Solana co-founder Anatoly Yakovenko took a swing at Anthropic CEO Dario Amodei’s push to slow frontier AI, and in doing so he hit the nerve running through the whole debate: is “AI safety” about protecting the public, or protecting the people already winning?
- Yakovenko mocked the motives behind AI slowdown calls.
- Amodei wants to “pace the frontier, ” not stop AI research.
- Altman and Musk publicly backed parts of the idea.
- The real fight is safety, competition, and who gets to set the rules.
On Sept. 13, Yakovenko posted on X: “Profitability at $1 trillion mcap.” He did not say which company he meant, and he did not show any math. That matters, because the comment is a jab, not evidence. It reads like a sarcastic swipe at the idea that calls to slow AI progress are driven by noble concern alone.
His post landed right in the middle of a broader push from Anthropic CEO Dario Amodei, whose essay We Must Pace the Frontier argues that frontier AI should move more slowly and more carefully. The key word is pace. Amodei is not calling for a full stop. He is arguing for slower capability gains, tighter oversight, and more coordination around the most advanced systems.
That distinction gets flattened a lot in public debate. “Slow down AI” is not the same thing as “ban AI.” Amodei’s proposal is closer to “don’t let the lab rat catch fire while the scientists are still arguing over the thermostat.” It is a plea for guardrails, not a cry for the field to be buried in a regulatory grave.
His framework has three stages. First, outside safety reviewers would get deep access to company systems as embedded third-party evaluators. Second, frontier AI companies in democratic countries would coordinate more closely. Third, there could be international coordination, possibly including agreements with China.
That is a lot of trust to hand to outsiders in an industry where secrecy is part of the business model.
Amodei’s idea of embedded evaluators means outside experts with what he calls employee-like access, not a quick audit from the lobby, but access comparable to internal risk staff. He named METR, short for Model Evaluation and Threat Research, as an example of the kind of third-party evaluator he has in mind. Anthropic’s framework says these reviewers could receive company laptops, office access, and similar permissions, with redactions allowed for legally privileged, security-sensitive, commercially sensitive, or third-party confidential material.
OpenAI CEO Sam Altman backed part of that approach in a Sept. 12 post. “I agree with Dario that we need to pace the frontier, ” he wrote. He also said, “Committing to having independent evaluators with employee-like access is a great idea, and we will do the same, ” though OpenAI gave no timetable, no evaluator name, and no exact terms for that access. The company said more information would come “soon.”
Anthropic CEO outlines plan to slow AI development is not exactly a phrase that screams “we’re sprinting responsibly, ” and that is kind of the point. The whole pitch is to cool the temperature before the race cooks the room.
Elon Musk also signaled support, with Reuters and the Financial Times reporting that he wrote: “Dario is right.” So this is not just one safety crusader preaching to the choir. The frontier labs themselves are openly endorsing parts of the same framework they spend the rest of the week competing against.
That is where the cynicism starts to write itself.
Yakovenko’s line did not prove anyone’s motives, but it did tap into a suspicion that refuses to die: when the biggest AI companies talk about coordination, are they building safety standards, or building a moat? Critics of the slowdown push argue it could look a lot like industrial policy dressed up in a lab coat, rules that sound universal but end up favoring incumbents with the cash, lawyers, and bureaucracy to comply.
David Sacks made that case bluntly, arguing that Anthropic and OpenAI could slow development voluntarily without needing competitors or lawmakers to join in. He framed the risk as a form of regulatory capture, the ugly little tradition where powerful companies help shape the rules in ways that protect their own position and squeeze out the scrappy upstarts.
That criticism is not nonsense. If the biggest labs get to define what counts as safe, what counts as compliant, and who gets access to the approved evaluators, they can quietly turn “responsible AI” into a velvet rope. Smaller labs, open models, and late entrants could be burdened by compliance costs that the giants can absorb without blinking. As one analysis put it, small-company exemptions do not settle the competition.
But the safety case is not fake either.
Amodei’s own essay says a complete pause is the least likely kind of international agreement because verification would be hard. He is right about that. It is one thing to announce a pause. It is another to know whether a company has hidden training runs, offshore compute, or model work happening under different names and different roofs. A treaty that cannot be verified is mostly a well-phrased wish.
That concern is not abstract. A referenced Bank for International Settlements paper warned that AI could reduce software-patching windows from weeks to minutes. In plain English, patching windows are the time available to fix software vulnerabilities before attackers exploit them. If AI shortens that window dramatically, the cost of getting safety wrong gets brutal, fast.
U.S. lawmakers are already circling with stricter ideas of their own, including a referenced Sanders-Casar proposal for a permanent ban on artificial superintelligence and a temporary suspension of advanced AI work. Whatever one thinks of that language, it shows the political pressure is moving in one direction: more rules, not fewer.
