来自 LILA Science 的 Rafa Gomez-Bombarelli 和 Andy Beam 讨论未来实验室应如何像数据中心一样，以科学数据作为新前沿来扩展 AI 规模。
Rafa Gomez-Bombarelli and Andy Beam from LILA Science discuss how the lab of the future should feel like a data center, scaling AI with scientific data as the new frontier.
要点 · TL;DR
用实验产生的科学数据扩展 AI 是互联网数据枯竭后的下一个前沿。 Scaling AI with scientific data from experiments is the next frontier after internet data exhaustion.
通用科学推理模型因跨领域迁移而优于专用模型。 A general scientific reasoning model outperforms domain-specific ones due to cross-domain transfer.
未来的实验室应像数据中心：密集、节能、无人值守。 The future lab should be like a data center: dense, energy-efficient, and lights-out.
核心观点 · Key points
苦涩教训和Scaling（规模扩张）是关键：可扩展的通用方法胜过专用方法。 The bitter lesson and scaling are key: general methods that scale beat specialized ones.
科学是通过实验验证的强化学习进行后训练的无限token生成器。 Science is an infinite token generator for post-training via RL with experimental verification.
未来的实验室应像数据中心：密集、节能且无人值守。 The lab of the future should resemble a data center: dense, energy-efficient, and lights-out.
通用科学推理模型因跨领域迁移而胜过专用模型。 A general scientific reasoning model beats domain-specific models due to cross-domain transfer.
迭代速度比大规模并行化对科学发现更重要。 Iteration speed is more important than massive parallelization for scientific discovery.
反共识 · Contrarian takes
互联网数据已耗尽；下一个Scaling（规模扩张）轴是实验生成的数据。 The internet data is exhausted; the next scaling axis is experimentally generated data.
Lyra不是生物技术公司；模型本身是主要资产，而非任何疗法。 Lyra is not a biotech company; the model itself is the primary asset, not any therapeutic.
由于能力曲线呈S形，AI驱动实验室的安全性现在就必须重视，而非以后。 Safety in AI-driven labs is critical now, not later, due to sigmoid capability curves.
强化学习中的奖励破解可能有益；有时跳过推理步骤反而得到更好结果。 Reward hacking in RL can be beneficial; sometimes skipping reasoning steps yields better results.
材料科学比生物学更难，因为缺乏统一原理和自动化。 Materials science is harder than biology due to lack of unifying principles and automation.
本期章节 · Chapters（共 37）
引言与背景Introduction and Backgrounds
LILA论点：苦涩教训与数据扩展LILA's Thesis: The Bitter Lesson and Data Scaling
通过RL进行科学后训练与可验证奖励Post-training via RL with verifiable rewards in science
API与人机协作API and human-in-the-loop
AI社区认知转变与团队建设Shift in AI community awareness and team building
数据暴露与验证的基础设施Infrastructure for data exposure and verification
验证时间与向超级智能的转变Time spent on verification and shift to superintelligence