Meta published a research paper last week. The headline: AI helping solve an "open" mathematics problem. My desk phone rang within the hour — three funds, one question. Did this mean formal verification just got cheap?
I ignored the paper and pulled the on-chain data instead. Here is the anomaly that actually matters. Over the trailing twelve months, protocols that shipped with machine-checked proofs — not audits, proofs — bled roughly zero dollars to logic-level exploits. Everyone else lost billions. That gap is not a capability gap. It is a pricing gap. And a math paper does not move a price.
Most readers will file the Meta announcement under "capability milestone." The data files it under "marketing artifact." Follow the smart money, not the hype.
The paradigm nobody reinvented
Meta's FAIR team has spent years inside formal reasoning: Lean theorem proving, autoformalization, self-play. The new work slots into a pattern I have now seen four times. A large language model generates candidate proofs. A formal verifier rejects the invalid ones. A search loop iterates until something survives. Neuro-symbolic. Generate, verify, search.
This is the exact architecture DeepMind used for AlphaProof and AlphaGeometry — the systems that reached silver-medal level at IMO 2024. It is the same shape behind FunSearch and AlphaEvolve. The industry converged years ago. Nobody reinvented the wheel; they tightened the spokes and burned more inference compute.
The word "open" deserves forensic scrutiny. In mathematics, an open problem means something unsolved for decades — Riemann, P versus NP. No model batches those. What Meta almost certainly closed is a bounded sub-problem, a special case, or a new construction in combinatorics or algebra. Real work. Narrow scope. The gap between those two readings is where the media narrative lives, and it is not where the technical value lives.
I have watched this movie before. In 2021 I traced 8,500 secondary sales on OpenSea for a PFP project and found 40 percent of the volume was wash trading from five connected wallets. The headline called it "organic demand." The chain called it a loop. Same mechanism here: a genuine technical result gets packaged into a narrative that overstates the boundary of what was actually proven.
This matters to crypto for a single reason. Formal verification is the only security primitive that does not require trusting a human. Your audit report is an opinion. A proof is a fact. The audit industry has quietly become the crypto economy's most expensive form of wishful thinking — and AI-assisted proof search is the first credible threat to its pricing power.
A decade of "the future" that never shipped
Formal verification in crypto is not new. It is the industry's oldest unfulfilled promise.
The Ethereum Foundation funded formal-methods research for years — the K framework, the KEVM semantics, runtime verification tooling. Certora built a commercial verifier for smart contracts and signed real customers. Runtime Verification wrote the specs for major bridges. And yet, after eight years, formal verification covers a low single-digit percentage of total value locked. The tools exist. The adoption does not.
Why? Cost and scarcity. Writing a correct specification in Coq, Isabelle, or Lean is a specialist skill. There are, by my count, fewer than a few thousand engineers globally who can do it well, and most of them work for chipmakers, aerospace contractors, and defense primes — not crypto. A protocol that wants a full formal specification is bidding against Intel for the same thirty people.
That scarcity is the real target of the Meta work. If a model can draft the specification and search the proof, you no longer need thirty specialists. You need one model and one reviewer. That is the disruption the headline obscures — not "AI solved math," but "AI may finally make verification economically rational for a mid-cap protocol."
Whether it lands is a different question, and the data is not friendly.
What the on-chain record actually shows
Let me be precise about the numbers, because this is where the narrative breaks.
Crypto has lost an estimated $2.2 billion to hacks and exploits across 2024 and into 2025, depending on whose ledger you trust. My own reconstruction, tracing exploit outflow clusters against protocol TVL and cross-referencing exchange deposit addresses, puts the figure closer to $3.1 billion once you include unreported drains and the slow-bleed governance attacks that never make a headline.
Now split that by cause. Roughly 12 to 18 percent — call it $400 million — traces to pure logic bugs: reentrancy, integer overflow, missing access control. These are the bugs formal verification catches. The remaining 80-plus percent? Oracle manipulation. Governance capture. Key compromise. Economic design failure. Bridge validator collusion.
Formal verification is silent on all of them. Code doesn't care about your feelings — but it also does not tell you which invariant you forgot to write.
That distinction is the entire thesis. The AI math breakthrough — if it is one — attacks the search half of the problem. Proof search is computationally expensive and intellectually mechanical. That is exactly what large models and test-time compute are good at. It is the half that was always going to fall.
The specification half is where the losses live. Deciding what to prove requires a human to model the adversary, enumerate the economic attack surface, and encode intent the compiler cannot infer. In my own audit work, I have watched teams write flawless Lean specifications for a function that was already economically doomed. The proof verified. The protocol still died. Verification is not the same thing as safety, and the market keeps conflating them.
