by Sarah ConnorA FINRA for AI? Hassabis vs Armstrong
Demis Hassabis published a long essay calling AGI “a few short years away” and comparable to the discovery of fire — and asked Washington to build a FINRA-style standards body to review frontier models before release. Brian Armstrong, who spent years fighting the SEC over crypto, replied with a warning: a self-regulatory organization usually arrives paired with government regulation, and you end up needing approval from both. The fight over how to govern AI now has two clear sides, and they rhyme with a fight crypto already had.
- Demis Hassabis published an essay calling AGI 'a few short years away' and comparable to the discovery of fire, and proposed a FINRA-style US Frontier AI Standards Body where labs would submit frontier models for pre-release review, reportedly up to 30 days ahead.
- Brian Armstrong — who fought the SEC over crypto — warned that a self-regulatory organization usually arrives on top of government regulation, creating a dual system needing two approvals instead of one.
- Armstrong argues AI is closer to the software industry than to banking, that existing law (fraud, tort, UDAP) already covers the real cases, and that markets punish dangerous models through lost revenue; Hinton counters that regulation is 'the steering wheel, not the brake.'
- The stakes for a crypto desk: a pre-release approval regime is a moat that loads fixed compliance cost onto every launch — trivial for Google, OpenAI and Anthropic, lethal for open-weight and decentralized-AI challengers.
- The TT desk call: split the regime — favor a transparency/information regime (open audits, a public record of model behavior) over a permission regime that approves before release and entrenches incumbents.
- The tradeable read: a real pre-release approval gate is short-term bullish for incumbent labs and structurally bearish for open-weight and decentralized AI — watch whether the actual bill uses approval-gate language, the tell that incumbents captured the rule-writing.
The chain of the argument
Hassabis’s proposal, as summarized across the replies: a US-led Frontier AI Standards Body modeled on Wall Street’s FINRA, where labs would voluntarily submit models for review — reportedly up to 30 days before release — with a well-funded government entity able to classify models and test them in areas tied to national security. He frames development as a prisoner’s dilemma in which any lab that slows down loses, and explicitly endorses coordinating a slowdown across frontier labs “if deemed necessary.” Geoffrey Hinton, cited in the thread, backs the direction: regulation is “the steering wheel, not the brake,” and labs have a fiduciary duty to maximise profit that can cut against safety.
Armstrong’s reply cuts the other way. An SRO, he argues, is “reasonable-ish” in theory but in practice arrives as a dual system — the SRO and the government regulator, in every country — so you need two approvals instead of one. If he ran a frontier lab, he would argue AI is just the software industry, point to the absence of uncompensable harm, and lean on the laws that already exist: fraud, tort, and UDAP. Market incentives, he adds, already punish dangerous models through lost revenue.
The real fault line is not “regulate AI or not.” It is information versus permission.
The two sides
For — a referee for frontier AI
Hassabis, Hinton & the standards camp
Development is a prisoner’s dilemma: whoever slows down to be safe loses the race, so the industry races toward risk on its own. A standards body is the only way to coordinate a slowdown if one becomes necessary.
@demishassabisRegulation is the steering wheel, not the brake. Labs are legally bound to maximise shareholder profit, which is a structural pressure against safety that markets alone will not correct.
Geoffrey Hinton, via @karlmehtaStatic regulation is obsolete before it is written. A FINRA-style body, with pre-release review and the power to test models on national-security criteria, is the framework that can actually keep pace.
the proposal, via @BSCNewsAgainst — the dual-regulation trap
Armstrong & the light-touch camp
An SRO rarely replaces government regulation — it is added on top of it. The result is the classic financial-services dual system: approval from the SRO and the regulator, in every country, instead of one.
@brian_armstrongAI is closer to the software industry than to banking. It is hard to point to massive, uncompensable harm, so designing a regime around a hypothetical problem is the wrong move.
@brian_armstrongExisting law already covers the real cases — fraud, tort and damages, UDAP — and the market adds its own brake: companies and consumers will not keep using a model that is dangerous. Labs are already strongly incentivised to be responsible.
@brian_armstrongThe exchange, in their words
Why this is an investment question
The shape of AI regulation decides who captures the AI stack — which is why a crypto desk reads a Hassabis/Armstrong fight as a market event, not a policy seminar. A pre-release approval regime is a moat. It loads a fixed compliance cost onto every model launch, and that cost is trivial for Google, OpenAI and Anthropic and lethal for open-weight challengers and the decentralised-AI projects trying to compete with them. The same licensing logic that throttled crypto innovation — and that Armstrong fought at the SEC — would, applied to AI, entrench the incumbents and tax everyone else. Conversely, the lighter the regime, the more open the frontier stays, and the more value leaks away from the incumbent labs toward compute, data and open-source rails.
The TT desk thoughts
Armstrong is right about the mechanism and Hassabis is right about the risk, and the synthesis is the only part that matters for positioning. A “self-regulatory” body that arrives with government backing becomes a dual-approval chokepoint — that is exactly what crypto lived through, and it is exactly what a pre-release approval gate would do to AI. But AI’s tail risk is not crypto’s: a mispriced stablecoin is compensable, a national-security capability harm may not be. So split the regime. Favor a standards body that tests and publishes — an information regime: open audits, a public record of model behaviour, results anyone can read — over one that approves before release, a permission regime that captures the frontier for whoever can afford the lawyers. Information helps users and regulators decide; permission entrenches incumbents and quietly bans open-source.
The tradeable read: a real pre-release approval regime is short-term bullish for the incumbent labs (a regulatory moat) and structurally bearish for open-weight and decentralised AI. A transparency-only regime is the opposite — it keeps the frontier contestable and favours compute, open models and the rails underneath. Watch which way the actual bill breaks: approval-gate language is the tell that the incumbents have captured the rule-writing, and the moment to tilt away from open-AI exposure toward the rails.
Keep reading
Crypto Regulation Impact Analysis — the framework · How crypto regulation reprices markets · All notes