Ground Truth

Silicon, capital, and what intelligence cannot manufacture
Corpus The Next Fifteen Years / blog

Underneath every post is a ~100,000-word document built the unfashionable way: from physical constraints upward, with scored probabilities, named falsifiers, and a quarterly indicator dashboard. Posts here take one claim out of it, put current numbers against it, and say plainly what would prove it wrong.

No takes without an order book behind them. Each piece ends on the one line worth keeping. New here? The framework is the map: what each named concept claims, and how the posts fit together. And the claims that can be caught failing live on the scorecard, each with a probability, a resolution date, and the rule it will be graded by.

Posts

Fast Takeoff Has Three Brakes, and Two Are Not Technical The scenario where AI research feeds on itself is the highest-variance question in the whole forecast, and the number that would answer it is measured only inside frontier labs. Here is how to watch it from outside, and why the likeliest fast outcome is a fast decade, not a discontinuity. 8 min
The Robots Are Arriving as the Workers Leave Every advanced economy is running out of workers on a schedule that was fixed decades ago. That cancels the AI displacement story in aggregate, leaves it untouched in composition, and turns the 2030s labor problem into a shortage in exactly the work AI does worst. 8 min
Six of the Seven Are Owned by Capital The list of things intelligence cannot manufacture is also a list of assets, and assets have owners. Six rows belong to portfolios. One belongs to people, and the least intuitive prediction in the corpus rides on it. 7 min
The Robot Is Cheap. The Robot-Hour Is Not. A $20,000 humanoid looks like $3-an-hour labor until you divide by utilization and success rate instead of by calendar. The real crossover arrives first where labor is expensive to employ, not expensive to pay, and that ordering is checkable. 7 min
The Productivity Boom Is Waiting on Retirements Factories owned electric motors for twenty years before productivity moved, because the men who built their careers on the old floor plan were still running the plants. The AI J-curve resolves on the same clock, and the clock is management turnover. 7 min
Sell Ground Truth Generation is free and checking is not, so the durable businesses of the next decade sell verification. Here is the build list, ranked by which of it actually gets funded by 2031, and which stays an essay. 7 min
The Better AI Works, the More It Costs to Finance A real productivity boom raises the neutral interest rate, which discounts every long-dated asset more harshly, including the AI build itself. The likeliest world is stranger: AI works well enough to matter, and rates stay low anyway. 7 min
Guilds, Mediated by Cryptography Every signal that is asynchronous and unwitnessed is dying, from the cover letter to the product review to the photograph. What replaces them is older than what they replaced, and the replacement has a distributional sting. 7 min
The Incident Picks the Rules A serious AI-attributed incident is likelier than not by 2031, and the regulatory architecture of the 2030s gets written in the 18 months after it. Which rules win was decided earlier, by what was sitting on the shelf. 6 min
The Cheapest Decade in History Will Feel Expensive AI pushes the cognitive half of the economy toward free while energy, housing, care, and the trades inflate. Households only experience the second half, the statistics mismeasure both, and that gap becomes the central politics of the 2030s. 8 min
The Most Important AI Regulator Never Ran for Office Through 2030, the practical limit on AI deployment in high-stakes domains gets set by insurance exclusions, aggregate limits, and a handful of reinsurers, not by any statute. The first correlated failure makes it official. 8 min
The AI Race Ends in a Refinancing If the 2027 to 2029 correction arrives, it starts as a credit event in the machinery financing AI datacenters, not as a bad quarter at a big lab. Here is the sequence, the casualty list, and what survives it. 10 min
Every Junior Job Was Secretly Two Products AI now does the work juniors used to learn on, so firms have rationally stopped hiring juniors. The bill arrives around 2035, roughly a decade after the last moment anyone can act on it. 13 min
You Cannot Own a Falling Floor Inference cost at a fixed capability fell 100 to 1000 times in two years, on a published schedule. Any product whose moat is model access has about two years of shelf life. 9 min
The Taiwan Risk Is Not the One You Are Watching The likelier failure is not an invasion but a quarantine, a war-risk insurance premium, and an export licence that takes eleven months instead of three. 7 min
AI Is Superhuman Wherever the Scoreboard Is Cheap AI is superhuman at competitive programming and merely fine at telling you whether to take the job. What separates the two is the cost of checking the answer. 11 min
No Night, No Neighbors Solar is the cheapest electricity ever sold, for about a quarter of the day. The cost of firming the other three quarters, plus the politics of every terrestrial site, is the case for putting datacenters in orbit, and in June SpaceX turned that case into a product. 11 min
The Bottleneck Migrates: Silicon → Electrons → Permits For thirty years the binding constraint was chip supply. By 2028 it is interconnection queues, turbine order books, and environmental review. 9 min

Follow

RSS · the corpus · ask it a question · source

select · Enter open · Esc close