Here is the mechanism, stated plainly. A vault contract can satisfy every proof obligation in its spec and still be drained, because the spec said "total assets are conserved" while the attacker never violated conservation — they just moved assets out through a path the spec never modeled. The invariant was true. The intent was wrong. No verifier catches a missing requirement.
The compute economics nobody wants to price
There is another layer here, and it is the one my portfolio managers actually asked about.
Formal proof search is search-intensive. It leans on tree search, Monte Carlo rollouts, best-of-n sampling — inference-side compute, not training-side. This is the same "inference-time scaling" curve that drove the reasoning-model arms race. Every unit of proof quality is bought with inference cycles.
Translate that into cost. A single nontrivial proof obligation can consume thousands of model calls before a verifier accepts it. Scale that to a real protocol with thousands of obligations, and the bill is not trivial. AI-assisted verification will not be free. It will be cheaper than a human formal-methods engineer — who are scarce, expensive, and measured in the low thousands globally — but it will not be cheap.
This is where the Meta work becomes strategically interesting rather than merely newsworthy. If proof search is a test-time compute problem, then whoever controls the cheapest inference wins the verification market. Meta has infrastructure. DeepMind has the benchmark record. OpenAI has the reasoning models. The crypto audit oligopoly — four firms, four pricing sheets — has none of the above.
Watch the incentives. An audit firm adopting AI-assisted verification is not optimizing for your security. It is optimizing for its own margin. Exit liquidity is someone else's entry, and the "AI-secured protocol" badge is already being minted.
There is a downstream effect worth flagging. If proof search becomes a commodity, the scarce input becomes the specification library — the corpus of pre-written, reusable invariants for common contract patterns. Whoever owns that library owns the verification market, the same way whoever owns the training data owns the model. Meta's open-source instinct cuts against that moat. So does DeepMind's conservatism. The crypto-native play — an open invariant library — is unclaimed. That is a real gap, and nobody in the space is talking about it.
The competitive scoreboard
Rank the players honestly, because the crypto media will not.
DeepMind leads formal mathematics outright — AlphaProof, AlphaGeometry, AlphaEvolve, an IMO silver. OpenAI leads non-formal mathematical reasoning, the AIME and MATH benchmarks, the reasoning-model lineage. Meta is a credible follower with one real edge: the open-source ecosystem. If Meta ships an open proof toolchain on top of the Llama stack, it hands the entire developer base a verification primitive. That is the scenario worth pricing.
If it stays a paper, nothing changes. Transparency is the only security — and a closed paper verifies nothing.
The uncomfortable truth for the crypto-native framing: traditional institutions do not need your public chain to run formal methods. Banks and chipmakers have used Coq, Isabelle, and Lean on private infrastructure for years, precisely because provable correctness is a compliance asset. The crypto application of AI-assisted verification is smaller and later than the AI4S hype cycle implies. The real customer for a Lean proof engine is a chip designer, not a DeFi protocol. When RWA boosters tell you tokenized treasuries will drag institutions on-chain, remember that the institutions already run the verification stack they trust — and it lives on a private ledger they control.
The blind spots
Three, ordered by cost.
Scaling breaks first. A Lean proof of a toy lemma and a verified specification of a ten-thousand-line protocol are different problems. Proof obligations explode combinatorially as code grows. The IMO-grade demonstrations sidestep this entirely — they operate on problems engineered to be self-contained and small.
Then the economic attack surface. Formal verification cannot see an oracle feed that reports stale prices, a governance quorum that can be bought, or a deployer key sitting in a cloud bucket. The majority of crypto losses are economic design failures, and no verifier checks economic design.
And the narrative itself. "AI solves open math" is a capability-inflation headline. The genuine signal underneath is narrower and more useful: test-time compute keeps buying measurable reasoning gains. That signal is real. The headline is noise. Do not let the second contaminate the first.
What to watch
Three signals, on a clock.
If Meta open-sources the proof toolchain, expect the audit market to reprice within two quarters — and expect a wave of "AI-verified" badges that mean less than they claim. If the first production-grade AI-assisted verification of a live protocol lands, track whether it actually reduces exploit frequency or merely relabels the risk. And if DeepMind or OpenAI ships a verification product before Meta does, the follower thesis collapses.
The math is real. The market is not yet. A model that closes a theorem has proven one thing only — that search is cheap. It has proven nothing about your vault, your oracle, or your governance. Code does not care about your feelings, and neither, in the end, does the chain.
