The Street Is the Laboratory Now

Science used to take its time behind a closed door. With AI, the experiment is running on our roads, in our phones and in front of all of us, at the speed of a funding round.

On 21 June 1948, at the University of Manchester, the Manchester Baby became the first computer to run a program stored in its own electronic memory. Freddie Williams led the project, Tom Kilburn wrote the program and Geoff Tootill had helped build the machine. That first program found the highest factor of 2 to the power of 18, and it took 52 minutes. It’s a story I know well, being from Manchester and having studied there, and it’s tempting to tell it as a tale of scientists working patiently in a lab until the idea was ready. The history is more interesting than that.

Within four months the government had asked Ferranti to build a commercial machine to Williams’ specification, and Tootill later moved to Ferranti to help with the transfer. Industry was in the room almost from the start. What strikes me now is how porous the boundary stayed. University, government and manufacturer built the thing together, the first Ferranti Mark 1 was delivered back to the university in February 1951, and the researchers there carried on using it and pushing it further.

Now compare that with a Sunday night in Tempe, Arizona, in March 2018.

In June I wrote that the direction is clear and the world is not ready, and that almost everyone I speak to is tired from living in the gap between the two. I left one thread hanging in that piece: why the gap opened so wide, so fast. I think a large part of the answer is that the weight of the research has shifted out of the university and away from that porous, shared arrangement. When I put together a timeline of the eighty years behind today’s language models, that shift was the shape of the whole thing. The ideas came from universities, industrial research labs, non-profits and people working on things nobody was paying them for, and then, from around 2012, the last flight of the staircase happened inside companies. The work that would once have spent a decade maturing in the open is now carried out by corporations, at corporate speed, on public infrastructure, with close to limitless capital behind it. And the rest of us have become the test environment.

Forty researchers in a season

In February 2015, Uber and Carnegie Mellon announced a partnership. Uber would open an Advanced Technologies Center in Pittsburgh and work alongside the university’s National Robotics Engineering Center on self-driving cars. By the end of May, Uber had hired away 40 of the centre’s researchers and scientists. The centre lost its director, key programme leads and, according to reporting at the time, more than $10 million in research funding. That September, Carnegie Mellon announced a $5.5 million gift from Uber to fund a robotics chair and three fellowships, which reads to me like an apology with a letterhead.

Three years later, one of that programme’s test vehicles struck and killed Elaine Herzberg as she wheeled her bicycle across a road in Tempe. The National Transportation Safety Board’s final report is careful and worth reading in full. The safety driver was streaming a television programme on her phone. Herzberg was crossing away from a crossing. But the Board also found that the system detected her 5.6 seconds before impact and never classified her as a pedestrian, that the Volvo’s own emergency braking had been switched off while the car was driving itself, and that Uber’s testing division suffered from an inadequate safety culture. The cars were only in Arizona because of a regulatory row. In December 2016, California revoked the registrations of Uber’s self-driving fleet for testing without the required permit, and Arizona’s governor welcomed the cars “with open arms and wide open roads” the next day. The rules themselves had become part of the competition.

I want to be fair to the people here. The engineers who walked down the road from Carnegie Mellon did not become worse scientists the day they changed employer. What changed was the building around them: the incentives, the deadlines, the investors and the competitors all pressing to be first. Uber sold the division in 2020.

Nobody signed the consent form

Anyone who has been near medical research knows how slow it is to test something on a human being, and why. There’s an ethics committee. There’s informed consent, written down and signed. There are phased trials that start small and grow only when the evidence allows. And there’s a tradition, going back to the Declaration of Helsinki in 1964, that the person taking part has to agree to it. It’s frustrating, it’s bureaucratic, and it exists because people were hurt when it didn’t.

For the last twenty or so years I’ve been an IT architect. My life has been governance boards, documentation, and justification. Whilst in many organisations this can be described as theatre, occasionally there is a question from a committee that cuts through the technology and sends a chill to your core. In those situations, the governance works.

A public road is plainly a different thing from a clinical trial, and a pedestrian isn’t a research participant in any formal sense. I’d go further and say that’s exactly the problem. When an experimental system leaves a controlled environment for a public one, the people exposed to its failures get none of the protections we built for experiments that involve human beings on purpose.

Elaine Herzberg never agreed to take part in an experiment. Nor did the woman in San Francisco on 2 October 2023, who was hit by a human driver and thrown into the path of a Cruise robotaxi. The car braked, stopped on top of her, and then, trying to pull over out of the traffic, dragged her around 20 feet. California’s Department of Motor Vehicles suspended Cruise’s permits and said the company had not shown it the footage of that second movement; Cruise disputed that. The federal regulator later fined Cruise $1.5 million after finding that its crash reports had left out the dragging, something it only discovered once it asked for the video. In December 2024, General Motors announced it would stop funding the robotaxi business, having invested more than $10 billion in it.

Roads are the obvious case because the harm is physical and the reporting is thorough. The same pattern runs through everything else. When a language model is released to hundreds of millions of people, its makers can watch how it behaves and where it fails at a scale no laboratory could reproduce, and depending on the product and the user’s settings, some of those conversations may also feed into the next version. In that broader sense the public is part of the test environment. Nobody hands out a participant information sheet. The closest thing we get is a terms of service agreement, which nobody reads, and a small grey label saying the answers may contain mistakes.

