A reading both sides predicted settles nothing.

If progress moves at the pace of the steps AI cannot take ⁴, the scarce thing is the next step a model can be taught, and that can be paid for, with practice tasks (exercises a model is scored on while it trains) and graders (programs that do the scoring) sold to the labs, and with the work of fitting a model to the steps a firm's own jobs need.

• ○can't be tested yetLabs and companies keep paying more for the steps models cannot yet take, even as each new model takes more of them on its own.

The research AI can automate is narrow, and the judgement of a few top researchers, not more agents, still sets the pace.

• ○can't be tested yetAI's help with AI research brings steady gains, not accelerating ones.

Scaling laws promise better guessing of the next word, not any particular new ability, and good data is getting ever costlier.

• ○can't be tested yetBetter data, more than better model designs, drives how much a model learns from its training, so progress comes to rest on finding new data.Claimed by Dwarkesh Patel and Jerry Han

Practice tasks written by outside contractors cannot teach a model long work whose results are hard to check, in fields such as biology and security; only records of how experts actually do that work can.

• ○can't be tested yetTraining for long work whose results are hard to check, in fields such as biology, security and matters of taste, comes to rest on records of real work inside firms, because tasks staged by contractors stop teaching models.

Once many agents can work on a task at once, what limits the work is the cost of the computing and the systems around it, not what a single model can do.

• ○can't be tested yetWith many agents working on a task at once, most valuable office work comes to turn on the cost of computing and the systems around it, not on whether models are good enough.

The tests that define the frontier, the work it is judged on and the customers who pay are mostly American, and the most advanced chips stay hard for Chinese labs to get.

• ●holding so farChinese labs do not take the lead at the frontier while the best chips, the tests that matter and the paying customers stay American.

A data centre can be financed and designed quickly, but the grid is planned for slow change with warning, and electricians and welders are few; neither can be hurried.

• ○can't be tested yetData centres arrive faster and with less warning than the grid's planners assume, so getting connected holds back new American sites.Claimed by Byrne Hobart
• ○can't be tested yetToo few electricians, welders and other tradespeople hold back the building of data centres, so the largest technology firms keep paying to train them.Claimed by Byrne Hobart

The labs and cloud companies already test and watch their own models, and can give that away to keep their customers.

Data, workflow and the customer relationship are hard to copy, even when the model underneath is not.

• ○can't be tested yetIf training a model further on a firm's own work keeps getting cheaper, apps train their own instead of renting the best.Claimed by Sean Cai

Platforms will set rules for how outside agents may use them, much as websites tell search engines' crawlers what they may read.

Teaching models long work whose results are hard to check needs practice tasks and graders that people must build.

• ●holding so farLabs spend more each year with the firms that sell practice tasks, graders and expert data.

Where a computer can check the answer, as in mathematics and code, a model can practise without people writing the answers, and the checking can spread from there.

Hosting models, selling data and custom training each lead to the same product, an open model fitted to a firm's own work and billed by use.

• ○can't be tested yetLeading firms that host models, sell training data or do custom training, such as Baseten, Mercor and Applied Compute, converge on one business, tuning open models for firms and selling them the use.
• ○can't be tested yetFirms outside the labs become large buyers of training data and practice tasks, so the labs stop being almost the only customers.Claimed by Sean Cai

Each new general model from the labs learns many fields at once, so a model trained for one field keeps its edge only until the next release.

• ○can't be tested yetStart-ups that build a model for one field fall behind as the labs' general models improve, outside a few regulated niches where data is scarce.

No government agrees a pause, but the largest labs, loosely in step, slow the public release of their most powerful models.

• ○can't be tested yetA formal pause looks unlikely; instead the biggest labs, loosely in step, slow the public release of their most powerful models.

Labs compete for users and revenue, so a model kept back is revenue handed to a rival.

Interest rates, the hiring boom after the pandemic and its unwinding hit young workers first, whatever AI does.

Handing a task to an agent is easiest where people already hire agents and can say simply what they want.

• ○can't be tested yetAgents take hold first in job search, property and other matching markets.

When everyone has the same models, what a firm knows about its own needs is its edge, and asking the market for help gives it away.

• ○can't be tested yetFirms keep inside the work that would reveal what they know, and independent agents dealing for firms in open markets stay rare.

A task is handed over when checking the result costs less than doing it. Hadfield and Koh argue about accountability; this site carries the argument to the order of adoption.

• ○can't be tested yetMore knowledge work is held back by the lack of a check than by what the work requires, so what a firm keeps is the checking and the answering.This site's reading, not the author's

Every firm's work is different, and fitting a model to it takes engineers who understand both the model and the firm, of whom there are few.

