On a humid morning in late March 2023, I was sitting in a co-working space off Nanjing Road in Shanghai, halfway through auditing a cross-chain bridge, when a colleague slid her phone across the table. On the screen was a headline: AI leaders pause development. She looked at me the way people look when they think history just turned a corner. "They actually stopped," she said. "The smartest people in the world agreed to stop."
I remember the exact feeling in my chest โ that strange vertigo of wanting something to be true so badly that you almost refuse to check. I did check. I opened the actual open letter, scrolled past the headline, and read the sentence that the headline had quietly amputated: an open letter calling on all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4. Not "paused." Not "will pause." Calling on. A petition. A plea. A strongly worded email with 1,100 signatures attached to it.
The gap between "called for a pause" and "paused" is not a semantic quibble. It is the entire story โ and almost nobody who shared the headline that week ever went back to close it. That gap is what I want to dissect here, because I have spent the better part of a decade watching a smaller, pettier version of the same deception play out inside crypto โ and the lessons transfer with almost embarrassing precision.
The Letter, and What It Actually Asked For
The document itself deserves more respect than the coverage gave it. It was published by the Future of Life Institute, a nonprofit that had been quietly funding AI-safety research for years before safety was fashionable. Its signatories โ and this is the part the headlines flattened โ included Elon Musk, Steve Wozniak, Yoshua Bengio, Stuart Russell, and a long tail of researchers who genuinely do lose sleep over alignment. The letter's argument was not Luddite. It was almost boringly reasonable: frontier labs are in a race, races compress safety thinking, so let's agree to a coordinated six-month freeze on the training of systems more powerful than GPT-4, during which governments and labs can build shared safety protocols and verification standards.
Notice how many load-bearing qualifiers are packed into one sentence. Training, not deployment. More powerful than GPT-4, not all AI. Six months, not forever. Coordinated, not unilateral. And crucially โ the thing the headline erased โ calling on, not committing to.
That last distinction is the whole ballgame. "Calling on" is the grammar of a petition, and a petition is a piece of social technology, not a piece of infrastructure. It has no enforcement layer. There is no slashing condition. There is no oracle that detects a violation and triggers a penalty. It is a smart contract with an empty require statement. And yet, for about a week in the spring of 2023, the crypto-adjacent corner of the internet โ my corner โ treated it as if the chain had halted.
Why does this matter to someone who spends his days thinking about DAOs and decentralized identity rather than transformers and RLHF? Because the exact same failure mode โ mistaking a norm for a mechanism โ is the disease that has hollowed out governance in this industry. I have watched grant committees convene, pass resolutions, and dissolve with nothing changing on-chain. I have watched "community votes" that were ratification theater, where the multisig had already decided before the forum thread was posted. The AI pause letter was crypto's governance theater performed on a bigger stage, with better production values.
Context: Why a Blockchain Person Was Watching AI at All
Let me be honest about the bias I'm bringing. I am not an AI researcher. My master's is in applied mathematics, and my working obsessions are game theory, cryptographic commitment schemes, and the question of how communities coordinate without a central authority. I came to this letter not as a machine-learning person but as a governance person. What I saw in it was a beautiful, tragic test case of the same problem that has defined my career: how do you make a promise credible when the promiser has every incentive to break it?
By early 2023 I had already spent two years in the wreckage of the FTX and Celsius collapses, publishing a series I called Anatomy of a Collapse. That work taught me one thing above all: the failures that matter are almost never technical. They are incentive failures dressed up as technical accidents. Celsius didn't fail because the code broke. It failed because a charismatic human held keys he was never meant to hold, and the structure around him let him hold them. FTX didn't fail because of a bug. It failed because trust was centralized in a person pretending to be a protocol.
So when the world's most powerful AI labs signed a piece of paper promising restraint, my trained reflex was not "how inspiring" but "who holds the keys, and what happens if they simplyโฆ don't?"
And here is the uncomfortable symmetry. The AI safety argument and the decentralization argument are usually treated as rivals. The AI-safety crowd says: this technology is too powerful to be permissionless. The decentralization crowd says: any technology powerful enough to threaten you is a technology where concentration of control is itself the threat. I think both camps are half-right, and the missing half is exactly the half that the pause letter implied but never built.