Peer review has been replaced by the launch event.

I have a small irritation with the word “beta”, and I’ll indulge it for a paragraph. In software it used to mean a version given to people who had chosen to try something unfinished. Somewhere along the way it became a disclaimer you could attach to a product used by the whole world, which lets a company ship first and apologise in the release notes. That is a very long way from Kilburn’s team debugging their machine in a university building, with nobody’s safety on the line except their own patience.

The other thing the old arrangement gave us was scrutiny. Work was published and other people tried to break it. Claims that couldn’t survive that tended to die quietly in the journals rather than loudly on the high street. The 2026 Stanford AI Index found that industry produced over 90% of notable AI models in 2025. Commentary on the report notes that basic facts about several of the largest systems, such as their training data and parameter counts, have stopped being disclosed. Peer review has been replaced by the launch event. You find out what a system can do, and what it gets wrong, at roughly the same moment as everyone else.

Add almost unlimited money to that, and you get something the history of science has rarely had to deal with: experiments that are too big to run quietly and too commercially important to slow down.

What the universities lost

The Carnegie Mellon raid looks less like an outlier every year. In 2023, Nur Ahmed, Muntasir Wahed and Neil Thompson published a paper in Science on exactly this. The share of new AI PhDs in North America going into industry rose from 21% in 2004 to around 70% by 2020. Industry, they argued, now controls the three ingredients modern AI research needs: computing power, large datasets and skilled people.

In March 2026, economists Ufuk Akcigit, Craig Chikis, Emin Dinlersoz and Nathan Goldschlag went further, using employment records for 42,000 AI researchers. They found that among the top 1% of publishing AI scientists, the annual pay advantage of working in industry had grown fivefold since 2001, to around $1.5 million. When a researcher leaves a university for a large firm, their paper output falls by 65% and their patenting rises by 530%, as the authors explain in a more readable summary. That last figure is the one I keep coming back to. The knowledge still gets produced; it just stops being shared.

That leaves universities training people for jobs they can’t match on salary, with far less computing power than the firms that hire them away.

Industry paid too

You could read all of this as a story where universities lose and corporations win. I don’t think that’s right, and the evidence is in who keeps walking out.

Geoffrey Hinton spent most of his career as a professor in Toronto before his small spin-out company was bought by Google. In May 2023 he left Google so he could speak about the risks of AI “without considering how this impacts Google”. He was clear that Google had behaved responsibly. The point is that one of the most qualified people on earth to warn us decided he needed to leave his employer before doing it. A year later Jan Leike, who co-led OpenAI’s team working on the safety of future systems, resigned and wrote that “safety culture and processes have taken a backseat to shiny products”. The team was dissolved within days.

Robotics has its own version. In November 2025, Figure AI’s former head of product safety, Robert Gruendel, sued the company, alleging he was fired after warning executives that its humanoid robots were strong enough to fracture a human skull. Figure denies it and says he was dismissed for poor performance, and a court will decide. What caught my eye was a detail in the complaint: it sets his warnings against company values that included, as quoted in the filing, “Move Fast & Be Technically Fearless”. It’s a reasonable enough motto for a start-up writing software, and an unsettling one for a company building a machine meant to stand next to a person in their kitchen.

And the money that makes all this possible is not as stable as it looks. In June 2025, Meta’s new superintelligence lab went after researchers from rival labs with packages Sam Altman described as $100 million offers, a figure Meta said misrepresented how the deals were structured. By October, Meta was cutting around 600 roles elsewhere in its reorganised AI operation, including in its long-standing fundamental research group, while the new lab building its next models was spared. GM’s $10 billion on Cruise ended in closure. Uber sold the division it built from Carnegie Mellon’s people. Industry has spent enormous sums buying the scientists and then, too often, lost the thing that made them valuable: the freedom to say “not yet”.

Where my argument is thinnest

I should be honest about the weak spot. Universities were never quite the calm, balanced places I’ve been describing. Academic science has its own history of harm, of careers built on results that didn’t replicate, and of slowness that had more to do with committees than caution. The Manchester Baby itself grew out of wartime radar work, funded by a government in a hurry. And Stanford’s own figures show that AI models built specifically for science mostly come from collaborations between universities and companies, which is closer to the Ferranti model than to the Uber one.

So I’m left with a question I can’t answer. If AI had stayed in that shared, porous arrangement for another fifteen years, would the world be better prepared for it, or would we simply be having this conversation in 2040 instead? I lean towards the first. I don’t know it.

What I’d want back

I use AI every day and have no wish to put it back in the box. What I’d like back are the habits science built up over a few hundred years of universities, habits we discarded because they were slow. Consent before the public becomes the test. Independent people allowed to look inside the systems and publish what they find. Researchers who can raise a concern without having to resign first. And public investment in university computing, so that the people asking careful questions have something to ask them with.

The world is not ready, and I stand by that. Part of the reason is that we took the experiment out of the laboratory before it was finished, and we are all now standing in the road.

A second piece looks at the other end of the university: what happens to research when the teaching business that pays for it comes under strain, and what a degree is worth when the jobs at the far end of it are disappearing.