Oxford Paper: ChatGPT Would Have Told the Wright Brothers Not to Fly

A summary of Felin and Holweg’s “Theory Is All You Need”: LLMs predict from past data, while human breakthroughs start when someone refuses the consensus the data encodes.

Oxford researchers Teppo Felin and Matthias Holweg argue that large language models can only mirror the past, and that every real breakthrough begins when a human decides the existing data is wrong.

A paper published in Strategy Science has gone viral after a detailed thread by AI curator Alex Veremeyenko argued that large language models are mathematically incapable of genuine invention.

The paper, Theory Is All You Need: AI, Human Cognition, and Causal Reasoning, flips the title of the landmark transformer paper Attention Is All You Need. Its claim is blunt: AI predicts from the past. Humans reason forward into the future. Those are two different kinds of thinking.

Felin, of Utah State University and Oxford, and Holweg, of Oxford’s Saïd Business School, are not arguing that AI is useless. They say models will dominate routine decisions that extrapolate from what has already happened, which is most decisions. What they reject is the idea, associated with Daniel Kahneman, that humans should be replaced by algorithms whenever possible.

A child vs. 13 trillion tokens

The authors start with scale. A large language model is estimated to train on roughly 13 trillion tokens. A human reading 150 words a minute would need about 164,000 years to consume that much text.

A child hears around 20,000 words a day, about 36.5 million words in the first five years, and still ends up with language that goes far beyond anything they heard.

The model, they argue, learns which words tend to follow other words. It becomes a mirror of what people have already written. It does not build a theory of how the world works, so it cannot step outside its training data.

Galileo would have lost the vote

The paper’s sharpest thought experiment is historical.

Imagine an LLM trained in 1633 on every scientific text written up to that point. Ask it about Galileo and heliocentrism. Thousands of years of geocentric writing would swamp Galileo’s ideas. The model would tell you he is wrong. It would also rate Tycho Brahe’s astrology as more credible than the claim that the Earth moves — because more people had written about astrology.

Flight, nine weeks after “never”

Then there is heavier-than-air flight, the example the authors use to name what they call the data–belief asymmetry.

In 1888, scientist Joseph LeConte looked at bird data, noted that no bird above 50 pounds could fly, and concluded humans could not either. Lord Kelvin, then president of the Royal Society, said he had not “the smallest molecule of faith in aerial navigation.” The New York Times estimated in 1903 that flight was one to ten million years away.

Nine weeks later, the Wright brothers flew.

They did not have better data. They had a theory. They broke flight into three problems, lift, propulsion, and steering, built their own wind tunnels, and generated the data that did not exist yet.

Wilbur Wright wrote in 1900 that he had been “afflicted with the belief that flight is possible to man.”

Every prediction machine on Earth would have told him no.

Why a model cannot start there

The authors’ point is design, not attitude. A system trained to minimize surprise cannot, by design, begin with a belief the existing record says is false. Breakthroughs start with that belief.

AI, in their account, is backward-looking and imitative. Human cognition, at the moments that matter, is forward-looking and causal: a theory that tells you what new evidence to go looking for.