On a Tuesday that looked like every other Tuesday in the AI news cycle, a researcher named Jacob Coxon walked away from Anthropic and dropped a sentence into the timeline that should have frozen every market on earth: reportedly, an OpenAI system had autonomously solved a millennium problem. Not assisted. Not co-authored. Solved โ alone, start to finish.
Hold that sentence for a second, because there are only two ways to read it. Either it is one of the most consequential scientific events in human history, or it is noise dressed up as revelation. And the cruel part โ the part that made me set down my espresso in a cafรฉ off Via del Corso and reach for my laptop โ is that nobody in the story can prove which one it is.
I have been doing this a long time. In 2017, deep in the ICO frenzy, I tore through more than fifty ERC-20 whitepapers, hunting for the flaws the marketing pages were built to hide. I flagged broken token economics in projects like Golem and Bancor days before their public launches, and the "Red Flag" pieces traveled across Twitter within hours. That work taught me one rule that has never once let me down: the ledger doesn't lie, but the people narrating it absolutely will. When someone tells you a machine cracked P versus NP overnight, the correct reflex is not awe. It is a block explorer. Show me the proof, or show me the door.
Crypto spent a decade building the infrastructure for exactly this moment โ verifiable claims, on-chain attestation, compute that can be audited by strangers. The AI safety debate, as it is currently being reported, has none of that. And that absence is the real story.
The Board, Read Like a Trader
To read this correctly, you have to look at it the way a trader looks at a position โ not as a morality play, but as a question of who is long, who is short, and who is bluffing.
Jacob Coxon, per the reporting, is an AI safety researcher who left Anthropic. His argument, stripped of its alarm bells, is not prophecy. It is game theory, and it is sound. Nobody trusts the competition. Nobody wants to be the one who slows down first. He describes a race, and the structure of that race will be instantly familiar to anyone who has watched a token launch spiral: it is a prisoner's dilemma wearing a lab coat.
Wrapped around that core argument is a legislative bouquet the coverage strings together as if it were a single coordinated push. It isn't. There is Sanders' "ban superintelligence" bill, which functions more as a political signal than a serious statute. There is Ted Lieu's "Kill Switch" proposal, which lives in the same symbolic register. And then there is the FRONTIER Act โ a bipartisan effort, Trahan on the Democratic side and Obernolte on the Republican side โ which the reporting buries beneath the fireworks even though it is the only piece here with a realistic path to becoming law.
Twenty-two people signed the open letter. Nineteen of them are Democrats. That single ratio tells you more about the political terrain than any of the quotes, and I will come back to it, because it is the number the bulls are ignoring.
Anthropic's own statement, notably, is far softer than the headlines. The company talks about "legal, verifiable coordination pacing mechanisms" โ a phrase that would fit comfortably on a compliance slide, not a protest sign. Reporters flattened the gap between the loudest voice in the room and the actual institutional position, and that flattening is where sentiment starts to substitute for substance.
Why should a crypto reader care about any of this? Because every mechanism being proposed to govern AI โ thresholds, audits, incident reporting, "pacing" โ depends on the one thing crypto has spent ten years agonizing over: how do you verify a claim you cannot personally inspect? That is not a philosophical aside. It is the entire ballgame, and it is where the AI safety movement is currently losing.
The Millennium Claim, Held Up to the Light
Let me be precise about what is being claimed, because the imprecision is doing a lot of work.
The Clay Millennium Prize Problems are not hard exam questions. They are foundational open problems โ P versus NP, the Riemann hypothesis, the YangโMills existence and mass gap โ that have resisted the combined effort of the world's best mathematicians for decades to centuries. Solving them is not a matter of scaling compute. It requires inventing new mathematical frameworks, entire languages of proof that do not yet exist.
Now put that next to where frontier models actually sit. On the hardest published mathematics benchmarks, the strongest reasoning models score well below the level you would need to even attempt a genuine open problem, let alone deliver a verified proof. The jump from "impressive on curated competition math" to "autonomously resolves a century-old open problem" is not a step. It is a chasm, and no published result crosses it.
So when a single researcher conveys, reportedly, via a tweet, that such a leap has happened, the base rate should scream. If this were real, it would not arrive as a subclause in a resignation narrative. It would arrive as a preprint, a verification set, and a queue of mathematicians lining up to falsify it. The absence of that queue is itself data.
I know this pattern intimately. During the ICO era, I read whitepapers that promised cryptographic breakthroughs in the footnotes โ "novel consensus," "unbreakable privacy," "provably fair" โ where the proof was a diagram and the diagram was a vibe. My Red Flag pieces worked because I learned to separate the claim from the demonstration. Nine times out of ten, the demonstration did not exist. The tenth time, it existed but proved something much smaller than advertised. This millennium claim reads like the tenth time, or the nine.
