Investment strategy · 14 Sep 2026
How to defeat Moloch: an investment strategy
Which businesses belong, what to build in what order, and how to keep power accountable as they succeed. The League’s global venture strategy.
Moloch: the coordination failure that makes almost everyone a villain and anyone the hero
The Antimolochian League · Bay Area
You are not the only one who sees it.
An AI lab can know a longer safety test is necessary and still fear a rival will ship first. A creator can leave a platform but lose years of work. A building can need a safety upgrade while every owner waits for someone else to commit. The League gathers builders, users, counterparties, and institutions to turn binds like these into ventures, commitments, and systems that make people more capable together.
The bad outcome can be obvious and still be hard to avoid.
The Antimolochian League is a Bay Area project for people building practical alternatives. We identify what someone would lose by choosing differently, then look for a change in the rules, ownership, or infrastructure that gives them a real choice.
What Moloch means
Sometimes everyone can see the damage, yet the system keeps rewarding the choices that cause it.
Moloch is not a villain, a company, or a political faction. It is a name for that recurring pattern. Allen Ginsberg used the figure in Howl. Scott Alexander later used it to describe situations where sensible choices, made one at a time, produce a result almost nobody wants.
In plain English
You may know what the group should do and still be unable to risk doing it alone.
Scott Alexander developed this use of Moloch in “Meditations on Moloch” (Slate Star Codex, 2014).
Liv Boeree
“It’s the god of unhealthy competition. It’s not the god of competition. Competition is a neutral thing.”Liv Boeree · On Bankless: Why Can’t We Have Nice Things (January 2023) · On her Win-Win channel: Why Beauty Filters Are Wrecking Our Brains · Win-Win, with Daniel Schmachtenberger: Moloch, AI & Capitalism
Our starting point
Who has to take the risk first? What do they lose if others do not follow? Which rule makes that loss unavoidable?
We focus on the part that can be changed. A creator should be able to carry verified work and reputation to another service. A small supplier should not have to repeat the same expensive compliance audit for every buyer. A group willing to fund a shared safety improvement should be able to commit on the condition that enough others join them.
These situations look similar from a distance, but they need different tools. Sometimes people need proof that others will move with them. Sometimes a gatekeeper profits from making departure expensive. Sometimes the reward for breaking ranks is so strong that only an enforceable rule will work. A short research note explains those differences and the limits of what a small project can change.
AI safety belongs in the last group. A lab can believe restraint is prudent and still fear that another lab will race ahead. Governments, insurers, standards bodies, and the labs themselves have the power to change those incentives at scale. The League’s narrower contribution is to make claims easier to check and to build tools that help people keep control of their data, work, and choices as AI spreads. The painting below shows the larger argument.
What exists today
One prototype: take your reviews with you.
x4me2 is a live prototype for recovering reviews a person wrote on other platforms. It checks authorship and hosts the recovered record outside the platform where it began.
Payments from AI systems are planned, not live. The current test is simpler: can someone recover a useful, verifiable record without the original platform’s permission, and what does each completed recovery cost?
Tyler Martin writes and edits the League’s public work. Each research page names the drafting tools used and the date of the piece.
Latest from the League
Working notes for people who need to change the rules, not merely describe the problem.
Research is useful when it changes a decision: what to build, what to test, which claim requires proof, or who needs to join the work.
Dated terms extract · 27 Aug 2026
How frontier labs use what you paste
Training, retention, review, and account terms are separate choices. This is a dated reading guide, not a vendor ranking.
Working note · 28 Aug 2026
Free tokens. Max exposure.
What a cheap model endpoint changes, what it does not, and where an agent’s permissions—not the model—decide what leaves a machine.
The argument in the painting
Keep feeding the system, walk away, or build something better.
The center shows people doing what the system rewards even as it consumes them. The right offers a clean escape for those who can afford to leave. The left shows the League’s wager: replace the dependency with something ordinary people can use, then make sure they can leave the replacement too.
The left side · the League’s answerBuild something better.
People replace old dependencies with open tools, portable records, and ways to deal directly. The alternative succeeds only if it gives people more control and preserves their right to leave.
The centerKeep feeding it.
Each person does what the system rewards: comply, submit, buy, repeat. Any one choice looks reasonable. Together they keep the machine running.
The right sideWalk away.