Core Analysis: The Anatomy of a Pause That Never Halted
The Signature Economy
Let me start where the coverage started โ the names โ because the list of who signed is more revealing than any argument in the letter's body.
Elon Musk signed. Steve Wozniak signed. Begnio signed. Russell signed. But look at who is not on that list with any decision-making authority over a frontier training run. Sam Altman's signature is absent. Demis Hassabis's is absent. The heads of production at OpenAI, Google DeepMind, Anthropic, and Meta's FAIR group either declined or were never going to sign. The letter's gravitational center was intellectuals, investors, and founders of adjacent companies โ not the people who actually order the GPU clusters.
This is not a knock on the signatories. It is a structural observation, and it is the single most important thing I took from the episode. A call to pause has zero binding force on anyone who did not sign, and the people who most needed to be bound were precisely the people who did not sign. The letter was, in game-theoretic terms, a coordination proposal aimed at players who had correctly calculated that defecting was dominant. A frontier lab that unilaterally froze for six months while its competitors did not would burn hundreds of millions of dollars in compute and lose the talent that follows momentum. No rational agent in that position freezes. The letter was a beautiful piece of values-first reasoning addressed to a commons that had already been enclosed.
I have seen this exact dynamic inside DAOs. A working group drafts a governance proposal urging treasury diversification. It circulates. It gathers endorsements from respected community members. And then nothing happens, because the mandate has no teeth โ the multisig signers were never persuaded, or the whales simply don't vote, or the proposal's author quietly leaves for a better-funded protocol. The endorsement count is a vibes metric. It is not the mechanism.
No Enforcement Layer, No Threshold, No Oracle
Suppose, for the sake of argument, that every frontier lab had genuinely wanted to comply. How would anyone have verified it? This is where the letter's engineering gaps become glaring, and where my applied-math brain started taking notes.
A meaningful pause needs four components, and the letter specified none of them. First, a metric: what exactly counts as "more powerful than GPT-4"? Is it measured in training FLOPs, benchmark performance, parameter count, emergent capability thresholds? The letter gestured at capability but never operationalized it. Second, a verification oracle: who runs the audit that confirms a lab did not secretly train past the threshold? Training runs are private, compute is private, and there is no equivalent of a block explorer for a secret dataset. Third, a penalty: what is the cost of cheating, and who levies it? A letter has no slashing condition. Fourth, an exit condition: what does the world look like after six months, and who certifies it?
Without all four, a "pause" is a mood, not a mechanism. It is the difference between a promise and a covenant.
This is where the AI-safety conversation and the crypto conversation stop being rivals and start being the same conversation wearing different hats. The reason proof-of-work worked is not that miners were honest; it's that dishonesty was more expensive than honesty, and everyone could verify the work. The reason RetroPGF โ Optimism's retroactive public-goods funding โ has a fighting chance where every other grant committee I've ever audited withers into nepotism is that the evaluation is downstream of observable shipped work. You cannot retroactively claim credit for something you didn't build, because the artifact is on-chain. The mechanism binds the participants to the claim.
Now hold that next to the pause letter. The claim โ "we will not train past a threshold" โ is unobservable, unverifiable, and unenforced. It is the inverse of a proof. It is an assertion of good faith in a domain where good faith is exactly the thing that gets arbitraged away.
What Actually Happened, in Numbers
Here is the part that closes the case, and it is the part I bring as a person who has spent a decade watching narratives diverge from data.
The six-month pause window โ roughly April through October 2023 โ was not a period of hesitation. It was one of the most frantic periods of frontier model release in the industry's short history. In that window and the year that followed, we got GPT-4 Turbo, then GPT-4o, Google's Gemini line, Meta's Llama 3, Anthropic's steadily compounding Claude releases, and a multilingual flood of open-weight models. Whatever the signatories intended, the training did not stop.
