In October 1903, a New York Times editorial put a working flying machine somewhere between one million and ten million years away. About two months later, two bicycle mechanics from Dayton flew one at Kitty Hawk. As forecasts go, that one aged about as well as a vendor roadmap.
Teppo Felin and Matthias Holweg build their paper, "Theory Is All You Need," around that gap. It appeared in Strategy Science in 2024. Their argument is that a language model learns by predicting what comes next in existing text, so its beliefs follow the data it was trained on. A person can hold a belief the data contradicts. They call that mismatch a data-belief asymmetry and treat it as the place new knowledge comes from. The Wrights had a theory that flight came down to lift, propulsion and control, and they generated the measurements that did not yet exist.
Brian Roemmele's write-up has the paper making the rounds again, with the takeaway that models will not independently produce the next breakthrough. I will sign off on that and underline "independently." The paper also left three songs about flying stuck in my head, so here is a Night Flight block.
Night Flight ran on USA Network from 1981, four hours of music videos, cult films and comedy on Friday and Saturday nights, all of it material the daytime schedule had no slot for. Somebody bet that an audience was awake for it, and the bet held for years. I read that as a data-belief asymmetry with a laugh track. What follows is three songs, no commercial breaks, and a theory at the end.
Test Pilot
TV on the Radio has a song called "Test Pilot." The way I read the lyrics, the band understood that miscommunication and passion can put you in the ground, and that the repair starts with a moment of clarity and balance before anyone touches anything. After that comes the healing and the moving on, and a willingness to work through it and keep flying in rough skies. It is a better post-incident review than most I have been handed.
The Wrights were test pilots in the literal sense. Their 1901 glider produced about a third of the lift their borrowed numbers promised. They did not ask the numbers for a second opinion. They threw out the published tables, built a wind tunnel that winter, and tested around two hundred wing shapes themselves. Then they climbed on top of the answer.
That last step carries a risk a language model does not share. In 1908, Orville was flying a demonstration for the Army at Fort Myer when a propeller failed. His passenger, Lieutenant Thomas Selfridge, died, the first fatality in a powered aircraft. Orville survived with injuries. A test pilot's performance review has an unusually short feedback loop. What the brothers did next follows the song's order of operations: they worked out what had failed, redesigned the aircraft, and went back up. The 1909 Flyer went to the Army.
I bring it up because Felin and Holweg's point about intervention has a physical side. At some stage the theory gets tested by something with a body at stake, or at least a budget. The model can propose the experiment. A person signs for it, and a person sits through the review afterward.
Learning to Fly
Pink Floyd's 1987 single of that name grew out of David Gilmour's real flying lessons, and Gilmour later described it as the band learning to fly again after Roger Waters left. I think about that every time a company tells me it has restructured around AI. Half the engineering team got a severance letter, the other half got a license and a memo, and everyone was informed they would be learning to fly again.
The vendors say the models are learning too. On May 20, 2026, OpenAI announced that one of its models had disproved Erdős's unit distance conjecture (we understand OpenAI is prone to embellished media hype; a Led Zeppelin goes over better). The proof was checked by mathematicians OpenAI brought in, and the account of how it was produced sits in papers OpenAI hosts. I will wait for that process to be open to outside inspection before I lean on it. Even in OpenAI's own telling, its team picked the problem and posed the question, and people edited and refined what came back. So I hold the strongest version of the paper's claim loosely, and I hold this announcement looser.
There is a limit inside that example worth keeping. A proof can be checked by other mathematicians and by software. Kitty Hawk could only be checked by flying. Where the world has to be poked to find out the answer, whether that is a market or a factory floor, someone has to choose which experiment to run, and that choice is a theory. OpenAI says something close to this in its own announcement: people pick the problems that matter, interpret the results, and decide what to pursue next.
Which is why I file most of the current hysteria under noise. The models are very good at compressing what we already know and getting it to us faster, and that is worth real money. On their own they will not hand us a new energy source or a new medical paradigm. Someone has to look at the existing record, decide it is incomplete or wrong, and then go build the wind tunnel.
Learn to Fly
The Foo Fighters video for the song is a parody of Airplane!. Two cabin cleaners spike the coffee, the passengers go down, and the band, having passed on the coffee in favor of liquor, ends up landing the plane. I have sat through enough steering committees to recognize the staffing model.
Every competitor can buy the same model, trained on the same public record. If your AI strategy amounts to asking it what the data says, you will get what your competitors get. Felin and Holweg's business conclusion is that advantage comes from theories specific to the firm. What separates one company's results from another's is the belief it holds that the data has not caught up to, and whether it runs the experiment to find out. It might be a customer segment you think is underserved, or a process you think can run at half the cost.
Before approving an AI initiative, I ask one question in writing: what do we believe that the data does not yet show, and what result would tell us we are wrong? Then I ask who is flying. If no one can answer both, the initiative has a tool and no test pilot, and the plane is being landed by whoever skipped the coffee. Use the tools. The model is an instrument on the panel. The pilot still flies.
The 1903 editorial was written by people looking at a long list of failed flying machines. The Wright brothers looked at the same failures and read them as a list of problems to solve one at a time.
That is the block. The Night Flight programmers and the Wright brothers shared a habit: a belief the numbers had not caught up to, and the nerve to put it on the air.