AI and the Pattern Matching Conditions Part 1: Three AI Debates, One Answer

There are three separate conversations about AI happening right now, and they barely acknowledge each other. And all are answered by the Pattern Matching Conditions, a framework for predicting which jobs AI will displace and where we should see rapid growth.

The first is about jobs. Labor economists, policymakers, and worried workers argue about whether AI produces mass displacement or manageable transition. The second is about capability. Researchers and forecasters argue about whether AI is approaching general intelligence or plateauing into useful infrastructure. The third is about money. Investors and analysts argue about where the trillion dollars of AI capital expenditure will actually produce returns and where it is bubble.

Each conversation has its own experts, its own vocabulary, and its own running disputes. What none of them seem to notice is that they are all asking the same underlying question: where, specifically, will AI produce transformative outcomes, and where will it not? And I have come to believe, after a year of writing through these debates twice a week, that all three questions have the same answer.

The Four Conditions

A domain will see genuinely transformative AI outcomes when four conditions hold simultaneously. The work has binary or near-binary outcomes, results that are simply correct or incorrect. Extensive high-quality training data exists, decades of documented examples for the model to learn from. The work is fully digital, executable on a computer without physical-world interaction. And, the condition that turns out to be load-bearing, an automated verification oracle exists: a mechanism that can check whether an output is correct instantly, at machine speed, without a human in the loop.

Where all four hold, expect rapid compounding progress, genuine displacement pressure, and real economic value creation. Where even one fails, expect slower, human-mediated, friction-heavy progress regardless of how impressive the underlying models look in demos.

The oracle deserves a sentence of explanation because it is the mechanism that makes everything else work. Modern AI systems generate candidate outputs at extraordinary speed. Speed alone produces nothing useful, a billion unverified diagnoses is just a billion guesses. But pair that speed with an automated checker, a proof verifier, a test suite, a simulation, and an exhaustive search, and it becomes functionally indistinguishable from insight. The proof is valid whether or not the system understood what it was doing. Without the oracle, speed produces plausible-sounding output that still requires expensive human judgment to evaluate. With it, fast becomes better. The oracle is what converts capability into outcomes.

The Evidence, Across All Three Debates

Watch how the same four conditions sort the evidence in each conversation.

In the capability debate, the most significant AI result to date arrived last week in formal mathematics, when an OpenAI model disproved an Erdős conjecture that had stood since 1946. Mathematics is the domain where the four conditions hold most completely: proofs are binary, the literature is vast, the work is entirely symbolic, and proof verification is automated. The breakthrough arrived exactly where the conditions predicted it would, and the 150 mathematicians who signed the Leiden Declaration days later were correct that it does not generalize, because the domains they were defending it from, and the domains the hype extends it to, lack the oracle. Meanwhile, coding benchmarks have saturated, with success rates rising from 2 percent to 94 percent in three years, in the other domain where test suites provide an automated oracle. Medicine, law, and strategic planning, the domains where transformative claims are loudest, have produced impressive demonstrations and no verified breakthroughs, because no oracle exists to separate their signal from their noise.

In the jobs debate, displacement pressure is concentrating in precisely the occupations where the conditions hold: document review, entry-level coding, routine auditing, standardized underwriting, and tax preparation. The recent New York Times profile of Schneider Electric captured both sides of the line in a single article. Call center agents, whose hardest work is contextual judgment with no oracle, were complemented and became more productive, with the least experienced workers gaining most. The wiring of a standardized industrial contactor, a binary, documented, fully engineered task, was automated outright. Same company, same technology, opposite outcomes, sorted exactly by the conditions. And at the macro level, eighteen months into the loudest displacement predictions in history, U-6 unemployment sits at 8.1 percent, real personal income is at an all-time high, and GDP is growing at 3 percent, because the occupations satisfying all four conditions are a bounded slice of the economy, not half of it.

In the investment debate, the pattern is the same but almost nobody is using it. Analysis focuses on which labs have the most impressive models, when the more decisive question is which application domains have the verification infrastructure to convert model capability into compounding outcomes. The revenue is real where the oracle is real: coding tools, security analysis, computational chemistry with automated assays, protein structure prediction. The losses are accumulating where deployment ran ahead of verification: the enterprise pilots that never left pilot, the agents that demo well and cannot be trusted, the investment-driven layoffs awaiting a capability that has not arrived. Capability is priced. The oracle boundary is not.

One Mechanism, Three Predictions

If this unification is right, it makes falsifiable predictions in all three debates, and stating them now is the point of this series.

On capability: the next major verified AI breakthroughs will arrive in oracle domains, formal mathematics, computational chemistry, materials simulation, and not in medicine, law, or strategy, regardless of model generation. On jobs: displacement will remain concentrated in the four-condition occupations, employment elsewhere will remain robust, and the macro data through the next several quarters will continue to show no economy-wide rupture. On value: the durable AI revenue through the end of the decade will come disproportionately from oracle-domain applications, and the write-downs will come disproportionately from deployments that lacked verification infrastructure.

If verified breakthroughs start arriving in oracle-free domains, if broad displacement appears in occupations failing the conditions, or if the durable revenue ends up concentrated where no oracle exists, the framework is wrong, and I will say so.

The next piece in this trilogy maps how oracle domains connect to everything else, the spillover chains through which a mathematics breakthrough does, eventually, reach a hospital, and why some of those chains move in years while others take decades. The third piece turns to what all of it means, for workers deciding what to learn, for policymakers deciding whom to help, and for anyone trying to tell the difference between the AI outcomes that are coming and the ones that are only being promised.