Anthropic chief executive Dario Amodei has answered criticism that his approach to AI policy would concentrate power in a few companies, saying the debate is being framed the wrong way.
The dispute followed comments from Atreides Management founder Gavin Baker, who said, according to an earlier Chain News report, that multiple trusted sources had told him Amodei had at one point said internally that Anthropic could someday become 「the world’s only private company」. Baker described that as an 「Anthropic maximalist」 worldview, one in which the most extreme outcome leaves only Anthropic and governments.
Baker said the issue goes beyond the exact quote
Baker argued that the key question was not simply whether Amodei had literally said Anthropic might become the only private company on earth. What mattered, in his view, was that several very serious people in Silicon Valley had heard some version of the idea and found it believable.
He said that credibility came from the fact that the claim did not appear inconsistent with Amodei’s public worldview. Amodei has repeatedly warned about AI risks tied to cybersecurity, biotechnology, loss of control, and concentration of power, and has called for stricter governance and oversight. Baker said he does not deny those risks and even believes Amodei is acting in good faith.
Two competing approaches to AI safety
Baker reduced the current AI safety debate to two paths.
- One view holds that AI is so dangerous that the strongest systems should be kept in the hands of a small number of supervised companies and governments through regulation.
- The other view holds that because AI is so dangerous, it should not be controlled by a small number of companies or political figures and should instead be distributed as widely as possible.
Baker clearly backed the second position. Citing Meta CEO Mark Zuckerberg, he said that if one concludes AI is dangerous enough that 「the only safe path is extreme concentration of power」, that conclusion is itself suspect.
He argued that a better answer is a world with more AI, where different models, companies, and value systems check one another instead of placing the strongest systems inside a handful of frontier labs.
Baker said Amodei’s regulatory case has not won the field
Baker then aimed his criticism at Amodei’s policy line more directly. He said it was now 「safe to say Dario has lost this debate」 because the strict AI regulation Amodei has advocated over time has not become the industry’s main direction.
Baker also said recent events had strengthened the political legitimacy of open-source AI and made it harder to move tougher regulation forward in the near term.
His sharper concern was that Amodei’s repeated emphasis on AI danger could backfire. Baker pointed to growing political opposition to data centers in the United States and predicted that anti-data-center groups could eventually use videos of Amodei warning that AI might harm humanity as campaign material.
That, Baker said, creates a contradiction: warnings meant to produce a safer AI industry could end up serving as arguments against the infrastructure that industry needs.
Baker still said he remains positive on Anthropic
Baker made clear that criticizing Amodei’s regulatory philosophy did not mean he was bearish on Anthropic. He called the company an astonishing business with deep competitive advantages.
He added that a few months ago he had viewed one of Anthropic’s biggest risks not as OpenAI or Google, but as the possibility that Amodei’s forceful rhetoric about AI risk could eventually lead governments to consider 「nationalizing」 Anthropic.
From Baker’s standpoint, if Amodei is set to lead one of the world’s most important technology companies, he should speak more actively about AI’s positive value as well.
Amodei called the framing a false binary
Responding directly on social media, Amodei rejected Baker’s attempt to cast the issue as a choice between regulating and concentrating AI on one side and opening and distributing AI on the other.
He said that was a false binary. In his view, Silicon Valley often treats regulation as identical to regulatory capture and concentration, but real institutions are more complicated than that. Good rules, he said, can constrain the power of large companies rather than entrench it.
To explain the point, Amodei used the court system as an example. Formal judicial systems can look bureaucratic and even elitist, he said, yet compared with a model built entirely on crowd power, they can do more to protect weaker individuals. Institutions do not always centralize power; well-designed institutions can move power away from individuals and toward rules.
Anthropic says frontier labs should face higher compliance costs
Amodei said Anthropic has deliberately tried to avoid policy proposals that would let regulation become a moat for large AI companies. Instead, he said, the company wants frontier AI firms to bear heavier regulatory costs while easing restrictions on smaller competitors.
He pointed to California’s SB 53 and the earlier SB 1047, both of which included thresholds tied to revenue or model training costs that could exempt smaller companies. He said Anthropic’s recent push for model testing follows the same logic: models closer to frontier capability should face tougher testing requirements, while systems behind the frontier should face fewer limits.
Amodei argued that this kind of structure is actually worse for frontier labs such as Anthropic and OpenAI, while giving more room to challengers and open-weight models.
Amodei said scaling laws are the real source of concentration
The deepest split with Baker sits there. Baker worries regulation will create concentration. Amodei said advanced AI already has concentration built into it because of scaling laws.
As model capability rises, he said, so do the needs for compute, data, chips, and capital. If the cost of training the most advanced systems keeps climbing, the number of companies able to compete at that level naturally shrinks. Open model weights can only address part of that problem.
He said the organizations that can run those models at scale will still be the large technology companies or hardware suppliers that control GPUs, data centers, and capital.
For Amodei, the real policy task is to design institutions that can do three things at once:
- reduce cybersecurity, biological, and loss-of-control risks from AI;
- limit the power of frontier AI companies;
- preserve competitive space for smaller firms and open-weight models.
He disputed the claim that the regulatory path has failed
Amodei also pushed back on Baker’s argument that the regulatory camp has already lost. He said some policy directions now being discussed by the U.S. government are close to what he supports, including testing the most advanced AI models before deployment and testing open-weight models as they begin approaching frontier capability.
He also said he supports Google DeepMind CEO Demis Hassabis’s proposal for a FINRA-like AI regulatory body. Compared with six months ago, when the technology sector broadly opposed state-level regulation and the federal government lacked a clear plan, Amodei said the policy environment has already shifted.
Risk messaging and public distrust
The second major disagreement between the two men concerns Amodei’s public posture on AI. Baker said Amodei has leaned too heavily on the dangers and is now hurting the AI industry itself. Amodei disagreed.
He said he has tried to keep a relative balance between AI’s risks and benefits. He pointed to his essay Machines of Loving Grace, which he wrote because he felt the AI industry had not offered a compelling enough picture of the future. In that essay, he predicted AI could help cure most human disease in roughly 5 to 10 years.
Amodei did acknowledge that the public now holds a high degree of distrust toward AI. Still, he said the core issue is not that AI companies talk too much about risk. The deeper problem is a broader crisis of trust around technology, business, and government, with many people suspecting that new technology will end up benefiting only a small minority.
Promises are not enough
That is why, in Amodei’s view, the answer is not more upbeat AI marketing. He said that telling people 「AI can cure cancer」 has, for many, turned into a cliché. What would actually change public opinion is curing cancer for real.
He added that the most reasonable criticism of Anthropic and the wider AI industry right now is that AI companies have made enormous promises without yet producing matching results.
Amodei said Anthropic is rapidly increasing its investment in biotechnology and healthcare. He said he hopes major results will emerge over the next few years, and that some early signs could appear in the next few months. Until those results arrive, he said, he does not want to market unfulfilled promises and does not plan to stop discussing the real risks AI may bring.