And the infrastructure tells the story even more plainly than the models. NVIDIA's data-center revenue, which is the closest thing we have to a public, hard-to-fake ledger of how much frontier training is actually happening, went from a single-digit-billion quarterly figure in early 2023 to a demand curve so steep that delivery times for H100-class accelerators became a months-long queue. Capital expenditure at the hyperscalers climbed into the tens of billions per year. If a real pause had occurred, these numbers would have flinched. They did the opposite.
I want to be precise about what this proves and what it doesn't. It doesn't prove the signatories were insincere โ many, individually, were entirely sincere. It proves something more interesting and more useful: a values declaration that lacks a mechanism is, in practice, indistinguishable from a marketing document, and the market treats it exactly that way. The market did not read the letter as a stop sign. It read it as a press release.
The Rebranding Parallel: Bitcoin Layer 2s and the Honor System
I have written before, and will keep writing, about a phenomenon that the pause episode clarified for me: the way narratives get wrapped around things that don't match them. My long-standing irritation here is the swarm of projects calling themselves "Bitcoin Layer 2s" that are, under the hood, Ethereum-side systems with a Bitcoin sticker applied to the marketing. The technical substance โ the bridge assumptions, the security model, the sequencer authority โ is Ethereum-shaped. The branding is Bitcoin-shaped. And crucially, the real Bitcoin community, the people who actually run nodes and think about the base layer's security guarantees, mostly do not acknowledge these projects as Bitcoin at all. The label and the mechanism are decoupled.
The AI pause letter belongs to the same genus. The label โ "AI leaders pause development" โ described a governance event that did not occur. The mechanism โ a voluntary petition with no enforcement โ was nothing like the label implied. A reader who only saw the label would conclude that the frontier had frozen. A reader who checked the mechanism would conclude that nothing had changed.
This is why I keep hammering on verifiability as the first principle, not a nice-to-have. In crypto we learned the hard way that "trust me" scales terribly. In AI safety, the entire frontier is still running on "trust me" โ and calling it a pause.
The Fragmentation Lesson: Slicing a Fixed Pie
There is a second parallel that the pause episode surfaced, and it concerns the word "scaling."
My recurring complaint about the current Layer-2 landscape is that we now have dozens of rollups and sidechains all competing for substantially the same pool of users and liquidity. That is not scaling; that is slicing an already-scarce resource into thinner and thinner fragments. Each new chain multiplies the surface area for bridges, the number of liquidity silos, and the cognitive load on the user, while the underlying user base barely moves. The narrative says "abundance." The data says "fragmentation."
Now look at the AI safety world through that lens. Every lab publishes its own safety framework. Every lab has its own red-team process, its own interpretation of "aligned," its own threshold for what's too dangerous to deploy. There is no shared measurement, no interoperable standard, no common oracle. It is dozens of safety frameworks over a single scarce resource โ public trust โ and each framework further fragments what could have been a shared verification layer.
The pause letter's stated purpose was to unify โ to buy time so that a common safety protocol could emerge. What actually emerged was more fragmentation, because fragmentation is what you get when coordination is voluntary and voluntary coordination is not anyone's dominant strategy. The DAO world knows this in its bones. It is why bridge standards took so long to congeal, and why every "universal" governance framework fractures the moment a whale finds it inconvenient.
The Thing That Actually Binds: Verifiable Humanity and the Truth Layer
I don't want to leave this as pure critique, because critique without a constructive direction is just cynicism with better vocabulary, and I have spent too many bear-market nights refusing to become a cynic.
Here is where I think the AI-safety conversation and the crypto toolkit genuinely converge, and it is the work I actually do now. The deepest risk from advanced AI is not that a model will develop a secret desire to escape โ that's the cinematic version. The deepest risk, the one that is already here and compounding, is epistemic: an internet where you can no longer tell whether a human or a model produced a given piece of content, a given identity, a given claim. Deepfakes are not a future threat; they are a present flood. And once the cost of fabricating a convincing human collapses to near zero, the thing that becomes scarce and valuable is proof of origin.
