Hook
A regulator is an oracle.
It ingests claims from the world, compresses them into a state variable — compliant or not — and downstream agents price that variable into their behavior. Everything we know about oracle design therefore transfers intact. Feed latency. Reporter incentives. Corruptibility of the trusted set. The gas cost of being wrong.
On September 14, year unspecified in the wire copy I pulled, Barack Obama urged Democrats to prioritize AI regulation, warning that without urgent action and a clear plan the technology carries danger. That is the entire item. No venue. No named policy instrument. No enforcement mechanism. No year.
Most of the crypto press ran it as politics. I read it as a systems specification, because a regulatory agenda is a specification. And this one is missing its load-bearing member: the measurement layer. Not the agency. Not the statute. The thing that produces the numbers the statute keys off.
Regulating a system you cannot measure is not regulation. It is a press release with a budget line.
Context
The policy history defines the state space, so start there.
October 2023: Executive Order 14110 signed, keying reporting obligations off a training compute threshold — 10^26 floating-point operations — and requiring red-team results for frontier models, with NIST producing an AI Risk Management Framework around it. August 2024: the EU AI Act enters into force, tiering obligations by risk class. September 2024: California's SB 1047 clears the legislature and sits in a veto window; it is vetoed on September 29. January 2025: EO 14110 is rescinded. July 2025: the federal posture flips to an AI Action Plan oriented around acceleration.
Fifteen months. Precautionary to deregulatory. The regulatory regime is a protocol with an admin key held by whoever wins the election, and it has been upgraded twice.
This is why the missing year in the wire item is not pedantry. If the quote is September 2024, it lands seven weeks before a US presidential election, EO 14110 is live, SB 1047 is undecided, and "prioritize AI regulation" is a mainstream position inside the governing party. If it is September 2025, EO 14110 is dead, the federal posture is explicitly deregulatory, and the same sentence is a faction talking to itself. Same words. Orthogonal meaning. Any analysis that does not mark this is guessing with confidence.
Obama's own position is not new. In 2016 his administration published "Preparing for the Future of Artificial Intelligence" and "AI, Automation, and the Economy." A decade of consistency tells you the speaker is not reacting to a news cycle. It does not tell you a mechanism has been specified. Politicians rarely specify mechanisms. That is the analyst's job.
Core
Start with what a regulator must actually ingest.
A rule that says "frontier models must not be trained above X compute" requires a claim: this training run used Y FLOPs, and Y < X. A rule that says "deployers must assess dangerous capabilities" requires a claim: this red-team search found no uplift in CBRN, cyber-offense, or autonomous replication. A rule that says "systems must be safe" requires a claim that nobody currently knows how to produce.
Every one of those is an attestation — a statement about a computation, produced by an entity that is not the regulator. The regulator does not observe the computation. It observes reports about the computation. In systems terms, the regulator is a consumer of an oracle feed, and whether AI regulation works reduces entirely to the design of that feed.
Three feed architectures exist.
Self-report. The producing entity publishes its own numbers. Trusted reporter, node set of one. Cheap, fast, and the only architecture currently deployed at scale — EO 14110's reporting obligations were essentially this plus a government channel. The feed is as honest as the least honest reporter, and the reporters are competitors whose payoffs are maximized by under-reporting. I have watched this topology before. An oracle with five node operators who all sit on the same board is a centralized feed wearing a multisig costume, and the market prices it as decentralized right up until the day it isn't.
Federated audit. Third-party evaluators, accredited labs, a permissioned set. This is where NIST-adjacent frameworks and the EU AI Act's conformity-assessment model land. It fixes the single-reporter problem and inherits the accreditation problem: who accredits the accreditors, and what happens when the auditor's revenue depends on the audited.
Trust-minimized attestation. The claim is bound to the computation cryptographically. The regulator verifies a proof instead of trusting a narrator. This is the only architecture that survives adversarial incentives, and it is the one that does not exist.
Trust is a vulnerability, not a virtue. Anything that depends on it gets exploited by whoever finds the cheapest violation.
Now the technical reality check, because this is where the policy conversation goes soft.
You cannot inspect a weight tensor for intent. Interpretability is not at the level where a regulator reads parameters and observes deception, and it will not be for years. So any workable attestation must be process-based — commit to the compute, the data provenance, the evaluation results — rather than outcome-based. That has a consequence the safety literature under-prices: a process-based gate is a resource gate. A FLOP threshold selects for capital, not for caution.
