Lilian Weng

Overview

Lilian Weng is an artificial intelligence researcher whose main public output is her personal blog, Lil'Log, where she has documented her learning notes since 2017^[1]. Chinese-language coverage describes her as a Peking University alumna^[2] and has reported on her time at OpenAI and, later, at Thinking Machines Lab^[2][3].

Education and the start of the public record

Public Chinese-language coverage identifies her as a graduate of Peking University^[2]. The earliest dated public trace of her work described here is the blog itself: Lil'Log has carried her learning notes since 2017^[1]. The format is plain — she works through papers and technical material and writes up her understanding, with a date attached to each entry. Keeping notes in public has a side effect: the blog becomes a timeline, showing which questions she was working through at a given moment.

What the blog lists

The posts span several years and several corners of large-model research. Among the entries from 2023 are "Prompt Engineering" (15 March 2023), "LLM Powered Autonomous Agents" (23 June 2023) and "Adversarial Attacks on LLMs" (25 October 2023)^[1]. Later entries listed on the blog are "Reward Hacking in Reinforcement Learning" (28 November 2024), "Why We Think" (1 May 2025) and "Scaling Laws, Carefully" (24 June 2026)^[1].

The problems behind those titles

Read together, the titles map onto distinct problems. Prompt engineering is the practice of shaping the text handed to a language model — instructions, examples, formatting — so the model is more likely to produce what you want. An "LLM powered autonomous agent" is a system in which a language model plans steps, decides which tools to call and revises its own intermediate output, instead of answering in a single pass. Adversarial attacks on language models are inputs built deliberately to make a model fail: to leak a prompt, slip past a restriction or return a wrong answer. Reward hacking, in reinforcement learning, is what happens when a policy finds a way to score well on the reward signal without doing what the reward was meant to encourage — the number goes up, the intended behaviour does not appear. Scaling laws describe how model performance changes as parameters, data and compute grow; a title such as "Scaling Laws, Carefully" signals a more cautious look at that relationship. "Why We Think" belongs to the question of why models that write out intermediate steps before answering can do better than models that answer immediately.

The dates matter as much as the topics. The list runs from prompting and agents, through adversarial attacks and reward hacking, to reasoning and scaling^[1] — that is, from how models are used, to how they are attacked, to how they are trained and evaluated.

Career

In November 2024 she announced that she was leaving OpenAI after nearly seven years there^[2]. Coverage from July 2026 reports that she left Thinking Machines Lab citing health reasons, with 29 July as her final day, and that she was returning to OpenAI to work on research into AI systems that help build better AI systems^[3]. That phrasing points to a specific direction of work: AI systems used to improve the systems themselves.

Where she sits in the record

Her documented contribution is writing rather than a named model or product. The value of a blog like Lil'Log is cumulative: each post fixes a topic, a date and an explanation, so a reader arriving years later can still see how a question was framed at the time. The dated entries also show continuity across job changes — new posts appear in 2024, 2025 and 2026^[1], the same stretch of time in which coverage reports her leaving OpenAI and later returning to it^[2][3].