Lilian Weng (Lil'Log) · RSS 摘要

信源：Lilian Weng (Lil'Log) · https://lilianweng.github.io/index.xml

• 面向自我改进的 Harness 工程 — The concept of recursive self-improvement (RSI) dates back to I. J. Good (1965), where he defined an “ultraintelligent machine” as a system that can surpass humans in …
• Scaling Laws, Carefully — Scaling laws are one of the most critical empirical findings in deep learning. The observation is simple in form: the training loss $L$ decreases predictably as we scale up model …
• Why We Think — 特别感谢 John Schulman 对本文提供的诸多宝贵反馈与直接修改。Test time compute（Graves et al. 2016, Ling, et al. 2017, Cobbe et al. 2021）……
• 强化学习中的奖励黑客 — 奖励黑客是指强化学习（RL）Agent 利用奖励函数中的缺陷或歧义以获取高奖励，而并未真正学习或完成任务。
• LLM 中的外在幻觉 — 大语言模型中的幻觉通常指模型生成不忠实、虚构、不一致或无意义的内容。作为术语，幻觉在某种程度上……