Lilian Weng

wiki
, , ,
1

Former VP of Research and Safety at OpenAI, co-founder of Thinking Machines Lab, and author of Lil’Log — the most widely read technical ML blog in the field, written since 2017 as personal learning notes that became a standard reference across the AI community.

Profile

Field Detail
Nationality Chinese-American
Current Role Co-founder, Thinking Machines Lab; Distinguished Fellow, Fellows Fund
Research Areas AI Safety, Reinforcement Learning, LLM Alignment, Reward Hacking, Autonomous Agents, Robotics
Education B.S., Computer Science (China); Ph.D., Computer Science (US) — see note
Blog Lil’Log — lilianweng.github.io
X / Twitter @lilianweng
GitHub @lilianweng
Google Scholar scholar.google.com — 48,000+ citations

Note on education: Sources conflict on specifics. One profile reports a bachelor’s from Beihang University and a PhD from Carnegie Mellon University; another reports a bachelor’s from Peking University and a PhD from Indiana University Bloomington. Neither has been confirmed by a primary source. Her doctoral research covered reinforcement learning, robotics, and network science.

Overview

Lilian Weng is a Chinese-American AI safety researcher and one of the most influential writers in the field, best known for two parallel contributions that reinforced each other over seven years at OpenAI: her technical leadership building the company’s safety infrastructure, and her blog Lil’Log, which she has maintained since 2017 as personal learning notes and which became the most-read technical ML blog among practitioners. She joined OpenAI in 2017/2018 on the robotics team, worked on the Rubik’s Cube dexterous manipulation project, transitioned to applied AI research, led the Safety Systems team (80+ scientists), and was appointed VP of Research and Safety in July 2024 — before departing in November 2024. Business Insider named her to its 2024 AI Power List. She is a co-founder of Thinking Machines Lab, Mira Murati’s AI startup, alongside John Schulman, Barret Zoph, Alec Radford, and others. She describes “persistence” and “humility” as her core traits, both rooted in early encounters with mathematics competitions where she confronted the reality of talent disparities and learned to approach hard problems as a “normal person who has to give it her all.”

Early Life & Education

Weng grew up in China, where she developed an early interest in mathematics and a consuming curiosity across many subjects — she has described devouring encyclopedias as a child, from astronomy to astrology. She participated in mathematics competitions from a young age, an experience she credits with shaping the “humility” that characterizes her work: facing steep talent curves, she internalized that sustained effort and clear thinking matter more than innate advantage.

She completed a bachelor’s degree in Computer Science in China (Beihang University per one profile; Peking University per another) and then a doctorate in Computer Science in the United States, with research spanning reinforcement learning, robotics, and network science. Her interest in deep learning, which she encountered after her doctoral research, ignited the curiosity that ultimately led her to OpenAI.

Career

Pre-OpenAI Industry Experience

Before joining OpenAI, Weng worked as a data scientist and software engineer at several Silicon Valley companies, including Meta (then Facebook), Dropbox, and Affirm. She has also been described as having held a research science role at Snapchat. This practical engineering experience — building ML systems in production environments before entering an AI research lab — shaped her distinctively applied orientation toward safety and alignment questions: she consistently framed safety not as a theoretical concern but as a product engineering challenge.

OpenAI (late 2017 / 2018–2024)

Robotics Team (2018–2021)
Weng’s first role at OpenAI was on the robotics team, working on some of the most demanding physical AI problems in the company’s early history. The flagship project was the Rubik’s Cube dexterous manipulation system — a multi-year effort to train a robotic hand with reinforcement learning to solve the puzzle, demonstrated publicly in 2019 using domain randomization across hundreds of simulated configurations to achieve transfer to the physical robot. She has described the emotional intensity of hardware research: “running physical experiments is a tremendously exciting, challenging, and emotional experience.” The robotics period grounded her subsequent safety work in a hands-on understanding of reinforcement learning failure modes, reward hacking, and the gap between simulation and physical deployment.