Leaving can protect the people able to do it. It does not change the conditions for anyone whose income, health care, or community still depends on the system.
One accelerationist answer is that faster technology can help replace failing institutions. The painting asks a different question: who owns the replacement, who sets its rules, and who can leave? The AI race makes that question urgent. The League reads AI 2027 and AI 2040: Plan A through a narrower problem: can labs verify shared limits and trust rivals to follow them? Governments and labs negotiate agreements at that scale. The League works on smaller pieces beneath them: portable records, independent checks, open interfaces, and alternatives that do not confiscate what people built when they leave.
The Moloch library
People discover Moloch through economics, game theory, sociology, and psychology. Here, some of our favorites.
Fritz Lang fed his factory workers to a furnace god in 1927, Allen Ginsberg turned the god on American industry in 1956, and Scott Alexander sharpened the name into an analytic tool in 2014. Since then the concept has spread through game-theory podcasts, AI-safety papers, and the mainstream press. This library collects the pieces we keep returning to, each with a summary of what it actually argues. Maintaining it is part of the League’s work.
Start here
Essay · 2014
Meditations on Moloch
The founding text. Alexander names the multipolar trap: a competition in which each participant must sacrifice something they value or lose to someone who already did. The same structure appears in arms races, markets, and evolution.
Video · 2021
Why Beauty Filters Are Wrecking Our Brains
The first entry in Boeree’s Moloch series and her best introduction to the concept: one concrete case, filters winning engagement face by face while shifting the baseline for everyone. The link starts where the Moloch explanation begins.
Talk · 2023
The Dark Side of Competition in AI
Brings the trap to the AI race: every lab expects its rivals to cut the corner first, so each cuts it in self-defense. Boeree reads the Montreal Protocol and arms-control agreements as evidence that coordinated exits have been built before.
Podcast · 2025
Moloch, AI & Capitalism
The flagship long-form conversation: how growth pressure and game theory generate the trap across the whole economy, and what AI does to a system already running on those incentives.
Short film · 2024
What is Moloch?!
A short film cut from Boeree’s interviews and writing, and the fastest route in for someone who would rather watch than read. Boeree has endorsed it as an accurate summary of her argument.
Browse the full library
Thirty-plus entries across essays, talks, research, press, and origins, each summarized in a line. Missing something? Send it through the form below.
How the League works
Find what keeps people stuck. Change that.
We begin with a person making a real decision. What would they choose if leaving did not cost their data, income, reputation, health coverage, or access? The answer points to the rule or product that needs to change, and to the venture that could change it. The foundation manifestos state what the League builds and how to take part.
Describe the bind
Name the decision, the cost of choosing differently, who sets the rule, and who gains from keeping it.
Change the rule
Make the relevant data, rights, benefits, or access portable, or remove the contract or standard that blocks a workable alternative.
Build something people can use
A theoretical exit does not count. People need to find the alternative, check the facts, agree to the terms, and complete the move.
Bring the people together
A mechanism needs founders, early users, the counterparty on the other side of the bind, operators to run it locally, and sometimes an institution to adopt a rule. The League aims to bring those people together around each test.
Publish the result
Report who gained control, who still pays, how many people used the alternative, and where the attempt failed, so the next builder can replicate it or skip it.
The boundary
People must be able to leave what we build.
A replacement has failed if it requires technical expertise, forces people to use AI, takes ownership of their identity, or makes one provider impossible to replace.
ChoiceAI may be available. It cannot be the price of admission.
ControlPeople can inspect decisions, correct mistakes, move data or reputation elsewhere, and appeal.
AccessDisability, income, language, available time, personal safety, and technical skill cannot be afterthoughts.
EvidencePublish what changed, who benefited, who paid, and what failed.
DependenceAny AI component must work across more than one provider and keep an open or no-AI fallback. People must be able to leave the system itself.
What success means
Can people make a better choice without paying for it somewhere else?
Speed is easy to count. We also ask whether people can leave, keep what they earned, correct a bad decision, and use the alternative without new risks being pushed onto someone else.
Useful workDid the same effort produce a result someone needed?
Room to chooseCan people switch services, jobs, or providers without losing their work, reputation, or access?
Shared resultDid the change make the better option practical for everyone who had to act?
DurabilityDid the alternative stay open, or did it become another gatekeeper?
An open question
When can a small, independent service become easier to use than a locked-in platform?
