On September 15, 2026, Liu Shengyu, a DeepSeek machine-learning systems engineer, warned that hypothetical control of the most advanced artificial intelligence—or artificial general intelligence (AGI)—by Anthropic could create a dangerous concentration of power. The warning, connected to a wider debate on September 16, was not a claim that AGI already exists or that Anthropic controls it. It was Liu’s political judgment about who should control systems that do not yet exist in the form he described.
That distinction matters. The headline-grabbing comparison was hypothetical; the governance dispute is real.
Liu’s warning is about concentration and access
Liu reportedly compared a future in which Anthropic controlled the most advanced AI or AGI with Nazi Germany acquiring atomic-bomb technology before the Allies. The analogy belongs to Liu, not to an established assessment of Anthropic’s intentions or capabilities.
His alternative is straightforward: cutting-edge intelligence should be available openly and affordably. Liu said he did not trust Anthropic or OpenAI to provide that kind of access. As a DeepSeek engineer associated with the company’s V4.1 model, he is also speaking from a particular position in the competition over how advanced AI should be developed and distributed.
The argument is about power as much as software. If a small number of private companies controlled the most capable systems, their decisions could shape access to intelligence, the rules governing its use and the direction of future development. Liu presents broader access as a way to resist that concentration.
That does not make open AI automatically safer. Openness and safety are different questions, and Liu’s warning does not provide a comparative safety evaluation of DeepSeek and Anthropic.
Amodei’s proposal is about speed and supervision
Dario Amodei, Anthropic’s chief executive, has argued for slowing the pace at which AI capabilities improve so that alignment, evaluation, operational security and oversight can keep up. “Pacing” does not mean stopping model training or halting technical progress. It means creating more time for safeguards as systems become more capable.
Amodei’s proposal has three parts:
- Embedded evaluators: independent third-party teams would receive ongoing, employee-like access to relevant tools, workspaces and processes inside frontier-AI companies. Their job would be to assess models and training pipelines, verify safety practices and report incidents.
- Coordination among democratic countries: frontier companies would work toward common safety standards and limits on unchecked progress.
- Broader international coordination: democratic and authoritarian governments would be brought into discussions, with verification treated as a central challenge.
Amodei’s central concern is not primarily who owns the most advanced system. It is whether capability growth—especially the possibility of recursive self-improvement, in which AI helps build the next generation of AI—could move faster than safety work.
The proposal has drawn support and criticism. Sam Altman agreed that the frontier should be paced, while Clement Delangue of Hugging Face argued for greater transparency. Investor Chamath Palihapitiya interpreted the proposal differently, warning that it could stop open-source development and concentrate technological and economic power with Anthropic. That disagreement mirrors Liu’s concern, even though Amodei presents his plan as a safety measure.
Two risks that overlap—but are not the same
| Actor and organization | Primary concern | Preferred response | What the statements establish |
| Liu Shengyu, DeepSeek | Concentration of advanced AI in a small number of private companies; limited access and affordability | Open and affordable access to cutting-edge intelligence | Liu’s reported warning and political position about control and access |
| Dario Amodei, Anthropic | Capability growth and recursive self-improvement outpacing safety and oversight | Slower capability progress, embedded evaluators, shared standards and international coordination | A published proposal for pacing AI development, not a plan to stop it |
The two positions meet around governance but propose different remedies. Liu worries about who gets to control advanced intelligence. Amodei worries about how quickly capability advances and whether safeguards can keep pace. Treating those as one argument obscures the trade-off at the center of the debate: wider access may reduce the power of a single company, while additional oversight may require giving selected institutions more access to the companies building the systems.
Neither position, by itself, settles the safety question.
Why Coxon and Hubinger entered the debate
The dispute also unfolded alongside warnings from people connected to Anthropic. Jacob Coxon resigned after three years at the company and criticized the race among frontier-AI firms to build self-improving systems. He had previously worked in pretraining research at OpenAI and Anthropic.
Evan Hubinger, Anthropic’s alignment-science lead, said he personally estimated the chance of AI causing human extinction within the next decade at more than 10%. That figure is Hubinger’s personal assessment, not a measured probability, an industry consensus or an established prediction. His comments distinguish the relatively low risk he associates with present models from the longer-term danger he links to recursive self-improvement.
Those warnings strengthen the case for taking the safety debate seriously, but they do not turn a scenario into a forecast. A statement about what advanced AI could do is not proof that the capability already exists.
What the controversy actually tells us
Liu Shengyu’s warning documents a dispute over the political shape of frontier AI: whether advanced systems should be concentrated inside a few companies or made broadly accessible. Amodei’s proposal addresses a different pressure: whether development can slow enough for evaluation and alignment work to keep up.
The most important conclusion is also the least dramatic. The controversy does not show that Anthropic controls AGI, that AGI has been achieved, or that DeepSeek’s approach is inherently safer. It shows that the AI industry is arguing over two connected questions—who should hold advanced capability and how much time society should demand before capability moves further—with no agreement yet on how to resolve either one.
