AI: From Lab to Geopolitics

In 1950, the mathematician Alan Turing stepped around an argument that had no end. Instead of asking what thinking is, he proposed a test: could a machine's written replies be told apart from a person's? He called it the imitation game. In the same paper, he wrote about machines that might learn.⁴

×

It was a way to investigate intelligence, not proof that a machine had it. That difference would run through the next seventy-five years.⁴

A machine that passes a conversation test has shown something about how it behaves in that conversation. It has not shown that it understands the world, that it is conscious, or that it will be reliable at anything else. Turing turned a philosophical question into one that could be tested. Many later claims about AI blurred the two again.⁴

Five years later, the field was given its name. In 1955, John McCarthy and his fellow organisers proposed a summer research project at Dartmouth College, to find out whether learning and intelligence could be described precisely enough for a machine to simulate them. They called it Artificial Intelligence. The workshop met in 1956.⁵

The ambition was vast: language, abstraction, problem-solving, machines that improve themselves. The workshop gave researchers a shared project and a name to rally around. What the founders did not have was an agreed method, and the field began as a programme of questions rather than a recipe for building anything.⁵

Two broad approaches took shape. One gave a machine explicit rules for a task. The other let it learn patterns from examples. The first needed a designer who knew the answer in advance. The second needed data, and a way to learn from it.⁹⁶

In 1986, David Rumelhart, Geoffrey Hinton and Ronald Williams showed how the second approach could work. Their method, back-propagation, adjusted the connections inside a neural network a little at a time, narrowing the gap between what it produced and what it should have produced. The network found its own internal features instead of being told them.¹⁰

That is still what "training" means: behaviour shaped by repeated adjustment against examples, not facts typed in by hand. And a network that learns well from one set of examples may still stumble on the next.¹⁰

Chess made the perfect proving ground: fixed rules, a finite board and a clear winner. For decades it had stood as a symbol of human intellect, and a benchmark that engineers could aim at. In 1996, in Philadelphia, the world champion Garry Kasparov beat IBM's Deep Blue by four games to two. A year later, in New York, they played again.⁹

This time, the machine won.⁹

On 11 May 1997, Deep Blue took the rematch by 3.5 games to 2.5. It could evaluate 200 million chess positions a second, on hardware and with methods built for chess alone. It could not hold a conversation, and it did not need to.⁹¹¹

×

The next breakthrough came from the other approach. In 2012, Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton trained a deep neural network to recognise images, using 1.3 million photographs sorted into 1,000 categories. Nobody wrote rules describing each object. The network learned its own features from the examples, and graphics processors made the computing practical.⁶

It was a combination that would define the decade: a learning method, a great deal of data, and hardware able to keep up.⁶

In March 2016, DeepMind's AlphaGo played Lee Sedol, the winner of 18 world titles, at Go. To put the challenge into context: Go has 10 to the power of 170 possible board configurations, which DeepMind describes as a googol times more complex than chess. AlphaGo combined neural networks with search, and learned partly by playing. It won by four games to one, in a match watched by more than 200 million people around the world.⁷

Each of these victories was real, and each was narrow. A machine could beat the best player in the world at one game and do nothing else. The useful question about any result was always the same: what task, under what conditions, measured how?⁹⁷

In 2017, a team of Google researchers published a paper called Attention Is All You Need. It introduced the transformer, a design that lets a model weigh how each part of a sentence relates to every other part, and that could be trained far more efficiently. The paper was about translation. The design would become the basis of a new generation of language models.¹²¹³

In 2020, researchers showed what that design could do at scale. GPT-3 had 175 billion parameters, ten times more than any previous model of its kind. To put that into context, the image network of 2012 had 60 million. GPT-3 had nearly three thousand times as many.¹³⁶

Size changed what a model could do. GPT-3 could take on a task from a plain description and a few examples, without being retrained for it. Its authors also set out its weaknesses and risks.¹³

