. Language models read contracts, draft credit papers, summarise supervisory letters, write code and answer customer queries. The International Monetary Fund measured this early: Kristalina Georgieva wrote in January 2024 that artificial intelligence would touch nearly 40 per cent of jobs worldwide, and around 60 per cent in the advanced economies, because that is where the share of cognitive work is highest. The number does not mean those jobs disappear. It means the work changes, and on the Fund's methodology about half of the jobs touched will be affected negatively, through displacement, wage pressure or redrawn responsibilities.

Erik Brynjolfsson at the Digital Economy Lab in Stanford has spent two years measuring what that looks like in practice, using American payroll records. Among employees aged 22 to 25 in heavily exposed occupations, meaning software development, marketing and customer service, employment has fallen relative to older colleagues in the same occupations. That shortfall against the older comparison group grew from 13 per cent to 15, then to 16, and stood at 19 in June 2026. The mechanism is worth noting: nobody is being dismissed, people are simply not being hired. A reduction that hurts nobody visible, because those affected never got inside the company in the first place.

The authors themselves stress that no economy-wide collapse can be read into this: “There is still no sign of economy-wide job destruction.” That is honest, and there is a second number that cuts the other way: the World Economic Forum expects 170 million new jobs worldwide by 2030 against 92 million displaced, a net gain of 78 million. That balance says something only about the total, and not about the entry-level position at issue here. The ifo Institute reported in August 2026 that German companies expect wages to fall rather than rise as a result of using artificial intelligence. Anyone who wants to talk about displacement need not wait for waves of redundancies. Watching the wage curve is enough.

Up to this point all of it concerns office chairs. What is being built in Beijing is the connection to everything else. A language model can describe a pipeline, assess a weld and write a repair manual, and it cannot hold the spanner. Between the model and the hand sits a problem the field calls embodied AI, and it has turned out to be the hard one. A text knows nothing of gravity. A body does.

That problem is exactly what gets worked on along a running track. Balance at just under forty-two kilometres an hour, control in real time, actuators that survive thousands of load cycles, batteries that last forty minutes, and a control system that learns from a fall instead of breaking on it. These are the same bottlenecks that have to be cleared before a machine walks into a workshop, sorts a warehouse or helps an elderly person out of bed. Jonathan Hurst, professor of robotics in Oregon and co-founder of the humanoid manufacturer Agility Robotics, put it like this at the close of the games: real progress will show up in warehouses and factories, where humanoids work independently for hours and generate actual economic value.

He is right, and it is a question of years rather than decades. Bank of America expects annual shipments of humanoid robots to rise from around 20,000 units in 2025 to ten million in 2035. Forecasts of that kind should be read with care. Morgan Stanley arrives at thirteen million units for the same year, Goldman Sachs at 1.4 million in an estimate from early 2024. Between the lowest and the highest figure lies a factor of seven. Goldman Sachs had previously multiplied its own forecast sixfold within a year, after manufacturing costs fell faster than expected. They agree on the direction and on not a single number.

Two objections stand against that direction, and both carry weight. The first is demographic: Germany will be short of millions of working people over the coming decade, and here automation is being sought rather than feared. The second is commercial. Today's humanoid robots last one to five hours per charge, carry roughly nine to fourteen kilograms and need an intervention after a few hundred operating hours, while an industrial arm in a cage runs for fifty thousand. None of those numbers carries the replacement of human labour. They carry something else: the beginning of a curve whose gradient a running track in Beijing has been showing us for a year.

The question nobody answers

Suppose all of that comes to pass. The models take over knowledge work, the machines learn to grip, and in five or ten years there is a device in the workshop, the warehouse and the care home that works more cheaply than the human being beside it. Then a question arises that has nothing to do with technology any more: what is that human being to live on? It repays reading the answers given by the people who build these devices.

Elon Musk gave his in May 2024 at the VivaTech technology fair in Paris. “In a benign scenario, probably none of us will have a job”, he said there, and went on: “If you want to do a job that's kinda like a hobby, you can do a job. But otherwise, AI and the robots will provide any goods and services that you want. There will be universal high income, not universal basic income, universal high income. There'll be no shortage of goods or services.” He put the probability of this benign scenario at about eighty per cent, and in the same breath he asked the follow-up question: “If the computer and robots can do everything better than you, does your life have meaning?”

In November 2025 he got more specific on a podcast: “I think long term, I think money disappears as a concept. Honestly, it's kind of strange. But in a future where anyone can have anything, I think you no longer need money as a database for labor allocation.” It is a pretty formulation, money as a database for the allocation of labour, and it comes from a man who on 24 July 2026 posted “(Former) Trillionaire” on his own platform, after his fortune had fallen by some $750 billion in six weeks. Three days later the same platform launched its payment product X Money.

Sam Altman has been setting out how it might work since 2021. In his essay “Moore's Law for Everything” he proposed an American equity fund that taxes companies and land rather than labour and distributes the proceeds directly to citizens, by his arithmetic $13,500 per adult per year. In 2024 he mused aloud about a variant he called Universal Basic Compute: instead of money, everyone would receive a share of the computing power of the best model of the day, to use, resell or donate to cancer research. You would then own, he said, a share of the productivity itself.

Dario Amodei has been the most direct of them. He runs Anthropic and is therefore one of the suppliers whose models make the shift described here possible in the first place. In May 2025 he told the news site Axios that artificial intelligence could eliminate half of all entry-level white-collar positions and drive unemployment to ten or twenty per cent within one to five years. His reason for saying so publicly: “We, as the producers of this technology, have a duty and an obligation to be honest about what is coming.” Fifteen months later American unemployment stood at 4.1 per cent. So far the forecast has no interim reading that speaks for it.

