Who's Afraid of Chinese Models?
17 hours ago
- The author initially believed tech, especially software's zero marginal costs, was fundamentally different, but later realized zero marginal costs underpin Aggregation Theory, leading to centralized value chains.
- AI is reintroducing old business principles like marginal costs, challenging the notion that open-weight models are free; they involve significant COGS (cost of goods sold) for inference.
- Tokens are not fungible commodities; intelligence derived from tokens is fungible. COGS for intelligence depends on factors like model footprint, inference efficiency, and token efficiency.
- Commodity markets operate where all sellers charge the same price based on supply and demand, with profits depending on cost structures; suppliers with higher marginal costs risk bankruptcy.
- Currently, the AI intelligence market isn't fully commoditized due to high demand and compute shortages, giving frontier labs like Anthropic and OpenAI high margins and low costs per unit of intelligence.
- Frontier labs may fear Chinese models due to training cost dominance, the non-perfect commoditization of intelligence (where data from inference improves models), and competition in customer experience integration.
- China's strategy involves commoditizing AI complements through open-weight models to boost its physical world dominance and weaken U.S. frontier labs, leveraging distillation attacks for cost advantages.
- Distillation attacks allow Chinese labs to use frontier models as teachers for rapid improvement, disadvantaging Western open-weight model makers who must follow stricter terms of service.
- Cybersecurity concerns arise as defenders may need to use Chinese models due to restrictions on U.S. models, highlighting the need for policy changes to ensure equal competition and access.
- The author argues against overreacting to Chinese models from an economic perspective but emphasizes cybersecurity risks and calls for U.S. policies that support innovation and defense capabilities.