• ○can't be tested yetEngineers who fit AI into each customer's own work stay scarce, so their pay rises or employers settle for weaker hires.
• ○can't be tested yetLabs keep buying the consultancies that put AI to work inside companies, to get their staff faster than hiring allows.Claimed by Byrne Hobart

When an agent does long, unpredictable work, the buyer wants someone with assets and a licence who answers for its mistakes, so AI reaches customers through such parties.

• ○can't be tested yetFirms increasingly buy AI's work through a party with assets and a licence that can be held liable, such as an established service firm.

Start-ups built around AI are cheaper and faster, and contracts and insurance can carry the liability.

An instruction to an agent cannot cover every case ⁶, so someone must decide in the cases it missed and pay when the work is wrong. Where a contract cannot be completed, ownership settles who decides ¹⁰³; so the party that bears the loss wants to own the operation, and a seller with nothing to lose is a poor party to buy from ⁷⁵. None of these writers makes the whole argument; this site joins them.

• ◑both sides expect thisInsurers go on writing AI's mistakes out of their standard policies, so whoever deploys an agent carries the loss unless it buys separate cover.
• ●holding so farBuyers built around AI go on purchasing firms that sell services, more each year than the year before.
• ○can't be tested yetService firms bought by owners built around AI earn higher margins afterwards than rivals that only buy the same tools, in trades not billed by the hour.

If insurers learn to price an agent's mistakes, a software seller can buy cover and warrant its work, as contractors do, and the buyer need not care who owns what.

As output per hour rises, people take part of the gain as leisure and share out the work that remains. Keynes argued it of machines in general in 1930; this site carries it to AI.

As machines make other things cheap, spending moves to work where the person is part of what is bought.

• ●holding so farCare, teaching, hospitality and other work where the person is part of what is bought take a rising share of spending and jobs as AI spreads.

Where better algorithms set the pace, wages jump and fall as tasks are automated ³⁹; where AI is sold with market power, Athey and Scott Morton add, workers can pay again when its price rises ⁴⁰.

• ●holding so farLabour's share of income falls as AI takes on more tasks, even as wages rise.Claimed by Pascual Restrepo
• ○can't be tested yetAs AI takes over paid work, governments move tax from wages towards spending, and later towards AI itself.Claimed by Anton Korinek and Lee M. Lockwood

AI captures the local knowledge that used to justify leaving decisions to people on the spot.

• ○can't be tested yetFirms using AI grow larger and take decisions further from the people on the spot.

What matters for most decisions is knowledge of particular times and places, which changes constantly and by its nature cannot reach the centre as statistics. Hayek argued it about central planning; this site carries it to AI.

• ○can't be tested yetLaw, contracting and other trades scattered by place stay split among many small firms, while the AI tools each trade uses gather around a few vendors.Claimed by Byrne Hobart

Sites that produce information are paid by visits, and AI that answers directly sends fewer of them.

• ○can't be tested yetPublishers lose the visits that pay for what they produce.

As answers replace visits, publishers charge the AI companies directly, by licence or by the page.

Most of AI's benefits are hard to see, while those facing automation outnumber those doing it, so politicians gain by opposing it.

• ○can't be tested yetAI grows more unpopular even as more people use it, giving politicians on both the left and the right reason to campaign against it.

Agencies, courts and industry standards set AI's rules before it becomes a voting issue.

The fastest way to teach a model a job is to record people doing it, so in some jobs the work becomes producing that record.

• ○can't be tested yetSome jobs move from doing the work to supplying the data that trains AI to do it.

Once a task costs less by machine than any worker would, services appear that were never worth paying a person for, and growth comes from filling them.

• ○can't be tested yetCheap AI makes many small services worth selling that were never worth paying a person for, and filling these niches is how AI lifts growth.

When AI does the specialised steps, the valuable worker is the one who can direct it across many of them.

• ○can't be tested yetThe gap in value between versatile workers who manage AI and narrow specialists keeps widening.

Codifying a firm's knowledge for AI is costly, lossy and needs upkeep, so scale pays where the work is generic, in computing and models, while firms closest to a trade's own knowledge win in applications.

• ○can't be tested yetFirms in AI-exposed industries do not all grow bigger by 2029; they consolidate in infrastructure and splinter in applications.
• ○can't be tested yetFirms that adopt AI move routine, recordable decisions to the centre and leave fast-changing and time-critical ones with local units; they do not centralise across the board.

AI lets one centre see and run many unrelated businesses, removing the overhead that made conglomerates trade at a discount.

• ○can't be tested yetGroups that run many unrelated businesses with AI come to be valued above the sum of their parts.
• ○can't be tested yetA group that runs many businesses with AI is valued above the sum of its parts only if it shares what each learns across them; a holding company that keeps them apart is not.Claimed by Soren Larson

Agents cut the cost of finding, bargaining and contracting, and they cut the cost of managing as well, so both sides of Coase's comparison fall ¹⁰². The authors who named the Coasean singularity put a question mark on it themselves and list the new frictions agents bring ⁹.

