Raw LLM Responses

Great episode but one thing needs correcting. The "80% of Silicon Valley startups prefer Chinese models" framing is misleading. Martin Casado from a16z - the original source - corrected this himself on X. The actual number: 20-30% of startups pitching a16z use open-source models. Of those, ~80% use Chinese ones. That's 16-24% of all startups. Not 80%. Big difference. And even within that slice the use case matters. Airbnb uses Qwen for one task inside a 13-model stack. Chamath moved some workloads to Kimi K2. These are cost optimization decisions on commodity inference tasks. Not "preferring Chinese AI." The cost differential is real. Nobody disputes $2 vs $15 per million output tokens. But framing it as "Silicon Valley prefers Chinese AI" conflates open-source model selection for non-sensitive workloads with some kind of strategic shift. It isn't. The moment you're in regulated industries, enterprise contracts, government, healthcare, finance - the picture flips entirely. Security, data residency, model governance, CCP-aligned content biases that travel with the weights. None of that disappears because inference is cheap. The real story here isn't "US vs China model preference." It's that open-weight models are eating the commodity layer while proprietary models hold the high-trust layer. That's an architecture story, not a geopolitical crisis.