Can Open Models Keep Up With OpenAI and Anthropic? | Nathan Lambert

在这期播客节目中，内森与主持人进行了面对面的交流，探讨了人工智能发展的宏观动态、前沿实验室、开源模型、中国人工智能生态系统，以及内森的新实验室 Trillium。 内森将模型的发展细分为三种资本形式：算力、数据和人才。尽管像 OpenAI 和 Anthropic 这样封闭的前沿实验室已经获得了巨额资金并进行了雄心勃勃的押注，但中国实验室面临着算力短缺，却拥有相当的人才。内森认为，根据科学文献，蒸馏并不是当前后训练的主要因素，他认为中国开源权重的发布在很大程度上是一种实际的商业化和分发策略。他指出，蒸馏只是服务于竞争，并帮助实验室弥合几个月的差距。 在比较开源和闭源模型时，内森解释说它们在不同的指数时间尺度上运行。闭源实验室是集成式的，推理效率高，并能快速触达高价值市场。开源和开放权重的模型经济需要更长的时间来构建，因为它涉及多个碎片化的参与者——构建者、推理公司和后训练初创公司——需要时间来匹配私营企业的数据并找到投资回报率。然而，开源模型对于初创公司生态系统至关重要，它为微调提供了可靠的帕累托前沿。 内森分享了对数据和环境的见解，强调前沿实验室在购买高质量数据和强化学习环境方面进行了大量投资，尽管这经常引入诸如奖励作弊等问题。为了应对这一问题，内森的新非营利组织 Trillium——以一种在树冠生长之前开花的春季短生植物命名——旨在为终端操作等高杠杆任务构建开源环境和数据栈。通过公开这些资源，Trillium 试图使研究民主化，揭开递归自我改进的神秘面纱，并帮助社会更好地理解自动化和智能体系统。 在谈到递归自我改进以及对闭源实验室拉开差距的担忧时，内森认为，虽然推理扩展和效率将继续急剧提高，但算力的指数级增长在性能上带来的回报却在递减，这使得突然的“快速起飞”可能性降低。不过，他警告说，虽然能力曲线图并不能令他担忧，但社会在很大程度上并未准备好应对随后随着开源模型赶上闭源模型而产生的跨公共网络安全和其他弥散性风险。 最后，内森谈到了预训练和行业的生态。他认为预训练科学仍然是可及的，艰巨的工作正在向后训练和编排智能体转移。尽管闭源实验室具有保密性，但知识自然会在科技界传播，从而使开源研究保持竞争力。归根结底，内森主张开展开放的研究、教育和基础设施建设，以确保更广泛的科技生态系统能够参与人工智能的发展并从中受益。

In this podcast episode, Nathan joins the host in person to discuss the macro-level dynamics of AI progress, the frontier labs, open-source models, the Chinese AI ecosystem, and Nathan's new lab, Trillium. Nathan breaks down model progress into three forms of capital: compute, data, and talent. While closed frontier labs like OpenAI and Anthropic have secured massive amounts of capital and taken ambitious bets, Chinese labs face compute shortages but possess comparable talent. Nathan argues that distillation is not a major factor in current post-training based on scientific literature, viewing Chinese open-weight releases largely as a practical commercialization and distribution strategy. He notes that distillation simply serves competition and helps labs bridge a few months of lag. When comparing open and closed models, Nathan explains that they operate on different timescales of exponentials. Closed labs are integrated, efficient at inference, and reach high-value markets quickly. The open-source and open-weight model economy takes much longer to build because it involves multiple fragmented players—builders, inference companies, and post-training startups—taking time to match private enterprise data and find ROIs. However, open-source models are essential for the startup ecosystem, providing a reliable Pareto frontier for fine-tuning. Nathan shares insights on data and environments, highlighting that frontier labs invest heavily in buying high-quality data and RL environments, though this often introduces issues like reward hacking. To counter this, Nathan's new nonprofit, Trillium—named after a spring ephemeral flower that blooms before the canopy grows—aims to build open-source environments and data stacks for high-leverage tasks like terminal operations. By making these resources open, Trillium seeks to democratize research, demystify recursive self-improvement (RSI), and help society better understand automation and agentic systems. Addressing RSI and the fear of closed labs pulling away, Nathan believes that while inference scaling and efficiency will continue to improve drastically, exponential increases in compute yield diminishing returns on performance, making a sudden "fast takeoff" less likely. However, he warns that while capability plots do not worry him, society is largely unprepared for cross-public cybersecurity and other diffuse risks that follow as open models catch up to closed ones. Finally, Nathan touches on pre-training and the human aspect of the industry. He believes pre-training science remains accessible and that the hard work is shifting to post-training and orchestrating agents. Despite the secrecy of closed labs, knowledge naturally travels through the tech community, keeping open-source research competitive. Ultimately, Nathan advocates for open research, education, and infrastructure to ensure the broader tech ecosystem can participate and benefit from AI progress.