AI Progress and the Geopolitical Implications of Chip Export Controls

AI Progress and the Geopolitical Implications of Chip Export Controls

Recent developments in the field of artificial intelligence have brought the role of export controls into sharp focus. The emergence of companies that, in some respects, rival the performance of leading AI models at a lower cost, has sparked debate about the efficacy of current policies and their impact on global AI leadership. Instead of focusing on whether certain countries are a direct threat to other AI companies, it is more helpful to look at whether recent technological releases undermine the case for export control policies on chips.

The Importance of Export Controls

Export controls serve a critical function: maintaining a lead in AI development for democratic nations. This isn't simply about stifling competition; it's about safeguarding technological advantages in the face of potential challenges. It can be argued that handing over advantages to other parties without good reason is inadvisable. AI companies must strive to have better models than others if they want to prevail.

Dynamics of AI Systems

Understanding the following dynamics of AI systems is crucial:

1. Scaling Laws: Scaling up the training of AI systems generally leads to better results across a range of cognitive tasks. The more is invested in training, the smarter the model becomes. These differences tend to have huge implications in practice, and thus companies are investing heavily in training these models.
2. Shifting the Curve: Innovations, whether in model architecture or hardware efficiency, shift the curve, allowing for better performance at a lower cost. New generations of hardware also have the same effect. This shifting of the curve causes companies to spend more, not less, on training models. Algorithmic progress shifts the curve even further. Shifts in the training curve also shift the inference curve, and as a result large decreases in price holding constant the quality of model have been occurring for years.
3. Shifting the Paradigm: Periodically, the underlying thing being scaled changes, or a new type of scaling is added to the training process. Reinforcement learning (RL) to train models to generate chains of thought has become a new focus of scaling. Companies are now working very quickly to scale up the second stage to hundreds of millions and billions, but it's crucial to understand that we're at a unique crossover point where there is a powerful new paradigm that is early on the scaling curve and therefore can make big gains quickly.

Interpreting Recent AI Releases

Recent releases from some AI companies highlight the interplay of these dynamics. While some may point to these advancements as evidence of successful innovation, a closer examination reveals a more nuanced picture. These releases represent an expected point on an ongoing cost reduction curve. What’s different this time is that the company that was first to demonstrate the expected cost reductions was from another country. US companies will soon follow suit — and they won’t do this by copying DeepSeek, but because they too are achieving the usual trend in cost reduction.

The Broader Implications

Despite these advancements, the overarching trend remains: companies are investing increasing amounts in training AI models. Efficiency gains are quickly reinvested into developing even more intelligent systems. This trajectory suggests that achieving generally superhuman AI will require vast resources and is most likely to happen in the next few years.

The critical question is whether other countries will have access to the necessary resources, particularly chips. If they do, the world could see a bipolar AI landscape, with potentially significant implications for global power dynamics. If not, the current AI leaders may maintain a unipolar dominance, at least temporarily.

Well-enforced export controls are the most important determinant of whether we end up in a unipolar or bipolar world.

Conclusion

Developments in AI should not be viewed as a reason to ease export controls. Rather, they underscore the importance of maintaining a strategic advantage in this critical field. The idea that the technology getting more powerful, having more bang for the buck, is a reason to lift our export controls makes no sense at all.

Ultimately, the choices made today regarding export controls will shape the future of AI development and its geopolitical implications. Which path do we want to take?

Comments are closed.