OpenAI VP Lilian Weng Departs and Shares Full AI Safety Talk Transcript

Yesterday, Lilian Weng, a Peking University alumna who served as OpenAI's research vice‑president for safety, announced her resignation, prompting a detailed look at her recent public activities.

At the 2024 Bilibili Super Science Evening event, Weng delivered a talk titled "AI Safety and the Way of 'Cultivation'," aimed at a general audience. She emphasized that AI safety is a shared responsibility, not just for researchers, and illustrated the potential benefits of safe AI in everyday life, such as intelligent home systems and health assistants.

The speech covered several core topics: the necessity of grounding AI development in robust safety foundations, the risks of bias in data (e.g., gender‑biased medical datasets), and the role of diverse, high‑quality data in training trustworthy models.

Weng introduced reinforcement learning from human feedback (RLHF) as a key technique for aligning AI with human values, explaining reinforcement learning with a simple dog‑training analogy and describing how human‑generated rankings can guide large language models toward better, safer behavior.

She also highlighted practical tips for prompting AI effectively, such as providing detailed context and carefully chosen keywords, and drew parallels between AI alignment and cultural tools like the "tightening rope" on Sun Wukong in "Journey to the West".

The article then shifts to recent turbulence at OpenAI: the departure of AGI Readiness lead Miles Brundage, co‑founder and chief scientist Ilya Sutskever, and safety lead Jan Leike, all citing concerns that product focus has overtaken safety culture.

These exits raise questions about the future direction of OpenAI's safety efforts, while the piece concludes by urging the broader community—especially active participants on platforms like Bilibili—to engage in AI safety, supervision, and alignment to ensure AI remains beneficial and trustworthy.