• ○can't be tested yetFirms in the industries most exposed to AI do not shrink towards nothing as agents cut the cost of dealing.

If a firm exists only because dealing with outsiders is costly, then agents that deal for almost nothing leave little reason to employ anyone. This is the popular reading of the Coasean singularity; no writer this page read holds it as stated.

Entry-level work is the simplest and the cheapest to hand to a model, so a large firm first stops hiring at the bottom, then loses its juniors, then its managers as later systems do longer and harder work, and last its executives ¹¹⁹. Its authors give no dates, call this the default course for large white-collar firms and not a certainty, and call the last stage their most speculative.

• ○can't be tested yetThrough 2028, firms that adopt AI shrink their entry-level and junior ranks before they remove layers of management.This site's reading, not the author's

Much of a manager's work is passing information up and instructions down. When a model can carry both, a firm needs fewer layers between the people who do the work and the people who answer for it ¹²⁶.

• ○can't be tested yetWithin a few years of adopting AI, firms have fewer layers between the people who do the work and the chief executive than similar firms that have not adopted it.

A firm is a coalition of people who want different things, and where no contract or rule settles a matter someone has to hold the right to decide and bear the result, a standing the law does not yet give AI ¹²⁴. Checking a model's work falls to the people in the middle ¹³¹. This site carries both further: the layers that stay are the ones that answer, sign or check, whatever their rank.

• ○can't be tested yetThrough 2030, firms that no person runs or answers for stay rare in lawful, regulated trades, because someone has to decide in the cases no instruction covers and answer for the result.This site's reading, not the author's

Company law already allows an entity that follows a program and that no person owns or controls ¹²⁸, and the authors of the pyramid essay expect a small number of firms to end with no human employees, the step they call their most speculative ¹¹⁹. The head of one lab expects a company of a single person to be valued very highly ¹²⁷. None says such firms will be common; this site states the strong form so it can be tested.

A platform with the most users gains most by serving good-enough models to all of them and selling better ones on top, whoever built them.

A neutral router that pools many firms' checked outcomes learns to solve their problems, new kinds above all, faster than any one firm alone, and pricing by outcome belongs with it.

The better a router works, the more easily buyers move to the cheapest model that is good enough, the harder sellers compete on price, and the less anyone needs a router.

When every input is needed at once, the one hardest to get sets the pace and its owner takes the profit, until the industry builds its way out.

• ◑both sides expect thisNVIDIA and AMD take more than 50% of the gross profit they share with Microsoft's cloud arm and the estimates for OpenAI and Anthropic.
• ×failing its testNVIDIA's lead over other chip makers in sales for data centres keeps widening as AI spending grows.
• ○can't be tested yetGovernments take direct stakes in the companies that make chips, power and models.

The spending on data centres has to be paid back by revenue that has not arrived yet, and the chips it buys lose value fast, so if the revenue comes late the investors take the loss.

• ◑both sides expect thisThe biggest cloud and platform companies spend more each year on buildings and equipment than AI earns in revenue.

Demand for computing runs into a supply chain that has not grown as fast, from the few firms that make the most advanced chips and the machines that make them to generators, cooling and skilled builders, so data centres are finished late while use keeps rising.

• ○can't be tested yetThe cloud companies start storing new chips instead of installing them, the sign Cahn watches for that data centres are running late ⁴⁸.

Customers could switch between models as easily as between clouds but mostly do not, and staying at the frontier costs too much capital for more than a handful of labs.

• ●holding so farOnly a handful of labs stay at the frontier, because it costs too much capital for more.

Competition and imitation pass most of an innovation's value to the people who use it, through lower prices; the innovators keep a sliver. Nordhaus measured it across American nonfarm business from 1948 to 2001; this site carries it to AI.

• ○can't be tested yetThe people who use AI keep far more of its value than the firms that make it.

Once a small model on a phone or laptop is good enough, running it there is cheaper and faster than paying a data centre for each request, and the meter stops.

• ○can't be tested yetMuch of AI's everyday work moves out of distant data centres onto people's own phones and laptops.
• ○can't be tested yetWork that must be decided faster than a message can reach a data centre, or on little power, is run on the spot, whoever holds the knowledge.Claimed by Soren Larson

The best models stay too large for a device, and people want the best model for work that matters.

The most capable models are made by a few firms in two states, and each state can license, limit or withdraw what those firms sell abroad, so a buyer's choice of model is bounded by where it sits.

• ○can't be tested yetWhich models a firm may use comes to depend on the country it is in and the countries it sells to.