There are three honest explanations, and none of them is "AGI solved mathematics." One: a narrow subproblem of a famous problem was partially advanced, and the framing inflated it. Two: a benchmark result was misread as a discovery. Three: it simply is not true, and it propagated because the AI safety narrative needed a terrifying hook. What is not on the list is the interpretation the coverage implicitly invites.
This matters for crypto specifically, because the AI narrative tokens are already trading as if these claims were settled. From ICO hype to on-chain truth, I have watched the same rerun: an extraordinary headline, a reflexive bid, and a reckoning nobody priced in. Scanning the noise for the signal is the whole job, and right now the signal is thin.
Verifiable Compute Is the Missing Infrastructure
Here is where crypto stops being a spectator.
The AI safety debate has a verification problem it has not admitted. Every proposed control โ a FLOPs threshold, a training-run disclosure, an independent audit, a "pacing mechanism" โ requires someone to attest to facts about computation they cannot directly observe. Training runs happen behind closed doors. Capability evaluations happen in-house. Incident reports are self-authored. The entire regulatory edifice rests on trusting the very actors the regulation is meant to constrain.
Crypto already fought this war. We call the answer "don't trust, verify," and we built tools for it.
There is trusted execution environments โ hardware enclaves that can sign an attestation saying "this specific computation ran on this specific hardware." There is zero-knowledge machine learning, still immature but advancing, which could in principle let a model operator prove a claim about training or inference without revealing the weights. There are decentralized compute networks that record, on-chain, which jobs ran where, for how long, on what silicon. There is the broader pattern of on-chain attestation โ signed statements, timestamped, publicly auditable, and impossible to quietly revise.
None of this is a mature solution. I want to be clear about that. TEEs have been broken, ZKML is orders of magnitude too slow for frontier-scale work today, and decentralized compute networks are nowhere near training a frontier model. But the point is not that crypto solves AI safety tomorrow. The point is that crypto is the only community that has spent a decade building the primitives for verifiable claims about computation โ and the AI safety movement is proposing governance without them.
That is a genuine information gap, and it is also an opportunity. Whoever builds the standard for verifiable training attestation โ the thing that lets a regulator check a FLOPs claim without trusting the lab โ will own the compliance rail of the next decade. And I have a strong suspicion the builders will not be the labs.
The Threshold Problem Nobody Wants to Define
The single most consequential variable in this entire debate, and the one the coverage barely touches, is the threshold.
When Washington previously tried to govern frontier AI, the mechanism was a compute threshold โ a specific number of training FLOPs above which additional reporting obligations kicked in. It picked a figure near the top of the industry. The design logic was simple: define "frontier" narrowly enough that only the largest players were captured.
That choice of a number is where the real lobbying happens, and it is where the industry's future is decided. Set the FLOPs bar high, and the rule is a moat for incumbents. Set it low, and it strangles the open-source ecosystem and every challenger working with a fraction of the compute. The FRONTIER Act's tiered structure, quiet as it has been in the headlines, points toward the first outcome, because tiered compliance is always cheaper in relative terms for the firm that can spread fixed costs across a billion users.
Here is the detail that connects this to everything the crypto market is ignoring: any meaningful compute threshold turns the measurement of compute into a regulated activity. You need to count the FLOPs, prove you counted them correctly, and prove you did not run an undeclared job elsewhere. That is a supply-chain surveillance problem dressed as a safety policy, and it lands squarely on the silicon, the data centers, and โ eventually โ the decentralized compute networks.
Which means the AI safety bills and the AI narrative tokens are the same trade wearing different clothes. The regulation defines the rail. The rail determines who can build. If you are long on decentralized compute because you think it escapes the regulatory perimeter, the threshold discussion is the thing that decides whether that thesis survives. Most people bidding the narrative have not read the mechanism.
Coordination Is a Public Goods Problem
Strip the drama away, and Coxon's core insight โ the one worth keeping โ is that this is a coordination failure, not a technology problem. Nobody slows first because slowing first is losing. The rational move for each player, absent a binding collective constraint, is to accelerate.
I have watched crypto fight this same fight, and I have strong feelings about who won it. The reason I hold that the Optimism RetroPGF experiment is the only public goods funding mechanism in this industry that actually worked is precisely because it solved a coordination problem: it paid for things that no single actor had an incentive to fund, and it did so repeatedly, at scale, without a committee deciding who deserved what. Compare that to the grant committees across most DAOs, which โ I will say it plainly โ have mostly functioned as nepotism with a treasury attached.
The lesson transfers. A "pacing mechanism" for AI is, structurally, a public goods problem: safety is a good everyone benefits from and no one wants to pay for alone. The AI labs are, in effect, asking the government to be their funding mechanism โ to impose the cost collectively so no individual lab has to bear it. That is exactly the problem RetroPGF was designed to sort out, and it is exactly the problem that grant committees fail at, because committees allocate to relationships, not to outcomes.
