Anthropic researchers and Wall Street investors engaged in a heated debate over whether AI should be subject to “centralized control” or “decentralized power.” The core disagreement lies in: as AI becomes more dangerous, does it actually require more centralized management, or should the concentration of power be avoided even more?
On August 14, a confrontation on X brought the core philosophical divergence of the AI industry to the surface.
On one side was Sholto Douglas, a post-training researcher at Anthropic and a key technical figure responsible for reinforcement learning. On the other side was Gavin Baker, founder of Atreides Management, a tech hedge fund manager overseeing billions of dollars. He continuously delivered systematic frameworks on AI infrastructure investment on the All-In Podcast this year and is regarded as one of Wall Street’s most AI-savvy investors.
The trigger was a statement Baker made on the All-In Podcast. He said he heard from multiple trusted sources that Anthropic CEO Dario Amodei had stated internally that Anthropic might ultimately become “the only private company left in the world.” Baker’s exact words were: “If this is true, I strongly suggest Dario stop saying things like this to anyone.”
Douglas directly refuted this on X: “Completely false. I like Gavin’s takes, but whoever told him this is lying, with the aim of making you believe a narrative that some people are desperate for you to believe.”
Then he added a sting: “The same people will tell you in one sentence that Anthropic has no moat, and in the next that it might be powerful enough to be dangerous. These two judgments cannot both be true.”
Baker did not push back hard. He replied thanking Sholto for clarifying the facts, but pointed out the bigger issue: “Many very serious people in Silicon Valley have indeed heard some version of similar claims, and they believe it because it aligns with Dario’s public statements.”
The true value of this confrontation lies not in whether Dario actually said those words. It lies in the fact that it ignited a deeper divergence regarding whether AI should be centralized or decentralized, and whether safety ultimately means control or competition.
Baker’s Framework: AI Is Too Dangerous to Be Centralized
Baker delivered the cleanest sentence of the entire debate on the All-In Podcast:
“People in Anthropic and the effective altruism movement believe this technology is too dangerous to be widely distributed. Whereas Zuckerberg, Elon, and Jensen believe this technology is too dangerous to be centrally controlled.”
Baker’s logic chain is as follows: Anthropic believes AI could be extremely powerful and dangerous, thus requiring prudent regulation and centralized control, with a few “responsible” labs to gatekeep. But Baker believes this logic will ultimately lead to one result: setting entry barriers through regulation, a few companies gaining de facto monopoly, open source being restricted, and competition being weakened. History has repeatedly proven that concentrated absolute power has never produced good results.
He analogized AI to gun rights: Americans have the right to bear arms, even though guns can be abused, because this right serves both defense and resistance against tyranny. Baker said he wants an AI that serves him personally, an AI aligned with his own values, rather than an AI aligned with what Dario Amodei thinks is “best for humanity.”
Baker strongly supports open-source models. Citing Dwarkesh Patel’s viewpoint, he pointed out that Anthropic’s Claude, because it has a built-in “Constitution” and alignment system, operates at the underlying logic according to what Anthropic thinks is correct. He doesn’t think the people at Anthropic are bad, but he believes it is dangerous to let any single company define “what is best for all humanity.”
He even gave a counter-intuitive investment judgment: the rise of open-source models does not necessarily harm the value of frontier labs. His analogy is the Manhattan Project. Oppenheimer and Feynman needed thousands of ordinary scientists to execute work; the elites became more valuable because they had more deployable “intellectual labor.” Similarly, open-source models (“150 IQ”) can be orchestrated and scheduled by frontier models (“250 IQ”), open source does volume, frontier does value.
Douglas’s Counterattack: The Market Is Solving Everything You Worry About Through Competition
Douglas’s rebuttal did not get entangled in specific gossip but directly attacked the logical foundation of Baker’s narrative.
He linked Dario Amodei’s long essay, then wrote: “The AI market is the most competitive market on earth right now, every one of the world’s largest companies is going all out to provide you with smarter, cheaper models. If all goes well, we will succeed in reducing the cost of everything down to the cost of energy. This is amazing, but it threatens many people’s old moats. They are scared.”
The subtext of this passage is rich. Douglas is saying: Baker and those behind him spreading the “Anthropic wants to monopolize the world” narrative are largely stakeholders whose old moats are threatened by AI. They hype Anthropic’s “ambition” and “regulatory capture,” with the true purpose of building momentum for open source, serving their own investment portfolios, or purely out of fear of change.
Douglas also pointed out a core contradiction in Baker’s argument: If you think Anthropic has no moat and will be crushed by open source, then you shouldn’t worry about it monopolizing the world at the same time. If you think it will become powerful enough to be dangerous, then you can’t say it lacks competitiveness. These two judgments are mutually exclusive.
In another tweet, Douglas wrote: “Indeed, if AGI is realized, capitalism will become very strange, the concept of ‘company’ might look completely different.” He linked Dwarkesh Patel’s article on corporate forms in the AI era, implying this issue is far more complex than the simple dichotomy of “centralized vs decentralized.”
A True Question That Was Avoided
The brilliance of the debate between Baker and Douglas lies in the fact that neither actually made obvious logical errors; their divergence lies in underlying assumptions.
Baker’s underlying assumption is: AI is powerful enough to be weaponized, therefore decentralized power is safer than centralized power, just as democracy is safer than enlightened autocracy. He holds a realistic attitude towards human nature: No one, no matter how well-intentioned, should possess unchecked superintelligence.
Douglas’s underlying assumption is: The intensity of competition in the AI market is itself the best check-and-balance mechanism. There is no need to “decentralize” AI power through political narratives because the market is already doing this. Google, Meta, xAI, OpenAI, DeepSeek, Kimi, GLM, the world’s largest companies and smartest teams are all rushing in the same direction. Under this competitive landscape, no single company can possibly “become the only private company in the world.”
Both avoided a sharper question: If AI really, as Dario said, could lead to 50% of white-collar jobs disappearing, then are “competition” and “decentralization” enough? Competition ensures efficiency, but it does not guarantee fairness. Decentralization ensures no single point of failure, but it does not guarantee systemic risk won’t occur.
Baker said he wants an AI that “serves him personally.” But if everyone has a superintelligence serving themselves, what will the world become? This is no longer a technical problem, or even a business problem, it is a political philosophy problem.
And on this issue, the divergence between an Anthropic researcher and a hedge fund manager may exactly map the most fundamental anxiety split humanity faces regarding AI: Does humanity fear a gatekeeper that is too powerful more, or a world without a gatekeeper more?
There will be no winner in this debate, but it at least makes both fears clear simultaneously.