Larger Than Live

Larger Than Live

The rockstar comparison is usually a way of saying someone is famous. That is the wrong axis. What changed for performers was two technologies, not one. Recording made the take durable. Distribution — radio, then MTV, then the internet — put that take in unlimited rooms without the performer travelling. A master in a vault is still one room. The superstar distribution is recording plus a pipe.

Superstars

Sherwin Rosen wrote this down. In “The Economics of Superstars” (American Economic Review, 1981) he showed that when technology lets one person’s output serve unlimited demand, small differences in talent produce enormous differences in reward. Marshall had seen the shape of it. Rosen made it a result. It is not a metaphor for this post. It is the economics the rest of the argument sits on. Recording without distribution does not serve unlimited demand. Radio, MTV, and the internet were the demand side of the same machine.

What that machine actually multiplied, in the acts that outgrew the stadium, was not range. It was truth-following with consequence: they would not sand the take when the room, the label, or the radio wanted something cheaper. The ones who flinched got recorded too. They did not become larger than live.

Presence

What used to cap the operator was the same thing that capped the singer: presence. An architect could be on one project. A surgeon could be in one room. Judgment did not record. Memos, hires, and policies faked scale. Headcount was how you pretended judgment had been copied.

Work scales as evaluation, not headcount. The unit of analysis is the endeavor: a firm, a case, a raise, a life’s work. You do not scale an endeavor by adding seats. You scale it by making the work evaluable.

Recording for judgment

A runtime of three things is the recording technology for judgment:

Specification is the cost function. Write the work so substitution preserves meaning, run it on that runtime, and one person’s decisions apply at institutional scale without that person being in the room. The record and the checks are the distribution channel: anyone can run them. An agent without a record a stranger can run is a tape in a drawer. Rosen then follows on the demand side: unlimited demand for a recorded take that actually ships, small differences in the operator, enormous differences in who is heard.

I built a production settlement platform solo, no team. The theory came out of that order, not the other way around. I have become functional in an answerable domain outside my trained expertise. The interpreter supplied the content. The hardness did not come from the interpreter.

The commercial name for the compliance consequence is compliance that compiles: an obligation goes in as a rule and comes out as a running system that emits its own proof. A signature on a control statement is still one room per night.

Answerable work

The work that resists automation is not the hardest work. It has two properties at once. The consequence is irreversible. It is attributable to a person or entity under law.

Hard systems programming has both. Settlement has both. Medical software, autonomous driving, autonomous robotics, and law have both. Project-management software and desktop tooling have neither. A missed standup is reversible and nobody is found against. A missed invariant in a settlement, a device, a vehicle, or a filing is not, and someone is.

A model can produce the artifact. It cannot hold the obligation. It cannot be the party a regulator finds against. It cannot choose which invariants are load-bearing. Answerability is not a capability, so no amount of model improvement touches it. It was never on the capability axis. That is why the operator remains.

Consequence-hardness

The scarce input is not breadth. It is consequence-hardness: the discipline acquired by doing hard work by hand somewhere failure costs something. You verify before you depend on anything. You keep a record because the record catches your own errors. You do not trust an output you have not checked.

Consequence-hardness is domain-general. Domain knowledge is domain-specific. The tooling supplies the second and cannot supply the first.

That is why an operator hardened in one answerable domain can become functional in another they have no training in. The interpreter supplies the content. The hardness supplies the rigor that makes borrowed content safe to rely on. Someone with the same tools and no prior consequence exposure has neither half.

This is a mechanism, not a boast about studying five fields. Hardness earned in one hard domain ports the moment an interpreter supplies the content.

The operator axiom is the same as in the formal paper. The theory requires a conscious operator who chooses what must remain true and stands outside the calculus. That is an axiom, not a result.

Truth-following under cost

Rarer than domain knowledge, and rarer than consequence-hardness, is truth-seeking and following it at all costs.

Consequence-hardness tells you what is true. This decides whether you act on it when acting is expensive. You can hold the scar tissue and still flinch. Most do, because the true answer costs a job, a round, a client, a relationship. That is the same property the best rock stars had. Recording did not invent it. Recording made it outsized. A runtime that records judgment will do the same. The operator who flinches will also scale. They will scale a comfortable falsehood.

Truth-seeking is a component, not a virtue. The runtime faithfully evaluates whatever specification it is handed. An operator who cannot accept a costly truth will specify a comfortable falsehood, and the runtime will scale it perfectly. The machine amplifies the operator’s honesty and dishonesty with equal fidelity.

A model will return a confident, fluent, well-formed artifact supporting whatever premise it is given. Comfortable falsehood is now cheap, polished, and infinite. Building the case for something false used to take work, and the work sometimes changed your mind. The agreeable path is now one prompt away.

It cannot be delegated. An agent has no stake in the answer. A system with no stake cannot supply the disposition.

The binding constraint has moved up the stack twice.

Each step up is less supplyable than the last.

Truth-following can only be verified retrospectively, through a record of costly choices. That is why an append-only record is one of the three legs of the runtime rather than merely memory. The record audits the operator, not only the work. It is the only instrument that does. The component that binds hardest is the one that is unobservable in the present tense.

The supply of judgment

Rosen is demand. The matching claim on supply: the runtime that scales judgment suppresses the production of judgment.

Lisanne Bainbridge, “Ironies of Automation” (Automatica 19(6):775–779, 1983): automate most of the work and the human remains responsible for precisely what cannot be automated, while no longer practising the skills their ongoing work used to supply. Her conclusion was that such operators need more training, not less, for the rare crucial interventions. The tooling removes the hand-work that produces consequence-hardness. The residual stays with the operator. The practice that used to produce the next one does not.

Bainbridge did not have the second mechanism. Surface breadth is now cheap. Assisted breadth feels like cross-domain judgment, including to the person holding it, so they never do the work. You cannot develop hardness you do not know you are missing.

The same tooling that multiplies an existing operator suppresses the formation of the next one. The supply is non-renewing. The moat widens. The replacement rate is currently negative.

Reversing it requires reintroducing the hard reps. Aviation mandated manual flying after automation-dependency accidents. Software has done nothing equivalent. The test is whether consequence-simulation or mandated-manual-practice pedagogy emerges in these domains. If it does not, the rate stays negative.

Where it breaks

The superstar distribution is brutal rather than admirable. Describing extreme concentration of reward is not endorsing it. Recording did that to singers. A runtime that records judgment will do it to operators.

The operator axiom cuts both ways. A runtime evaluating a wrong specification produces confident garbage at institutional scale. That is worse than the org chart it replaced. Fewer people will have practised enough to notice the specification was wrong.

What it predicts

Near-term, scorable:

  1. Solo operators delivering institution-grade work in answerable domains, at a quality that survives audit, without a department around them.
  2. Diligence asking to run the artifact rather than to meet the title.
  3. The replacement rate of consequence-hardened operators remaining negative unless mandated-manual-practice or consequence-simulation pedagogy shows up outside the domains that already have it.
  4. At least one real compliance process, in the next two years, accepting an executable proof in place of an attestation.
  5. The people who successfully cross into new answerable domains using these tools will be those with prior consequence exposure, not those with prior breadth.

The formal version is The First Scaling Theory. The accessible companion is Martian Moneyball.

References