That changed how people could use software. Instead of choosing a program built for one job, a user could describe the job in ordinary language. Summarising, translating and answering questions could all happen in the same place. But the answers came from a model trained to predict text, not from a database in which every answer had been checked.¹³

Then came a finding about size itself. In 2022, researchers trained models on people's demonstrations and rankings of answers. The people testing them preferred the answers of a model with 1.3 billion parameters to those of GPT-3: a model a hundred times smaller, trained on their feedback. Scale was not the only thing that mattered. What a model was trained to do mattered too.¹⁴

On 30 November 2022, OpenAI released ChatGPT as a research preview. Anyone could type a question, get a fluent answer, and ask a follow-up. OpenAI's own announcement warned that it could write plausible-sounding answers that were wrong, and could respond differently when the same question was worded differently.¹⁵

×

That warning belongs in the story. A fluent answer reads like an informed one. They are not the same thing. Training on people's feedback made the models better at following instructions; it did not make every sentence true. Judging an assistant still meant checking its evidence, not admiring its prose.¹⁵¹⁴

Every one of these advances ran on hardware. Building a model and using it are different jobs. Training adjusts its parameters, which by now numbered in the billions; using a trained model runs them to produce each answer. Both take computing power, and the most capable systems take a great deal of it. From the 2012 image network onwards, the breakthroughs came tied to particular machines, not to algorithms floating free of them.⁶¹³¹²

×

That made chips a matter of state. On 7 October 2022, the US Bureau of Industry and Security announced controls on advanced computing chips, supercomputers and the equipment used to make semiconductors. Its stated aim was national security: limiting China's access to capabilities that could support military applications. The rules took effect in stages over that month.¹

The controls did not simply ban a product. They reached into a supply chain: components, manufacturing tools, and some of the work involved in making advanced chips. Research papers could still cross borders. The hardware they depended on was a different matter.¹

Science joined the race too. In 2021, DeepMind researchers reported that AlphaFold could predict the three-dimensional structure of proteins, often with accuracy competitive with laboratory measurements, in a demanding blind assessment. It was a major advance with a precise meaning: predictions, checked against structures measured by experiment.¹⁶

A prediction can point the way. It does not replace the experiments that establish how a protein actually works.¹⁶

The next step was to let models act. In December 2024, Anthropic set out the difference between fixed workflows and what it called agents: systems in which a model directs its own process, choosing a tool, checking the result and deciding what to do next.¹⁷

×

An agent is software, not a new kind of person. What it can do depends on what it is allowed to touch. Reading a document, changing a file and making a payment carry very different risks. Anthropic's account stressed feedback from the environment, points at which to stop, and room for people to step in.¹⁷

The US National Institute of Standards and Technology had already mapped the wider problem. Its AI Risk Management Framework, released in 2023, treats trustworthiness as something to assess across a system's whole life and in the setting where it is used. Reliability, security, transparency, privacy and harmful bias are each measured on their own terms.⁸

For any organisation, a strong result in the laboratory is only the start of the evidence. A system can be useful in one place and dangerous in another, because the cost of a mistake differs. Its operators need ways to notice problems, and to act when conditions change. Managing the risk means looking at the institution around the model, not just the model.⁸

In November 2023, governments met at Bletchley Park in England for the first AI Safety Summit. Among them were the United States and China. They signed a declaration on the opportunities and risks of advanced AI, and agreed to work together to understand them.²

×

A declaration is not a regulator. Bletchley recorded shared concern and a promise of further work. It created no authority over any AI system, and each country kept its own rules and its own industrial ambitions.²

The European Union went further. Its AI Act entered into force on 1 August 2024, with its obligations phased in over the following years. It sorts uses of AI by their level of risk, and sets duties for the general-purpose models that can serve many uses at once.³

The dates matter. A law entering into force, one of its duties applying, and the whole of it taking effect are not the same moment. A history can say when the Act was adopted and how it was designed. Whether a particular system complies is a question for the rules as they apply today.³