Summary
Nathan Lambert, a prominent voice in open-model research, joins a16z for a wide-ranging conversation on whether open-weight models can keep pace with the closed frontier labs at OpenAI, Anthropic, and Google. The discussion coincides with the launch of his Trillium project.
Lambert frames model progress as a function of three kinds of capital: compute, data, and talent. He explains why organizational structure matters so much to frontier labs, how OpenAI’s reasoning breakthrough and Anthropic’s coding focus shaped their positions, and why Chinese labs such as Moonshot AI and Zhipu pursue open releases as a practical route to relevance. The conversation weighs China’s talent strength against its compute constraints.
The interviewers and Lambert also dig into how post-training has changed, from supervised fine-tuning toward reinforcement-learning-heavy environments, and why buying high-quality data is central to progress in domains that are hard to verify, such as law, healthcare, and biology. Distillation, the debate over fears about a hidden frontier, and cybersecurity preparedness come up along the way. A Cursor experiment on rewriting SQLite is discussed as evidence that fine-tuned open models can be good enough for demanding tasks. This is a useful listen for anyone tracking the open versus closed model race.
📺 Source: a16z · Published October 14, 2026
🏷️ Format: Interview