Proof-of-work taught us exactly what happens to resource gates. Thresholds concentrate. Mining pools, ASICs, industrial operations with compliance departments. Set a capability threshold at 10^26 FLOPs and you have not created a safety filter; you have created an optimization target with a magic constant inside it. Every contract I have audited that keyed off a ratio instead of an absolute value got exploited at the boundary. Boundary conditions are where the math is thin, and a magic constant in a statute is a boundary drawn before anyone measured the actual distribution.
What would the trust-minimized layer actually look like? I have spent most of the last decade on the primitives, so let me be specific rather than aspirational.
Inference attestation. Prove that output y is the result of running committed model M on input x. This is zkML, and it works — for small models. Proof overhead relative to inference cost is still large enough that frontier-scale inference proofs are a research problem, not a product.
Training attestation. Prove that weights W are the output of a training run over dataset D with hyperparameters H. This is the claim that matters, because it is the claim every compute threshold depends on. It also does not exist in any usable form. Proving a 10^25 FLOP training run requires proving the arithmetic of the forward and backward passes and the optimizer state transitions. The proof-to-computation overhead is still orders of magnitude. This is not a 2025 problem. It is not obviously a 2030 problem.
Provenance commitments. Hash the weights, sign the model card, publish the dataset manifest. Cheap, essentially solved, and almost universally skipped.
So the honest inventory is this: the regulator has no feed, the feed it needs is against an unsolved problem, and the interim architecture is the labs reporting on themselves.
Here is where the bull market is actively misallocating, and where I part company with most of the AI-plus-crypto narrative. Capital is flowing into decentralized compute marketplaces and agent tokens. Compute is a commodity; GPU hours are fungible and the market has already priced them competitively. The scarce primitive is not decentralized compute. It is verifiable compute — the binding of a claim to a computation such that verification does not require trusting the party that ran it. A decentralized GPU marketplace without verification is cheaper cloud with worse SLAs. It solves price. The governance problem is trust.
The industry keeps mistaking the two. Privacy is a protocol, not a policy — you do not bolt it on at the terms-of-service layer, you build it into the proof system or you don't have it. Safety has the same shape. A voluntary "responsible scaling policy" published by a lab is DAO governance with a multisig, and I have traced enough team wallets to know how that story ends. The foundation publishes the commitment. The multisig has four signers. Two of them are the same person's cold wallets. The commitment is a compliance shield, and the shield is doing exactly the job it was designed to do: absorbing reputational risk without changing the underlying computation.
Contrarian
The counter-intuitive position, and the one that will annoy both camps: regulation without verifiability is worse than no regulation, because it launders legitimacy.
A compliance stamp converts an unverifiable claim into a marketable one. I have watched this exact dynamic in smart contract auditing. A protocol pays for a two-week review, receives a PDF, and mints a badge. The badge becomes the fact. Users price "audited" as "safe," the audit firm's incentive is to keep the client, and the vulnerability ships anyway. Audit theater is not a crypto-specific pathology. It is what happens whenever the artifact produced is a certificate rather than a proof.
Apply that to AI. A frontier lab files a safety report. The filing becomes the fact. A regulator without the technical capacity to evaluate it endorses it, because endorsement is cheaper than evaluation and the alternative is admitting the state cannot measure the thing it is regulating. Twelve months later the deployment is in production, the stamp is on the model card, and nobody can reconstruct the reasoning behind the claim.
There is a second blind spot that neither camp prices, and it is structural. Enforcement requires evaluation. Evaluation requires frontier compute. The only entities with frontier compute are the entities being evaluated. This is circular, and it is not fixed by appropriating a larger budget. You cannot audit what you cannot run, and if you run it through the audited party's infrastructure, you have not audited anything. Any statute that does not resolve the compute asymmetry is a statute about paperwork.
Takeaway
The question is not whether AI gets regulated. It is the ordering.
If the rules arrive before the measurement layer, you get audit theater and a compliance moat, and the moat entrenches exactly the incumbents the rules were written to constrain. If the attestation layer arrives first, the rules can be thin and the enforcement can be real, because a regulator with a verifiable feed does not need to trust anyone's narrative.
Every invocation of "urgent action" and "a clear plan" is downstream of a technical question nobody has answered: can you prove a claim about a training run? Today, no. The proof overhead numbers are the leading indicator. Not the speeches.
Math doesn't care who wins the election. Watch the circuit sizes.