Applied AI Research Team (2021–2023)
In 2021, Weng transitioned to OpenAI’s Applied AI Research team, where she led research connecting the company’s emerging language model capabilities to downstream applications. This period aligned with the development of ChatGPT and GPT-4 and required combining research-grade understanding of model behavior with the applied engineering perspective to make systems work reliably and safely in user-facing contexts.

Safety Systems Team Lead (2023–2024)
In 2023, Weng was tasked with leading OpenAI’s Safety Systems team — at its peak comprising over 80 scientists, researchers, and policy experts. The team’s mandate was to safeguard against risks from frontier models, including adversarial attacks, jailbreaks, reward hacking in RLHF training, and evaluating catastrophic risk scenarios. She also served on the board’s safety and security committee, giving her formal institutional oversight responsibility for safety decisions at the highest level of OpenAI’s governance structure.

VP of Research and Safety (July–November 2024)
Following the departure of former safety lead Aleksander Madry, Weng was appointed Vice President of Research and Safety in July 2024, consolidating safety research across OpenAI under a single leadership. Her mandate included both managing existing safety systems and expanding preparedness evaluations for future models. This appointment came in the context of a broader wave of safety departures from OpenAI, including Jan Leike, John Schulman, and Mira Murati.

On November 15, 2024, Weng announced her departure on X: “After 7 years at OpenAI, I feel ready to reset and explore something new. I made the extremely difficult decision to leave OpenAI. Looking at what we have achieved, I’m so proud of everyone on the Safety Systems team and I have extremely high confidence that the team will continue thriving.”

Thinking Machines Lab — Co-Founder (2025–present)

Weng is a co-founder of Thinking Machines Lab, launched by former OpenAI CTO Mira Murati in early 2025. The founding team includes John Schulman, Alec Radford, Barret Zoph, and several other OpenAI alumni. Thinking Machines Lab positions itself as an AI research and product lab focused on human-AI collaboration. Weng’s background in safety systems, applied AI research, and robotics maps directly to the lab’s agenda.

Fellows Fund — Distinguished Fellow (December 2024)

In December 2024, Fellows Fund, an AI-focused venture capital firm, announced Weng as a Distinguished Fellow. She stated: “Fellows Fund’s mission and vision are very compelling. I’m honored to join this group of AI Fellows to support the next generation of AI founders.”

Lil’Log (lilianweng.github.io)

Lil’Log is Weng’s personal technical blog, running since 2017. She describes its origin plainly on the homepage: “I’m documenting my learning notes in this blog since 2017.” What began as a personal practice for consolidating her own understanding became, over the subsequent years, one of the most linked and cited resources in the ML research and engineering community — a canonical first reference for practitioners encountering new topics.

The blog’s authority rests on a combination of exceptional depth, pedagogical clarity, and research currency. Posts are typically 20–45 minutes long, systematically cover the relevant literature with formal notation, and are updated as the field evolves. They function as neither pure tutorials nor pure literature reviews, but as something closer to a graduate student’s ideal explanation of a topic — technically rigorous, self-contained, and driven by genuine curiosity about understanding rather than performance.

Selected major posts:

Post Date Significance
A (Long) Peek into Reinforcement Learning Feb 2018 Comprehensive RL overview; widely used as a course reference
Meta-Learning: Learning to Learn in Few Steps Nov 2018 Canonical introduction to meta-learning
Domain Randomization for Sim-to-Real Transfer May 2019 Detailed technical explanation of Weng’s robotics research context
What are Diffusion Models? Jul 2021 Became the standard introduction to diffusion models before Stable Diffusion; repeatedly updated through 2024
Contrastive Representation Learning May 2021 Comprehensive survey of self-supervised learning
How to Train Really Large Models on Many GPUs? Sep 2021 Co-authored an upgraded version with Greg Brockman on the OpenAI Blog
Large Transformer Model Inference Optimization Jan 2023 Systematized the inference efficiency literature
Prompt Engineering Mar 2023 One of the most-read introductions to the field; characterized it as “empirical science”
LLM Powered Autonomous Agents Jun 2023 Defined the planning/memory/tool-use agent framework that became dominant in subsequent agent research
Adversarial Attacks on LLMs Oct 2023 Comprehensive survey of jailbreaking and adversarial robustness
Thinking about High-Quality Human Data Feb 2024 Analysis of annotation quality, RLHF labeling, and the undervaluation of data work
Reward Hacking in Reinforcement Learning Nov 2024 Her final major OpenAI-era post; synthesized safety-critical failure modes
Why We Think May 2025 Analysis of test-time compute and chain-of-thought reasoning; thanked John Schulman for “super valuable feedback and direct edits”

