Liu Shengyu, a DeepSeek machine-learning systems engineer associated with the company’s V4.1 model, warned on September 15, 2026, that Anthropic controlling the most advanced AI or artificial general intelligence (AGI) could create a dangerous concentration of power. His warning is hypothetical and centers on access and control—not on a claim that Anthropic currently controls AGI.
The dispute matters because it combines two different fears. Liu is concerned that a small group of private companies could control the most powerful systems. Dario Amodei, Anthropic’s CEO, is focused on the possibility that AI capabilities could improve faster than safety research, evaluation and oversight.
Liu’s warning is about concentration and access
Liu said he does not trust Anthropic or OpenAI to make frontier AI sufficiently open and affordable. He supports making cutting-edge intelligence available to everyone on those terms, and compared the consequences of hypothetical Anthropic control of advanced AI with Nazi Germany acquiring atomic-bomb technology before the Allies.
That analogy describes Liu’s political and personal judgment. It does not establish that AGI exists, that Anthropic controls it, or that Anthropic intends to cause harm.
DeepSeek is associated with the open-model alternative Liu defends, but openness by itself does not establish a safety advantage. The argument is instead about who gets access to powerful systems and whether control remains concentrated among a small number of companies.
Amodei’s proposal is about speed and supervision
Amodei has argued that AI capability improvements should slow enough for safety work to keep pace. His use of “pacing” does not mean halting model training or technical progress. He proposes buying time for alignment, testing, interpretability and operational security.
His framework 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.
- Coordination among democratic countries: frontier companies would work toward shared safety standards and limits on unchecked progress.
- Global coordination: democratic and authoritarian governments would be brought into discussions, with verification safeguards treated as a central concern.
Amodei also points to recursive self-improvement—the possibility that AI systems help build more capable successors—as the risk that most worries him. He has said that gaining an extra year or two before systems reach critical capability levels could give alignment research more time to reduce the chance of a serious failure.
Two related risks, but not the same argument
| Actor | Main concern | Preferred response |
| Liu Shengyu and DeepSeek | A small number of companies could control the most advanced AI, limiting access and concentrating power. | Open and affordable access to cutting-edge intelligence. |
| Dario Amodei and Anthropic | Capability growth and recursive self-improvement could outpace safety work and oversight. | Slower capability gains, embedded evaluators, shared standards and international coordination. |
The positions overlap around governance, but they point in different directions. Liu emphasizes distribution of power. Amodei emphasizes control mechanisms around development. An open model may broaden access, yet that fact alone does not show that it is safer than a more closed system.
Why Coxon and Hubinger raised the stakes
The debate unfolded alongside warnings from people connected to Anthropic. Jacob Coxon, a former Anthropic researcher who had also worked at OpenAI, resigned after three years and criticized both companies for racing toward self-improving superintelligence.
Evan Hubinger, Anthropic’s alignment-science lead, personally estimated that AI causing human extinction within the next decade had a probability greater than 10%. That figure is his own assessment, not a measured probability or an industry-wide forecast. Hubinger distinguished the relatively low risk he associates with present models from the future risk of superintelligence emerging through recursive self-improvement.
Coxon’s warning and Hubinger’s estimate reinforce Amodei’s concern about capability growth, while Liu’s intervention adds a different question: who gets to control the systems if they become extraordinarily powerful?
What Amodei means by embedded evaluators
Embedded evaluators would be independent teams with continuing access to a company’s relevant systems and processes. Their role would include checking safety practices, reporting incidents and assessing models and training pipelines from inside the organization rather than through occasional external reviews.
The proposal has drawn support and criticism. Sam Altman has backed the need to pace frontier development, while Clement Delangue has argued that transparency can make AI safer. Investor Chamath Palihapitiya has criticized the proposal as a possible way to stop open-source development and concentrate technological and economic power with Anthropic.
That disagreement mirrors the larger dispute. Safety rules can limit dangerous capability growth, but the design of those rules also determines who has influence over the field.
