Azeem Azhar's Exponential View China’s catching up to US AI… Here’s why it won’t matter
71 snips
May 14, 2025 
Lennart Heim, a researcher at RAND Corporation specializing in AI and geopolitics, discusses why China's AI advancements won't eclipse the US's position. He emphasizes the crucial role compute power plays in AI development, detailing the economic impacts of training models. The conversation also covers the geopolitical dynamics of AI compute, debating the balance between centralized and decentralized resources. Finally, he touches on the importance of global collaboration in AI standards to navigate the complexities of technological competition.
AI Snips
Chapters
Transcript
Episode notes
Rise of Inference Compute
- Inference or test-time compute is rapidly growing as AI models become interactive and reason iteratively.
- This compute type is increasingly costly but crucial for advanced capabilities like step-by-step reasoning.
Compute Efficiency Drives Cost Drop
- Algorithmic and software efficiency improvements reduce AI compute costs by about 3x per year.
- Improvements make AI models cheaper to run over time despite increasing size and complexity.
AI Diffusion Trumps Model Lead
- AI's economic impact depends on diffusion and widespread deployment, not only on frontier model capabilities.
- The US's greater total compute capacity will accelerate AI adoption and economic benefits.