Anthropic CEO on Deepseek and Export Controls
Anthropic CEO Addresses Impact of Export Controls on Chinese AI Development
Dario Amodei, CEO of Anthropic, recently contributed to the ongoing discussion regarding the effectiveness of U.S. export controls on AI chips, particularly in light of the advancements made by Chinese AI firm DeepSeek.
Export Controls and DeepSeek's Progress
Amodei, who previously advocated for stricter export controls alongside former U.S. deputy national security adviser Matt Pottinger, posits that the current restrictions are indeed hindering the progress of Chinese companies such as DeepSeek. He asserts that DeepSeek’s models, when evaluated against those developed in the U.S., demonstrate a performance level trailing behind U.S. counterparts, considering the time elapsed since their respective releases.
“DeepSeek achieved a level of performance comparable to U.S. models that are 7 to 10 months older, and at a reduced cost – though not to the extent some have suggested,” Amodei explained. “[This represents] a predictable point on the continuing curve of cost reduction. The novelty lies in the fact that a Chinese company was the first to showcase these anticipated cost efficiencies.”
Comparative Model Performance
A comparison was drawn between DeepSeek V3, a prominent model from the company, and Anthropic’s Claude 3.5 Sonnet. Amodei estimates the training cost for Sonnet to be “a few $10M’s.”
Despite Sonnet’s training completion occurring 9 to 12 months prior, and DeepSeek’s model being trained more recently in November or December, Amodei highlights that Sonnet continues to outperform in numerous “internal and external evaluations.”
“U.S. companies are also consistently demonstrating the typical trend of cost reduction,” Amodei added. “The efficiency innovations pioneered by DeepSeek will soon be implemented by both U.S. and Chinese laboratories in the training of multi-billion dollar models.”
Future Implications and Policy Considerations
Amodei acknowledges DeepSeek’s team as “very talented engineers” and emphasizes China’s emergence as a significant competitor to the U.S. He anticipates a critical juncture dependent on the export policies adopted by the incoming Trump administration.
Prior to the transition, the Biden administration implemented new hardware export restrictions slated to take effect soon, but these could be modified or reversed under a Trump presidency.
Amodei believes that strengthening export regulations and limiting China’s access to “millions of chips” for AI development could allow the U.S. and its allies to establish a “commanding and long-lasting lead.” Conversely, a relaxation of these controls could lead China to concentrate more resources on “military applications” of AI technologies.
Concerns Regarding Military Applications
“This, coupled with its substantial industrial base and military-strategic advantages, could potentially enable China to assume a dominant position on the global stage,” Amodei warned. “It’s important to clarify that the objective isn’t to deny China, or any authoritarian nation, the substantial benefits in science, medicine, quality of life, and so forth that derive from highly advanced AI systems. Everyone should be able to reap the rewards of AI. The aim is to prevent them from achieving military supremacy.”
Calls for Stronger Measures
During a Senate hearing on Wednesday, Howard Lutnick, Trump’s nominee for commerce secretary, accused DeepSeek of intellectual property theft, suggesting a need for more robust measures.
“Our export controls, without the backing of tariffs, are proving ineffective – akin to a ‘whack-a-mole’ approach,” Lutnick stated. “Chinese tariffs should be set at the highest possible levels.”
As commerce secretary, Lutnick would play a pivotal role in enacting Trump’s plans to impose and enforce tariffs.
OpenAI's Perspective
OpenAI, a primary competitor to Anthropic, has also urged the Trump administration to adopt more assertive strategies to secure U.S. leadership in AI. A recently released policy document from OpenAI cautioned that if the U.S. fails to attract sufficient global investment for AI projects, funds will likely be directed towards “China-backed projects,” thereby “[strengthening] the Chinese Communist Party’s global influence.”
- Key Players: Dario Amodei (Anthropic), DeepSeek, Matt Pottinger, Howard Lutnick, OpenAI
- Central Issue: Effectiveness of U.S. export controls on AI chips
- Potential Outcomes: U.S. leadership or Chinese dominance in AI