The Atlas
A personalised reader over a knowledge graph. Every node links to its neighbours; tags are nodes too. Pick a thread.
- 001The Bottleneck Isn't Capital. It's Noticing.The Recognition Thesis
A small firm loses its options not in the auctions it enters and loses, but in the ones it never sees. The highest-return AI is cheap, unbroken attention.
- 002When Uncertainty Pays YouConvexity
Cap your downside, leave your upside open, and the world's volatility stops taxing you and starts paying you. The only question left is the shape of the payoff, not the odds.
- 003The Most Valuable Thing an AI Can Say Is "I Don't Know"The Invisible Frontier
The same model can make you forty percent better or nineteen points worse while sounding identically sure. Abstention is the only signal that tells the two apart.
- 004The AI Wins You Can Measure Are the Ones That Won't Make You Any MoneyDifferentiated Value
The productivity gains you can put on a slide compete away into price. The durable edge is the rare, weighty decision where your instinct never formed and no one's data can settle it for you.
- 005Cheap to Make, Expensive to CheckThe Verification Economy
Generative AI drove the cost of making work toward zero and left the cost of checking it where it found it. Whether you stay the director or become a limb turns on a single question: does a cheap verifier exist.
- 006The Reverse CentaurWho Really Decides
Keep every human on the org chart, add a second approver, and the machine can still end up the sole decider: a firm that checks whether it hit the target, never whether the target was right.
- 007Who Drew the Menu?Agenda Control
Hand someone three options and the sense of authorship climbs, the override rate climbs, and the real decision is already gone. Which three reached the table, and what got cut, was settled offstage before the card arrived.
- 008Every Company Wiki Is a Graveyard. The Bug Is the Timing.Knowledge Escrow
Documentation fails because the cost of packaging knowledge lands on the author at deposit, when it pays nobody. The model moves that cost downstream, to the moment someone actually asks.
- 009The Cell Is Not a DecisionRisk, Measured Wrong
The red-amber-green grid was built to decide who may sign off, not to measure anything. A published theorem shows that a five-by-five can rank a rare catastrophe worse than a tossed coin.
- 010The Quadrant Is a Price, Not a PlaceThe Map Redraws Itself
The rare-and-severe corner of the risk map is not fixed territory. It is a line drawn by two prices, attention and process, and AI is collapsing both, squeezing the corner from opposite walls at once.
Stubs — referenced, not yet written
- 100Shadow OptionSTUB
An option a firm already holds as a by-product of its existing positions (leases, relationships, licenses), but has not recognized; left unseen until it expires, it never economically existed. Recognition, not acquisition, is the bottleneck.
- 101Recognition EngineSTUB
The highest-ROI role for AI in a small firm: not automating frequent work but pointing continuous, cheap attention at discontinuous payoffs on both tails — surfacing options and exposures before either expires.
- 102Frequency-Impact MatrixSTUB
The grid where frequent problems get processes and rare-severe ones land on the owner by default. Its boundaries are economic, not natural: the rare-and-severe quadrant is sized by the price of vigilance, which AI collapses.
- 103Verifiable SpaceSTUB
AI replaces search, not judgment. It wins only where a proposal can be checked more cheaply than it was produced; outside that space, fluent output is a liability.
- 104Prevention ParadoxSTUB
The value of a loss averted is counterfactual and therefore invisible — you cannot point to the disaster that did not happen. So every recognition design must carry its own evaluation: backtests, canaries, exercises.
- 105Latent ExposureSTUB
A negative shadow option: an unrecognized commitment to a future loss. It is cheapest to restructure precisely when nobody is looking at it.
- 106Knightian UncertaintySTUB
Risk has a distribution; uncertainty does not. The genuinely unknowable tail cannot be priced by a model and stays with the owner — which is exactly where convexity pays.
- 107Reverse CentaurSTUB
The inversion of human-in-command: the machine holds the decision premise and sets the pace, and the human is reduced to an executing limb that supplies a signature.
- 108Decision PremiseSTUB
A decision is a conclusion drawn from premises. Whoever fixes the decisive premise — the framing, the options, the conclusion presented as fact — has effectively made the decision, whatever the org chart says.
- 109Range CompressionSTUB
Ordinal bins (low/medium/high) throw away orders of magnitude. The matrix fails worst exactly at the tails, where the difference between a large and a catastrophic loss is the whole decision.
- 110Real OptionSTUB
A small commitment that buys the right, but not the obligation, to a larger commitment later. Most arise as by-products of positions a firm already holds.
- 111Option ChainSTUB
Exercising one option creates new shadow options; strategy is a sequence of cheap steps, each unlocking the next, rather than a single committing leap.
- 112ConvexitySTUB
Capped loss with open upside. Under this payoff shape uncertainty flips sign: the more unknowable the world, the more the position is worth — uncertainty becomes an asset rather than a cost.
- 113Jagged FrontierSTUB
AI competence has a ragged, invisible edge. Inside the frontier it helps sharply; just beyond it, it harms quietly and confidently — and the boundary does not announce itself.
- 114Out of DistributionSTUB
A model knows only the frequent. The unprecedented lies beyond its reach by construction, and its confidence does not degrade gracefully there.
- 115Base Rate TrapSTUB
For a rare event, the false-alarm rate decides the system's worth, not its accuracy. A 99%-accurate detector of a one-in-ten-thousand event mostly cries wolf.
- 116Metacognitive DemandSTUB
Directing AI shifts the cognitive load from doing the work to judging your own output. You cannot catch what you are unequipped to evaluate — so oversight quietly fails.
- 117Prediction vs DecisionSTUB
AI cheapens prediction and thereby raises the price of judgment. The machine supplies the forecast; the commitment, the accountability, and the value-laden fork stay with the human.
- 118Differentiated ValueSTUB
Commodity capability is competed away to zero. The durable edge is the rare, consequential decision — which is exactly the work AI measurably wins at least.
- 119Rare for You, Common for the CorpusSTUB
What a single operator meets once, the training corpus has seen a thousand times — true for the documented and recurrent, false for the genuinely unprecedented.
- 120Generation-Verification GapSTUB
Producing work is cheap; checking it is not. Offload only where a cheap oracle exists to verify the output, or the saved generation cost reappears as verification cost.
- 121The 70% ProblemSTUB
AI delivers the easy 70% fast and confidently. The last 30% is the expert judgment you cannot offload — and it is where the value and the liability live.
- 122Three Lines of DefenseSTUB
Operations, oversight, and independent assurance. A solo operator is all three at once and structurally lacks the third — the independent check is the missing line.
- 123Tacit KnowledgeSTUB
We know more than we can tell. The undocumented operating logic living only in one person's head is the dominant, invisible exposure of a micro-firm.
- 124Illusion of ChoiceSTUB
A curated menu hides the decision better than a verdict does. Picking within a fixed set feels like authorship while the real choice — the set itself — was made upstream.
- 125Agenda ControlSTUB
Whoever fixes the set of options has decided. The option set is the most decisive and least-watched premise in any decision.
- 126DeskillingSTUB
Automate the routine and the skill that catches errors quietly atrophies. It is the time-bomb under every oversight scheme: the checker loses the ability to check.
- 127Affordable LossSTUB
Entry sized by the loss you can absorb, not by an unquantifiable return. The effectuation rule that makes convex bets repeatable.