These statements have something in common that is easy to miss on a first reading. There is no shortage of proposals; Sam Altman's fund is costed down to the tax base. What is missing is anyone to take them up. A basic income presupposes a body that decides it, funds it and pays it out. An equity fund needs a tax law. A share of computing power needs an office to allocate it. In every developed state that body is called parliament, and none of these proposals has reached one.

In the parliaments themselves the question has been dispatched for years, in both senses of the word. The European regulation on artificial intelligence, the most ambitious body of rules in the world on this subject, sorts applications by risk class, prohibits some and subjects others to transparency and oversight requirements. On the distribution of the proceeds it contains nothing. Not a word, not a clause, not a mandate to review.

The debate had once got further than this. On 16 February 2017 the European Parliament voted on a report by the Luxembourgish member Mady Delvaux on civil law rules on robotics. The draft contained two proposals that sound today as though they came from another age: an unconditional basic income and a tax on robot labour. Both passages were taken out in the vote. What remained was a call on the Commission to hold a debate on new employment models and on the sustainability of tax and social security systems. Mady Delvaux said afterwards that she was disappointed. Nine years on, no result of that debate has come to my attention.

The idea itself is older and did not come from social policy. Bill Gates put it this way in an interview in 2017: if a human being does $50,000 worth of work in a factory, that income is taxed, through income tax and social contributions. If a robot comes along and does the same, it should be taxed at a comparable level. South Korea is generally regarded as the first country to have responded to automation through the tax system, cutting the depreciation incentives for investment in automation, which is widely taken to be a robot tax and is not one.

In Germany, artificial intelligence appears in the current coalition agreement as a question of competitiveness and growth, with a digital ministry of its own and investment in computing capacity. A commission on reform of the welfare state has been appointed. It is dealing with cutting red tape, unifying definitions of income and merging benefits. In the publicly available documents I have found no connection between the two undertakings, that is between the technology being funded and the social system being rebuilt.

Here is my personal assessment of what was on show in Beijing, and of what is not being said in Berlin and Brussels. Perhaps none of it comes to pass. Perhaps robotics stays in demonstration mode for thirty years, perhaps the marriage of language model and body turns out to be the fusion reactor of computer science, forever twenty years away. I would hold that to be the friendliest of the possibilities, if not necessarily the likeliest…

The fifth blow

Before that, it is worth looking at what is happening here and now. According to the ifo Institute's business survey of May 2026, 54.5 per cent of German companies use artificial intelligence in business processes, against 40.9 the year before. That sounds like a breakthrough and mostly describes spread. As to the effect, a much-quoted, non-peer-reviewed study by the MIT initiative NANDA measured it last year: in 95 per cent of the corporate generative-AI projects examined, no effect on the profit and loss account could be demonstrated within about six months of going live. The authors expressly attribute that to the way these companies go about introducing the technology, and not to weak technology.

Taking that finding seriously leads to uncomfortable consequences: giving a pilot project a date on which it either goes into operation or is stopped; asking what it contributes to results before the licence is renewed, and living with the answer; touching the processes rather than the tools, because a machine that speeds up an unsuitable workflow does nothing more than that; and putting the employees whose work is changing into the same rooms where the change is decided, instead of informing them of the outcome afterwards.

The second addressee is politics, and there the task is larger by a wide margin. The distribution of the proceeds of a technology that replaces work is currently discussed on podcasts, at technology fairs and in articles and blog posts by company founders. Those are places where thinking can happen and unfortunately nothing can be decided. A parliament can do both. The European one made a brief start in 2017 with the unconditional basic income and the robot tax, and then voted both of them out again.

The list of tasks is known, it is merely untouched. To work out how a tax system functions when its most important base for decades was labour as a factor, once that factor partly disappears. To work out who owns the proceeds of a productivity that arises from capital and data rather than from employment. To work out what becomes of a twenty-five-year-old who is not hired, because the entry-level position in which one used to learn the trade no longer exists. These questions need no finished answer. They need a place where somebody is responsible for producing one.

That leaves the outlook I would rather not write, and which amounts to a derivation, not a finding. Nicolaus Copernicus took the centre from man, Charles Darwin his special standing, Sigmund Freud the command of his own head, and Rüdiger Safranski is describing right now how the mind is going the same way, the last proof of his uniqueness. Four blows and four losses, four defensive movements, and every time man remained at the end what he had been before: the creature that asks the questions.

A fifth blow would need no scientist to pronounce it, and it would come even if everything else went well. Suppose, then, that the distribution question were answered, that the high basic income arrived, that nobody went hungry or wanted for anything. The blow would be that man keeps his special standing, carefully preserved, well provided for and entirely without consequence. A specimen behind glass, biological evidence that he once existed, with a plaque beside him setting out what this species was capable of before the devices took over. He would be treated superbly. It would be the best-reviewed zoo in history.

Preventing that does not require anyone to stop a machine, as in the well-known blockbuster “Terminator”. It is enough to give the questions a place where somebody is responsible for the answer, instead of letting them be covered over by videos of robots running into a wall and breaking. Nicolaus Copernicus was laughed at, and Charles Darwin was drawn in cartoons with the body of an ape. The laughter did not move the Earth back into the centre.

In Beijing on Wednesday a robot ran the hundred metres in 8.64 seconds and then went into a wall. The year before, the same manufacturer had needed 21.50. Laughing at the fall is allowed. It is worth knowing what one is laughing at, and how quickly it might stick in the throat next time.