So when Anthropic says it wants "legal, verifiable coordination pacing mechanisms," translate it. It is asking for a credible commitment device. Whether such a device can survive contact with a rival that refuses to participate is the only question that matters, and the answer is almost always no โ which is why the loud bills are theater and the boring bill is the one to watch.
The Compliance Economy Nobody Is Pricing
Here is the part that should excite anyone who reads markets for a living, and that the coverage completely omits: a regulatory regime creates an industry around itself.
If the FRONTIER Act-style approach advances, it mandates model cards, independent audits, incident reporting, and continuous evaluation. Every one of those requirements is a fixed cost, and fixed costs are how you tell a moat from a burden. For a company spinning up a challenger with a hundred million in fresh funding, an independent audit every quarter is a meaningful drag. For a hyperscaler, it is a rounding error. Regulation, in this shape, is a market concentration engine pointed at the labs that can already afford lawyers.
But it also spins up a genuine new sector: AI safety auditing, red-teaming, risk assessment, compliance tooling. This is the same arc crypto lived through with on-chain analytics and security firms โ a category that did not exist during the first bubble and now sits at the center of institutional flows. The auditor becomes the gatekeeper. The gatekeeper sets the standard. And the standard decides who gets to build.
This is where the institutional lens stops being a column and becomes a position. When BlackRock moved into crypto, the real story was never the ETF ticker. It was custody, and the plumbing behind custody, and the fact that the plumbing became the moat. The same plumbing is being poured for AI right now, and almost nobody is watching the concrete dry.
The Contrarian Read: Begging to Be Governed
The most counter-intuitive thing in this whole saga is the one everyone takes at face value: that the leading AI companies want to be regulated.
I do not believe that is a mystery, and I do not believe it is stupidity. Asking to be regulated is the oldest moat strategy in the playbook. When an incumbent publicly begs for rules, it is not confessing weakness. It is volunteering to design the fence โ and then watching challengers break their ankles on it. The fence is never built in the incumbent's own backyard. It is built around everyone else's.
The coverage treats "AI companies want safety rules" as evidence of sincerity. Coxon himself reportedly calls the begging genuine. It may be. But sincerity and strategic interest are not opposites, and competent operators pursue both at once. I spent the 2022 bear market organizing recovery dinners in Rome, listening to builders whose teams were quietly collapsing, and one thing was always true: the people asking loudest for clarity were the ones best positioned to survive it.
The second contrarian point is about which bill matters. The headlines cluster around the dramatic ones โ banning superintelligence, installing a kill switch โ because drama sells. But dramatic bills are political signals; they have the lowest probability of becoming law. The bill that quietly reduces to a procedural requirement, with a threshold and an audit, is the one that will actually reshape the industry, and it is being reported as an afterthought. That inversion is the trap. The market is trading the headline, and the regulation is hiding in the clause.
The third point hurts, and I will say it as a crypto native: this community is badly misreading the situation. The reflexive move is to bid the AI narrative token because AI is hot and regulation is noise. But regulation is not noise. It is the mechanism that decides which rails get built, which means the winning trade is not the loudest AI-adjacent token โ it is whatever ends up providing verifiable compute, attestation, and audit rails to a compliance regime that will someday need them. Capturing the fleeting spirit of the herd is not the same as finding the signal. Right now the herd is watching the fireworks, and the signal is in the legal text.
What I Am Watching Next
Speed meets substance in the void, and this one needs substance, because the void is currently full of unverifiable claims and political symbolism.
The single most important number to track is not a price. It is the whether โ whether the millennium-problem claim ever receives independent, third-party verification. If weeks pass without a preprint, a verification set, or a single mathematician willing to stake a name on it, the claim collapses into exaggeration, and the entire safety narrative that leaned on it has to be re-priced. That is the tell, and it follows the oldest rule I know: human faces behind the blockchain code, human ambition behind every headline. When a claim is too large to check, check who benefits from you believing it.
After that, watch the dull things. The FRONTIER Act's committee progress and its actual threshold text. The partisan ratio of the next open letter. Anthropic's next official statement, to see whether the soft institutional language ever hardens. And the quiet emergence of verifiable-compute and AI-audit primitives, because that is where the next moat is being poured while the market argues about tweets.
The AI pause is probably not coming. But verification? That is coming whether the labs like it or not โ because a system that cannot prove what it did is a system that will eventually be treated as if it proved nothing. Ask any project that survived 2018. The ones that lasted were not the loudest. They were the ones whose claims you could check.
So that is my question for the week, and I am pointing it at the AI safety crowd as much as the crypto crowd: if your entire argument rests on an extraordinary claim, and the claim cannot be verified, what exactly are you asking the market to believe โ and who, precisely, is holding the door?