Large platforms and institutions solve real problems. They help people find a match, agree to terms, check identity or eligibility, and complete a handoff. They also set the rules and can make leaving expensive. Falling AI prices may let smaller providers do some of that work with less overhead.
The calculator below does not answer that question. Its examples are hypothetical, and it only shows how a set of invented assumptions interact. Every output is a scenario, not a measurement or forecast. Our first published measurement plan defines the real counts we will collect from review recovery on x4me2.
See the hypothetical examples and assumptions
A shared fire-safety upgrade
- Why people wait
- Owners in a multi-unit building agree that a sprinkler retrofit is necessary. No one wants to pay a first share while the project can still fail for lack of commitments from everyone else.
- What could change
- Each owner signs the same conditional agreement. Money is collected only if enough owners commit for the complete retrofit to proceed.
- What to measure
- Commitments completed, time to reach the threshold, cost of collecting them, and whether the retrofit was finished.
Portable platform reputation
- Why people stay
- A creator can download old posts, but another service cannot verify the reviews, audience relationships, or authorship history that give the work value. Leaving means starting from zero.
- What could change
- A signed, portable record lets competing services verify the history without asking the original platform for permission.
- What to measure
- Records recovered, authorship checks passed, time and cost to move, and the share accepted by another service.
A supplier repeating the same audit
- Why entry is expensive
- A small manufacturer must pay to prove the same safety and sourcing facts again for each large buyer. Bigger suppliers absorb the repeated audits; smaller ones never reach the approved list.
- What could change
- Buyers accept one signed evidence package from any auditor that meets a shared standard.
- What to measure
- Repeated audit cost avoided, time to approval, buyers accepting the record, and errors found after approval.
AI price at break-evenScenario: $22/M
Base cost per handoffScenario: $49
Completion rateScenario: 52%
The dashed line adds assumed costs pushed onto other people, such as lost portability or restricted access. The solid orange line includes only costs visible to the person making the decision.
The price of AI work falls from left to right. The shaded region begins where the hypothetical independent service costs less than staying with the current provider.
The point of the sandbox is to expose the assumptions, including how often an attempt succeeds and what it costs to verify a result. It does not predict that platforms, companies, or capital disappear.
In this invented scenario, the independent service costs less below $22 per million tokens.
How the sandbox works and where it fails
Calt(p) = p × 2.0M / 0.52 + $49
Completion rate: s = 0.25 + 0.75 × access × awareness. AI cost: token price × tokens per attempt ÷ completion rate. Both formulas are provisional choices, not findings.
The alternative’s base cost: checking the handoff, administering consent and appeals, serving people with access barriers, and moving records from the old system. Stronger protections cost more to operate and reduce the assumed harm.
The current provider’s visible cost: an assumed amount for search, bargaining, finance, contracts, enforcement, management, and profit, plus harms the decision-maker already recognizes. The current provider also uses AI, so its assumed cost falls as AI gets cheaper.
One shared slider: “People who recognize the cost of staying” changes both lines. It raises the alternative’s assumed completion rate and the effect on others counted against the current provider. Moving it therefore pushes the lines toward each other by design.
Resistance: the current provider may respond by making export, authorship checks, or switching harder. The slider adds up to $90 to the alternative’s base cost. The default of 0% is intentionally favorable to the alternative.
AI-provider dependence: the alternative is also vulnerable if it relies on one model company. The boundary above still applies: more than one provider, an open or no-AI fallback, and a preserved right to leave.
The idea fails if checking results cannot keep up with demand, access remains unequal, the service turns into surveillance, one provider takes control, or the move abandons useful records, obligations, workers, or vulnerable people.
Every adjustable dollar amount is a scenario input. No measurement or forecast stands behind the defaults.
Background reading: Hayek on dispersed knowledge; Ostrom and Williamson on alternative governance structures; NIST’s AI Risk Management Framework. The 2026 frontier-price band reflects published list prices across input and output tokens: Anthropic pricing; OpenAI API pricing.
Join the League
Join the Antimolochian League.
The League is a Bay Area community of people building ventures, commitments, and tools that get others unstuck. Bring a case, bring a project that is already changing a rule, a record, or who owns the data, or offer to be the counterparty a project is missing.
Describe the choice someone wants to make, what they lose by moving first or leaving, and who controls the rule. A real example is more useful than a theory, and it is how people join.