This is why I co-founded a community initiative I called Verifiable Humanity, and why we onboarded thousands of users to blockchain-based identity primitives that let a person prove they are a distinct, persistent human without doxxing themselves. It is not a silver bullet. It is a mechanism โ and that is the entire point. Whereas the pause letter asked labs to promise, verifiable identity proves. It puts a cryptographic commitment where a promise used to sit. In an AI-dominated information environment, the blockchain becomes the truth layer โ not because blockchains are magic, but because they are the only widely deployed infrastructure that makes certain kinds of claims expensive to fake.
There is a neat inversion here that I find almost poetic. The AI-safety movement's instinct was to slow down the machines through a voluntary pledge. The decentralization movement's instinct is to speed up verification by making the machines' outputs attributable to accountable origin. One approach asks power to restrain itself. The other builds a structure in which restraint is enforced by physics-like mechanism rather than by virtue. After watching the pause evaporate, I know which one I trust.
Contrarian Angle: The Letter Worked โ Just Not the Way Anyone Said
Now let me argue against myself, because the most dangerous thing I can do is build a tidy narrative and stop stress-testing it.
There is a strong case that the pause letter was a genuine and substantial success โ as a rhetorical device, not as a freeze. Notice what happened in the eighteen months that followed its publication. The EU's AI Act advanced from proposal to concrete, tiered regulation. The United States issued an executive order on AI that put frontier-model reporting and safety-testing obligations into the federal bloodstream for the first time. The UK convened an international AI Safety Summit. China implemented interim measures on generative-AI services. In other words, the letter's actual ask โ buy time to build safety protocols and regulatory frameworks โ arguably got its time, and the frameworks did start to appear.
That reading reframes the "failure" as a category error on my part. Maybe the letter was never meant to stop training. Maybe it was meant to shift the Overton window so that regulation became thinkable, and by that metric it punched far above its weight. A petition is not an enforcement layer โ but a petition is an excellent agenda-setting layer, and I may have been too quick to dismiss the value of agenda-setting because my biases run toward mechanisms.
Here is the sharper, more uncomfortable version of the contrarian point, though. If the letter's real function was agenda-setting, then its greatest beneficiaries may not be the public but the incumbents. A regulatory regime with reporting thresholds, safety audits, and compliance costs is, structurally, a moat. Large labs can afford compliance teams; a two-person open-source collective cannot. An incumbent that has already trained its frontier model benefits from a world in which future training is expensive and scrutinized. I am not accusing any signatory of bad faith in proposing this โ motivations are mixed, multiple, and mostly private. I am simply observing that the outcome of "safety regulation" and the outcome of "entrenching the leaders" look suspiciously similar from the outside, and this ambiguity is exactly why values declarations can't be taken at face value.
I have seen this in DAOs too. The "safety" framing โ protecting the treasury, protecting the community โ is the most reliable rhetorical cover for entrenching a small group's control. The proposal always sounds noble. The mechanism always consolidates power. This is why I trust RetroPGF's structure over any committee's stated intentions: retrospective funding rewards demonstrated contribution regardless of who you are, which is hostile to entrenchment in a way that prospective grants and voluntary pledges simply are not.
So which is it โ did the pause fail because it lacked teeth, or succeed because it lacked teeth but moved the agenda? Both, and the deeper point is that the ambiguity itself is the problem. A mechanism that can be read as either a noble failure or a cynical success is a mechanism that no one can hold accountable. That is precisely why I keep insisting that the first question about any governance proposal โ AI, DAO, or otherwise โ is not "is it good?" but "what happens if someone ignores it?"
Takeaway
The pause never happened. Not because the people who asked for it were lying, but because they designed a promise where the moment demanded a mechanism. The lesson I carry, and the one I'd hand to anyone building safety infrastructure for the AI era, is that the credibility of a commitment is a function of how expensive it is to fake โ and we have spent two decades, in this industry, learning exactly how to make certain claims too expensive to fake.
The frontier will not pause itself, and it does not need to. What it needs is a truth layer: verifiable compute, attributable identity, output provenance that survives the flood of synthetic content. The next time a headline tells you the most powerful actors in a field agreed to restrain themselves, open the document. Read the verb. Ask who holds the keys. And then ask the only question that has ever mattered โ what happens when they don't?
The answer to that question is the architecture we still have to build. Not a letter. A covenant.