Are We in a Race with AI? The Question of the Track

In 2018 OpenAI reported that the compute used in the largest AI training runs had been doubling about every 3.4 months since 2012, an increase of more than 300,000 times between AlexNet and AlphaGo Zero. Had Moore's law kept its usual pace, doubling every two years, the rise would have been only about sevenfold. Figures like these invite the story of a race: a hare bolts and leaves us behind. But a race presumes one track and one finish line, and that is what needs examining.

Two speeds, two yardsticks

The figures above measure the inputs of training: how much compute was spent. They do not measure understanding, wisdom or usefulness. We should not measure the machine with a ruler of distance and the human with a ruler of depth and then announce a winner. In his 1988 book Mind Children, Hans Moravec observed that it is comparatively easy to make a computer perform at adult level on intelligence tests or at checkers, but hard to give it the perception and mobility of a one-year-old. He saw abstract thought as "the thinnest veneer" of human thinking, resting on far older sensorimotor knowledge that natural selection had refined for much longer. That is Moravec's paradox: what is hard and easy for machines runs against our intuition.

Superiority depends on the kind of task

Fabrizio Dell'Acqua and colleagues from Harvard Business School ran a randomised field experiment with Boston Consulting Group involving 758 knowledge workers. Inside the "frontier" that the models handle well (18 tasks), users of GPT-4 completed 12.2% more tasks, 25.1% faster, at higher quality. On a complex managerial task outside that frontier, they were about 19 percentage points less likely to reach the correct answer than those without AI. The tool lifts those who know its limits and can mislead those who do not. The experiment is treated further in the article "Three Myths About Artificial Intelligence: Fluency, General Intelligence and Neutrality"; it is a single experiment with a 2023 model and should not be generalised without caution.

Does the acceleration have a ceiling?

Pablo Villalobos and colleagues estimate that the size of training datasets may approach the whole stock of public human text between 2026 and 2032 if current trends continue. This depends on assumptions, and they mention routes that might ease the effect, such as synthetic data and more efficient use of data, though none is guaranteed.

Brenden Lake and colleagues argue that deep networks differ from human intelligence: people build causal models of the world, start from intuitive theories of physics and psychology, and acquire and generalise knowledge quickly. This does not mean machines will never catch up, only that comparing the two on a single scale is imprecise.

The race that matters

The World Economic Forum's The Future of Jobs Report 2025, based on a survey of more than a thousand employers, estimates that about 39% of existing skills will be transformed or become outdated between 2025 and 2030, and that 59 of every 100 workers will need training by 2030. The report does not say the machine will outrun the human; it says the skills in demand are changing. The real contest, on the likeliest reading, is between those who keep learning and those who stop. That is our reading, not the report's wording.

What should we do?

• Make continuous learning your parent skill. Whoever has mastered how to learn does not fear that what they learned will date.
• Divide tasks deliberately. Hand the tool the repetitive work it does well by evidence, and keep moral judgement, grasp of context and responsibility.
• Test the tool on your own task. Success in a general trial does not guarantee success in yours.
• Protect depth. A tool's summary is no substitute for building firm understanding.

Perhaps the old tale teaches it best: the hare did not lose for being fast, but for assuming speed was enough. The machine accelerates where we point it, and pointing is a trust; whoever knows their destination, takes the means and relies on God is well placed to reach it, however slowly it seems.