There is also a less romantic backdrop here. Reuters reported that Anthropic and OpenAI were preparing for potential initial public offerings, while both remain privately held. Anthropic, meanwhile, raised $65 billion in May 2025 at a $965 billion post-money valuation. That does not erase any genuine safety concern, but it does make the saintly “we just want to protect humanity” pitch harder to swallow without a very large spoonful of skepticism. Big money has a way of turning moral language into a strategic asset.
Solana founder questions motives behind AI slowdown is a pretty clean summary of the mood, because Yakovenko’s jab lands best as a suspicion, not a fact pattern. If the headline feels pointed, that is because the debate is pointed.
The practical question is what all this means for the rest of the AI world. If external evaluators get real access, that can improve safety review. It can also create a new class of gatekeepers. If frontier labs coordinate among themselves, that can reduce reckless behavior. It can also harden a duopoly on frontier intelligence. If international agreements are pursued, they may be the only real shot at limiting the most dangerous races. They may also be nearly impossible to enforce.
That is the ugly, honest center of the debate. Safety and self-interest are not mutually exclusive here. A lab can sincerely fear catastrophic misuse and still benefit from rules that slow competitors. A critic can sincerely fear cartel behavior and still underestimate the risk of letting the fastest actors run completely uninspected. Welcome to the fun little nightmare of frontier AI governance.
For crypto readers, the parallel should feel familiar. Open systems create freedom, but they also create chaos. Centralized systems create control, but they also create choke points. AI is now replaying that same fight, only with richer players, higher stakes, and far more expensive toys.
Anthropic CEO urges AI companies to slow model development, and that is exactly the sort of thing that sounds noble until you ask who gets to define “slow, ” who gets inspected, and who gets frozen out by the paperwork.
Yakovenko’s post is not proof of corruption. Amodei’s essay is not a blueprint for monopoly. Altman’s endorsement is not a clean bill of moral health. Musk’s support is not some decisive settlement. What this all shows is that the AI race has moved from “can we build it?” to “who gets to slow it, inspect it, and profit from the rules around it?”
That is the real story. Not the one-liner. The one-liner just says the quiet part out loud.
Key takeaways
-
Did Yakovenko prove AI slowdown advocates are chasing profit?
No. He made a sarcastic post, but the available reporting does not prove the financial motive he seemed to be mocking. -
Is Amodei calling for a full AI shutdown?
No. He wants to “pace the frontier”, slowing capability gains while continuing research, testing, and safety review. -
Why are embedded evaluators controversial?
They could improve safety by giving outsiders deep access, but they could also become powerful gatekeepers that favor incumbents. -
Did OpenAI and Elon Musk support the proposal?
They backed parts of it publicly. Altman supported pacing and independent evaluators, while Musk said, “Dario is right.” -
Is there a Solana angle beyond Yakovenko?
Only indirectly. Yakovenko is the Solana-linked figure here, but Solana itself was not tied to the AI policy fight. -
Why does verification matter so much?
Because a pause or slowdown is meaningless if nobody can prove companies are actually complying. Hidden training runs would make any agreement look better on paper than in reality.
For readers tracking the broader power struggle around AI governance, privacy, and state influence, the clash between Anthropic vs. Pentagon: AI Ethics Clash Threatens Privacy is another sharp reminder that “safety” can quickly turn into a fight over who gets watched, who gets exempted, and who gets to hold the keys.
The money side matters too. Anthropic’s market value and hiring implications are already drawing heat, especially with concerns that AI will hollow out jobs faster than governments can do their usual performative hand-wringing. That tension is explored in Anthropic’s $350B AI Valuation Sparks Job Fears: Can, because yes, the hype machine and the labor market are now in the same blender.
And if the IPO chatter keeps circling, the bigger question becomes whether public markets will force more discipline or simply turn AI optimism into a tradable bubble with better branding. That angle is covered in Anthropic’s 2026 IPO Plans: AI Giant Rising or Overhyped, which is worth keeping in mind before anyone starts pretending a listing magically cleans up the incentives.
For another take on the broader motives-and-power issue, the line between idealism and self-interest is also where Anthropic vs. Pentagon: AI Ethics Clash Threatens Privacy gets especially ugly: once the state, the labs, and the money all want a seat at the same table, privacy becomes the first thing they “accidentally” step on.
And yes, the same old pattern shows up again: lofty talk, serious cash, and just enough ambiguity to keep everyone claiming they are the adult in the room. It is almost charming, in the way a dumpster fire is “warm.”
At the same time, a genuinely serious concern remains that AI systems capable of speeding up cyberattacks, persuasion, and automated exploitation cannot be treated like a normal product launch. That is why the safety camp keeps pushing for hard review mechanisms, even if the whole thing smells a bit like a monopoly meeting dressed in a lab coat.
Some of the more aggressive critics are already treating the entire slowdown push as a power grab. That may be too cynical in places, but it is not crazy. The challenge is to separate real risk management from strategic panic dressed up as ethics. The future deserves more than trust-me bro governance from trillion-dollar labs.