A model whose maker publishes it can be copied and run anywhere, so a rule on who may buy a closed model leaves a capable open one within everyone's reach; the writers who argue the other side themselves name open models as the hedge.

• ○can't be tested yetA firm shut out of the leading closed models runs an open one that is close enough for its work.

The labs' general skills converge, so what separates them is the specialist business records each has bought, and each grows strongest in the trades it bought most from.

• ○can't be tested yetThe leading labs may pull apart by industry, each strongest where it has bought the most specialist business records.
• ○can't be tested yetLabs buy more of their training data straight from leading app companies and specialists, squeezing the firms that sell general practice environments.Claimed by Sean Cai

Each field and each kind of data needs its own experts and its own checks, so no one seller can be best at all of them.

• ○can't be tested yetThe market for training data stays split among many specialist sellers rather than gathering around one or two.

Labs want one supplier that can recruit experts at volume, check quality and pass security reviews, and those costs favour the largest sellers.

Competition among the frontier labs keeps switching cheap, and bans on copying models and exclusive deals leak.

• ○can't be tested yetLabs' efforts to tie customers down, through exclusive deals and bans on copying their models, eventually fail, and moving between labs stays cheap.

Models differ in efficiency and in the kinds of input they handle, and a system built from several models needs each one to fit its part.

• ○can't be tested yetModels keep differing in efficiency and in the kinds of input they handle, so they do not become an interchangeable utility like electricity.

New capacity takes years to plan, power and build, so it cannot be multiplied overnight the way new fibre equipment multiplied internet bandwidth.

• ●holding so farThe price of rented computing does not crash the way internet bandwidth did, because physical capacity cannot be multiplied overnight.

AI's demand strains every link from memory to power at once; somewhere supply will prove easier to expand than expected, and prices there will turn.

• ○can't be tested yetSomewhere along the chip supply chain, a part AI has made scarce proves easier to expand than expected, and its price turns.

The lab with the most popular model spends its chips serving it, leaving less to train the next, while rivals with spare chips train theirs.

• ○can't be tested yetThe lab with the best model keeps changing, because the leader uses its chips to serve its popular model while rivals train the next one.

Each fall in price makes AI worth using for more work, and use grows faster than the price falls.

• ●holding so farFirms' spending on AI per employee keeps rising, because what they use grows faster than the price of each unit falls.
• ○can't be tested yetMuch narrow work AI does is turned into cheap fixed routines, so spending on each such task rises for a while and then stops.Claimed by Byrne Hobart

Work gets moved to cheaper models and fixed routines, so falling prices cut the bill faster than new uses add to it.

NVIDIA lends to and invests in the data centres that buy its chips in return for a share of their income, turning one-off sales into a stream and gaining a say over who gets capacity.

• ○can't be tested yetNVIDIA keeps financing its own customers in return for a share of their income, and uses such deals to gain a say over who gets computing capacity.

The largest technology firms build more computing than they need and rent the spare to rivals, turning each firm's bet into one bet, shared across the industry, that demand for AI keeps growing.

• ○can't be tested yetThe largest technology firms build more computing than they need and rent the spare to rivals, so the industry becomes one shared bet on demand for AI.

Assistants can see what people want to buy earlier than search does, so advertisers will pay to reach them, and ads fund the free versions.

• ○can't be tested yetAdvertising comes to pay for free AI assistants, because they can spot what people want to buy before search does, though keeping users' trust will be hard.

Coding agents let firms build their own tools instead of buying them, but the programs that hold a business's core records are hard to replace and every new tool needs them.

• ○can't be tested yetCoding agents slow new firms' purchases of ready-made software but make the programs that hold a business's core records worth more.

Labs need staff inside their customers to deploy their models, and that staff grows with the customers.

• ○can't be tested yetThe labs' margins are pulled down because the staff who deploy their models grow with their customers, which a pure software business avoids.

The best model's lead over open ones is short, so a firm that owns its intelligence owns a wasting asset; and most firms see nothing in their own work that others do not see, so a system that learns from many firms beats one that learns from one ⁴⁵. What holds most organisations back is workflow, authority and data, not intelligence ¹¹¹.

• ◑both sides expect thisUse of models stays spread across several labs and open models, with the concentration of routed tokens among labs below 0.25 index, and few firms train or run their own.
• ○can't be tested yetA model of a firm's own pays only where the firm senses something no one else can; elsewhere a system trained on many firms' results does better.Claimed by Soren Larson

A supplier of rented intelligence sees how its customer works, can raise its price once the customer's work depends on it, and can learn enough to compete ¹¹². Garicano and Saa-Requejo answer by renting the model while owning the rules and records above it and keeping a second supplier. This site carries their warning further, to the case for owning the model itself.