In the ever-evolving landscape of Artificial Intelligence (AI), recent developments by Chinese AI company DeepSeek have sparked discussions about the competitiveness of US AI firms and the effectiveness of current export control policies. This article delves into DeepSeek's achievements, their implications for AI development, and underscores the vital importance of export controls in maintaining a competitive edge for democratic nations.
Export controls aren’t merely about stifling competition; they’re a crucial tool for ensuring that democratic nations remain at the forefront of AI development. While competition drives innovation, it's essential not to inadvertently hand technological advantages to authoritarian entities like the Chinese Communist Party.
To grasp the significance of DeepSeek's achievements and the role of export controls, it's vital to comprehend three fundamental dynamics of AI systems:
Scaling laws dictate that, all things being equal, increasing the scale of AI system training leads to better results across a range of cognitive tasks. The more resources invested in training, the more adept the AI becomes at solving complex problems.
Innovations, whether architectural improvements or hardware upgrades, can shift the AI performance curve. A "compute multiplier" effect means that the same level of performance can be achieved at a lower cost. These efficiencies lead to greater investments in training smarter models. As Dario Amodei's team highlighted in their research from 2020, algorithmic progress significantly shifts this curve. The gains in cost efficiency are reinvested into training more intelligent models.
Occasionally, fundamental changes occur in the AI training process. Reinforcement Learning (RL) has emerged as a new scaling focus. Companies like Anthropic, DeepSeek, and OpenAI are using Chain-of-Thought prompting that is trained with RL to train models that exhibit greater reasoning and problem solving capabilites. Given that RL is still very early in its adoption, gains are achieved more cost effectively than for a larger pre-trained model.
DeepSeek's recent model releases, particularly "DeepSeek-V3" and "R1," provide valuable insights into the current state of AI development.
While DeepSeek's achievements are noteworthy, it's crucial to contextualize them. Claims that DeepSeek is doing for $6 million what costs US AI companies billions are inaccurate. For example, Claude 3.5 Sonnet, a mid-sized model from Anthropic, cost tens of millions to train and remains ahead of DeepSeek in many evaluations.
The ongoing trend in AI development involves companies investing heavily in training increasingly powerful models. Efficiency innovations are quickly adopted, and cost gains are reinvested into creating even smarter AI, underscoring the importance of export controls. According to research, the cost of training frontier AI models will likely increase to billions of dollars by 2027.
In 2026-2027, we could face two starkly different scenarios:
Well-enforced export controls are the only way to prevent China from acquiring millions of chips. These regulations will therefore be instrumental in determining whether we end up in a unipolar or bipolar world.
DeepSeek's achievements, enabled by access to a moderate number of chips, highlight China's capabilities. Furthermore, the necessity to smuggle some of them emphasizes the export controls are effective. Loopholes are being closed over time, thereby preventing China from getting a full fleet of top of the line chips. By closing loopholes more efficiently, the US can increase the likelihood of a unipolar world.
While DeepSeek's engineers are undoubtedly talented, their allegiance to an authoritarian government necessitates vigilance. Export controls are a critical tool for safeguarding national security. The rising capabilities of Chinese AI companies like DeepSeek underscores that US export controls are not something to roll back, but something to double down on to protect US national interests.
