Tell a British colleague their plan is quite good and you have told them it is adequate and you have reservations. Tell an American the same thing and you have told them it is very good. Same word, same sentence, opposite information. Neither party has any reason to check, because both of them heard a word they knew.
That is the smallest demonstration I know of something the AI conversation is getting wrong at scale. The words are shared. The world behind them is not.
Two assumptions are holding up most of the AI strategy documents I read: that capability is growing exponentially and will continue to, and that language is an adequate representation of the world. If both hold, the aggressive three-year plan is the correct plan. If either is shakier than it looks, a great many roadmaps are bets wearing a schedule's clothes.
Take the first. Václav Smil's work on how technologies develop is a useful corrective, mostly because it is unromantic: exponential growth in any physical system is always a temporary phase. Every technology traces an S-curve: slow start, acceleration, then a plateau imposed by physical or systemic constraint. The mistake is not being impressed by the acceleration. It is failing to recognise it as temporary, which is easy to get wrong from inside one, because from inside one every data point confirms it.
The honest counterargument deserves to come before my conclusion rather than after it. Progress can bypass a ceiling by leaping to a new curve: vacuum tubes to transistors to microchips, each generation hitting its limit and being succeeded rather than stopped. The industry is attempting that now, so I am not claiming capability stops advancing.
The claim is narrower, and it matters if you are signing anything. Each leap is a bet, not a schedule. An organisation that has built three years of plan on a particular bet paying on a particular timeline has confused a research programme with a delivery roadmap.
The second assumption is rarely stated out loud, which is part of how it survives. That a large enough model trained on enough text approximates general intelligence rests on a claim about language: that it carries the world well enough to stand in for it.
Language is an extraordinary human achievement and a thin, lossy representation of physical reality. Michael Polanyi made the point half a century before it became commercially relevant. We know more than we can tell. The engineer who senses a design is wrong before she can say why. The nurse who knows a patient is deteriorating ahead of the monitors. That is not decoration on top of explicit knowledge. It is the substance of expertise, and almost none of it was ever written down, because it cannot be.
A model trained on everything we have managed to tell is still missing everything we cannot. And if two dialects of the same language cannot agree on quite, the idea that a large enough corpus holds a stable map of the world might reasonably be asked to clear that hurdle first.
Brian Friel understood the problem long before anyone needed it for a strategy deck. Translations is set in a Donegal village in the 1830s, during the Ordnance Survey, and turns on surveyors mapping Ireland by replacing Irish place names with English equivalents. The new map is accurate. By the standards a map is judged on, it is better. And the local histories and sensory knowledge those names carried are gone, and nobody can quite say what has been lost until it already is.
That constraint is structural rather than a stage of development, which is why these systems behave as amplifiers rather than oracles. A lossy map cannot supply the judgment it leaves out.
Something to try this week. Pick one task you are planning to hand to a model. Go to the person who does it today and ask them one question: what do you know about this job that isn't written down anywhere?
Take notes for fifteen minutes. Whatever they tell you is the part of the map that does not exist, and it is the part your plan has quietly assumed away. If the list is short, you have found a good candidate for automation. If it runs long, you have not found a task. You have found a person.