624: Do Less Math in Computers

What It Covers

Chinese AI startup DeepSeek releases r one reasoning model matching OpenAI's o one performance at 3% training cost using restricted hardware, causing 17% NVIDIA stock drop and challenging assumptions about AI development costs, moats, and American technological supremacy in artificial intelligence.

Key Questions Answered

• •Cost optimization breakthrough: DeepSeek trained r one for approximately $6 million versus OpenAI's estimated $200 million by implementing mixture of experts architecture splitting models into specialized components and multi head latent attention compressing memory usage during inference, demonstrating dramatic efficiency gains through algorithmic innovation rather than hardware superiority.
• •Export restriction circumvention: Using NVIDIA h 800 chips instead of cutting edge GPUs due to US export controls, DeepSeek extracted maximum performance through low level optimization similar to console game development, proving technological constraints can drive superior engineering solutions and that hardware restrictions fail to maintain competitive advantages in AI development.
• •Open source strategy advantage: DeepSeek releases model weights under MIT license and publishes complete research papers, contrasting with OpenAI's closed approach. CEO states open source attracts technical talent through respect and accomplishment rather than creating protective moats, suggesting transparency builds stronger organizational culture and sustainable competitive advantages in AI.
• •Reinforcement learning without humans: R one zero eliminates expensive human feedback loops used in ChatGPT training, relying purely on reinforcement learning with reward functions for correct answers and proper formatting. This approach scales better than human generated question answer pairs, reduces costs significantly, and potentially avoids human bias limitations in model training.
• •Apple's AI vulnerability: Apple lacks core competency in LLM development despite hardware advantages like unified memory architecture and neural engines. The company started AI efforts years behind competitors, ships underwhelming Apple Intelligence features slowly, and risks disruption similar to Microsoft missing mobile computing, though platform control provides temporary protection.