Percy Liang 讨论了将 AI 模型扎根于共享现实以避免极化的重要性，并分享了他从早期 NLP 研究到基础模型的历程。
Percy Liang discusses the importance of grounding AI models in shared reality to avoid polarization, and shares his journey from early NLP research to foundation models.
要点 · TL;DR
前沿模型的评估需要中立第三方以确保透明度。 Evaluation of frontier models needs neutral third parties for transparency.
学术界应专注于新颖想法，而非仅仅扩大模型规模。 Academia should focus on novel ideas, not just scaling models.
语言模型必须与共享现实挂钩，以避免极化。 Language models must be tethered to shared reality to avoid polarization.
核心观点 · Key points
前沿模型的评估必须由中立的第三方进行，以确保透明度和合法性。 Evaluation of frontier models must be done by neutral third parties to ensure transparency and legitimacy.
学术界应专注于新颖想法和社会影响，而不仅仅是扩大模型规模。 Academia should focus on novel ideas and societal impact, not just scaling models.
语言模型必须与现实挂钩，以避免两极分化。 Language models must be tethered to shared reality to avoid polarization.
需要民主程序来确定 AI 系统中嵌入的价值观。 Democratic processes are needed to determine values embedded in AI systems.
数据归因可以为内容创作者提供公平补偿并提高数据质量。 Data attribution could enable fair compensation for content creators and improve data quality.
反共识 · Contrarian takes
开放模型不是开源；它们更像是开放的二进制文件，可审计性有限。 Open models are not open source; they are more like open binaries with limited auditability.
对 AI 风险的恐慌会集中权力并损害开放研究。 Fearmongering about AI risk centralizes power and harms open research.
用 LLM 模拟人类社会具有科学价值，但若作为证据则存在风险。 Simulating human societies with LLMs is scientifically valuable but risky if taken as evidence.
微调和预训练在生态系统中扮演不同角色；微调实现定制化。 Fine-tuning and pre-training serve different ecosystem roles; fine-tuning enables customization.
模型合并和版本控制可以借鉴软件开发实践。 Model merging and version control for models could mirror software development practices.
本期章节 · Chapters（共 21）
共同现实与极化Shared reality and polarization
播客与嘉宾介绍Introduction of the podcast and guest
早期研究兴趣与博士历程Early research interests and PhD journey
早期职业生涯与鲁棒性Early career and robustness
旧思想：句法与语义分析Old ideas: syntax and semantic parsing
焦点转移：从鲁棒性到基础模型Shift in focus: from robustness to foundation models
负责任 AI 与透明度Responsible AI and transparency
基础模型愿景Vision for foundation models
语言模型稳健评估Robust evaluation of language models
通用模型评估Evaluation with general-purpose models
评估与模型构建的资金失衡Funding imbalance in evaluation vs model building
对智能体模拟与生成式智能体的兴趣Interest in agentic simulations and generative agents
模拟人类系统的潜力与风险Potential and risks of simulating human systems
递归自我改进与安全Recursive self-improvement and safety
开放权重模型 vs 开源Open-weight models vs open source
预训练 vs 微调Pre-training vs fine-tuning
学术界在能力方面的角色Role of academia in capabilities
学术 AI 研究的计算需求Compute needs for academic AI research