The post “LLM Powered Autonomous Agents” (June 2023) is Weng’s most-cited blog work. It introduced a coherent framework — LLM as agent brain, augmented by Planning, Memory, and Tool Use — that became the reference conceptual structure for agent system design in both research papers and product development. The post’s influence on how the field describes and builds LLM-based agents is difficult to overstate.

Key Contributions

  • Lil’Log — Seven years of deeply researched, pedagogically exceptional ML blog posts that have shaped how a generation of researchers and engineers understand the field; “What are Diffusion Models?” and “LLM Powered Autonomous Agents” are among the most-read ML explanations ever written.

  • OpenAI Safety Systems — Built and led the team responsible for safeguarding OpenAI’s frontier models against adversarial attacks, reward hacking, and catastrophic risks; the external red-teaming programs and preparedness evaluations she built institutionalized structured safety evaluation at a frontier AI lab.

  • Rubik’s Cube Dexterous Manipulation — Key contributor to the OpenAI robotics project demonstrating dexterous in-hand manipulation through reinforcement learning with domain randomization; a landmark result in physical AI.

  • Applied AI Research at OpenAI — Led the transition of OpenAI’s research capabilities into applied systems, contributing to the development of the evaluation and alignment infrastructure underlying ChatGPT and GPT-4.

  • Safety Policy and Governance — Served on OpenAI’s board safety and security committee; her consolidation of safety research under a single VP role represented a structural commitment by OpenAI to integrated safety oversight.

Awards & Recognition

  • Business Insider 2024 AI Power List — Named for leadership in safe and ethical AI development.
  • Fellows Fund Distinguished Fellow (December 2024).
  • 48,000+ Google Scholar citations — reflecting the reach of both her research publications and the indirect scholarly influence of Lil’Log.

Key Relationships

  • Mira Murati — Co-founder of Thinking Machines Lab; their joint founding reflects a shared conviction that safety and research infrastructure are inseparable from capable AI development.
  • John Schulman — Co-founder of Thinking Machines Lab; thanked by Weng for direct edits on “Why We Think” (2025); a long-term intellectual collaborator across RL and reasoning.
  • Aleksander Madry — Predecessor as head of OpenAI’s safety preparedness work; Weng took over the preparedness team when Madry was reassigned in July 2024 before being elevated to VP.
  • Jan Leike — Former OpenAI safety lead whose high-profile departure preceded Weng’s appointment; both were part of the wave of safety-focused departures that characterized OpenAI’s 2024.
  • Greg Brockman — Co-authored an upgraded version of Weng’s “How to Train Really Large Models” post for the OpenAI Blog; one signal of her integration into OpenAI’s research communication culture.

Personal Style

Weng’s public presence is defined by a paradox: she is among the most-read technical writers in AI, yet her blog carries the persistent framing of personal notes rather than authoritative exposition. This humility is genuine and biographical. She has traced it to childhood mathematics competitions where she encountered the limits of her own talent early, learning to approach hard problems incrementally rather than heroically. She describes her core traits as “persistence” and “humility”; friends describe her as “organized” and “patient.” The blog’s opening disclaimer — noting grammar mistakes as evidence that ChatGPT is not involved — signals that the writing is deliberate and first-person in a field where AI-assisted polish has become the norm. Her safety philosophy parallels her personal epistemology: she has stated that “scaling is not the only recipe” and that alignment and safety are “the most urgent challenges right now,” a position held not from outside critique but from years building these systems from the inside. Her Digg vibe profile (dominant topics: reinforcement learning, AI safety, LLM alignment) reflects a communicator whose public identity has remained technically grounded throughout her institutional rise.

References