What China Does Not Want the U.S. to Know
In a leaked call transcript, we learn what DeepSeek's CEO thinks are China's weaknesses and strengths in the AI race with the United States | Edition #309
TL;DR
- DeepSeek's CEO identifies resources, specifically compute power, as China's most significant gap with the U.S. in AI development.
- The CEO believes the talent gap is minimal, with the perceived difference stemming from fewer experimental opportunities due to limited compute.
- China's largest models are an order of magnitude smaller than U.S. counterparts, requiring vast numbers of domestic or foreign GPUs for training.
- Huawei's limited production and lower capital investment hinder China's ability to train large-scale models.
- The CEO suggests Chinese AI companies will likely focus on cost-effectiveness and product development, potentially becoming major global producers.
- There is optimism regarding domestic chip development, with Huawei's capabilities potentially rivaling NVIDIA's, albeit with a time lag.
- DeepSeek prioritizes reasonable returns over maximizing profit, aiming for cost efficiency in services.
- China's AI development is estimated to be 1-2 years behind the U.S. in some aspects but uses significantly less compute.
- The CEO foresees no upper limit to language model scaling and emphasizes the future inevitability of embodied intelligence.
- High-cost data labeling, a U.S. strength, is a challenge for China due to capital structure limitations, with a focus on labeling lower-cost data.
- The most reasonable approach for China seems to be focusing on general AI Agents, with Coding Agent as a top priority.