The Insane Strategy Behind China's Secret AI Model

Summary

The episode investigates the stealth launch of 0xAlpha/OxAlpha, a viral free model on OpenRouter with huge free token allowances, strong coding benchmarks, and rumored continual learning. The hosts trace the model to Chinese AI lab Z.ai/Jipu and discuss China's broader state-backed strategy of cheap open-weight models, data harvesting, and price compression against Western frontier labs. They also flag the financial unsustainability of the lab and tease upcoming model launches from both China and the US.

• Stealth model 0xAlpha/OxAlpha launches on OpenRouter with 100 trillion free tokens and 1M context window.
• Benchmarks initially appear strong but real-world coding pass rate is lower around 63%.
• The model's tokenizer and responses align with GLM/Z.ai/Jipu, pointing to Chinese origin.
• Hosts debate whether daily improvement is continual learning or just checkpoint updates.
• China's state-backed AI strategy uses cheap open-weight models to compress Western model pricing.
• Jipu/Z.ai financials are seen as unsustainable without state subsidy.
• More AI model launches are expected from both Chinese and Western labs.