Lilian Weng
Last reviewed
May 31, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 · 1,788 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 31, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 · 1,788 words
Add missing citations, update stale details, or suggest a clearer explanation.
Lilian Weng is an artificial intelligence researcher known for her work at OpenAI, where she spent about seven years and led the Safety Systems team as Vice President of Research and Safety, and for her technical blog Lil'Log, one of the most widely read educational resources in modern machine learning [1][2]. She joined OpenAI in 2018, contributed to its robotics and applied research efforts, and built up the organization's safety engineering work before departing in November 2024 [1]. In 2025 she became a co-founder of Thinking Machines Lab, the artificial intelligence startup founded by former OpenAI chief technology officer Mira Murati [3][4].
Weng is recognized both for her research leadership and for her writing. On Lil'Log she has published long expository posts on subjects that include reinforcement learning, diffusion models, large language model agents, hallucination, and reward hacking, and these posts are commonly cited as study material by students, practitioners, and other researchers [2][5][6].
Weng completed an undergraduate degree at Peking University in China [4][7]. She then moved to the United States for doctoral study and earned a PhD from Indiana University Bloomington in 2014 [7][8]. Her doctoral research was in the area of network science and complex systems, and she worked within Indiana University's Center for Complex Networks and Systems Research under the supervision of Filippo Menczer [7][8]. Her early academic publications studied the spread of information and attention across social networks rather than the deep learning topics she later became known for [8].
After finishing her PhD, Weng worked in industry data science and machine learning roles. According to her professional record, she was at Dropbox from 2014 to 2016 and then at the financial technology company Affirm from 2016 to 2018 [7]. Her work during this period centered on applied machine learning and data science rather than fundamental research.
Weng joined OpenAI in 2018 [1][7]. Her early work there was on the robotics team. She was among the contributors to the project in which a robotic hand learned to manipulate and solve a Rubik's Cube, described in the 2019 paper Solving Rubik's Cube with a Robot Hand [1][9]. The system, built on a dexterous hand named Dactyl, was trained largely in simulation using a method the team called automatic domain randomization, which generated training environments of increasing variety so that a policy learned in simulation could transfer to physical hardware [9]. Weng has written that the effort took a long time to make the hand robust enough to operate in the real world [1].
As OpenAI shifted its focus toward large language models, Weng's role moved toward applied research. Around 2021 she helped form and lead the Applied AI Research team [1][10]. That group worked on product-facing systems and the infrastructure around them, including the fine-tuning API, the embedding API, content moderation endpoints, and applied safety frameworks [10]. This work sat at the boundary between research and the services that external developers used to build on OpenAI's models.
Following the launch of GPT-4 in 2023, Weng took over the company's Safety Systems team [1][10]. The team brought together OpenAI's deployment-stage safety work, including adversarial robustness, abuse prevention, and the technical safeguards applied to systems such as ChatGPT [1][10]. In her departure note she described the Safety Systems group as having grown to more than 80 scientists, researchers, engineers, and policy experts [1]. In August 2024 her title became Vice President of Research and Safety [1].
During her time at OpenAI, Weng's responsibilities connected several strands of the company's work. The applied research she led produced tools that developers used directly, and the safety work she later managed dealt with the risks of deploying generative models at scale, an area closely tied to broader questions of AI safety and AI alignment [1][10].
Weng has written a personal technical blog called Lil'Log, hosted at lilianweng.github.io, since 2017 [2][6]. She has described the blog as a place to document her learning notes, and she began it as a way to organize her own understanding when she was relatively new to deep learning [2][5]. Over time it grew into a reference that many people in the field read and recommend.
The posts are long and detailed. They typically work through a topic from its foundations, collect the relevant research, and present the mathematics and intuition together with diagrams. Among the most widely read entries are an early overview titled An Overview of Deep Learning for Curious People (2017), a survey of reinforcement learning titled A (Long) Peek into Reinforcement Learning (2018), and a 2023 post titled LLM Powered Autonomous Agents that became a common reference point for people building agent systems on top of language models [5][6][11]. Later posts have covered subjects such as reward hacking and, in 2025, a long piece on test-time reasoning titled Why We Think [12][13].
The blog is frequently described as one of the most influential technical blogs in contemporary AI, and its posts are used as teaching material in study groups and reading lists [6]. Weng has framed the act of writing thorough explanations as a form of self-education, on the view that explaining a concept clearly is a strong test of whether one understands it [6]. The combination of careful synthesis and free public access made the blog a common entry point for readers trying to follow fast-moving research areas.
Weng announced her departure from OpenAI on 8 November 2024, stating that her last day would be 15 November [1]. In a message posted on the social platform X, she wrote that after seven years at OpenAI she felt ready to reset and explore something new, and that she had made what she called an extremely difficult decision to leave [1]. She said she was proud of what the Safety Systems team had built [1]. An OpenAI spokesperson said the company deeply appreciated her contributions to safety research and to building technical safeguards, and expressed confidence that the Safety Systems team would continue its work [1].
Her exit came during a period in which a number of OpenAI's safety and research leaders left the company, including Ilya Sutskever, Jan Leike, Mira Murati, and John Schulman [1]. Some departing staff publicly raised concerns about how the company balanced safety against the pace of product development, though Weng's own announcement did not frame her decision in those terms [1].
In December 2024, shortly after leaving OpenAI, Weng joined the venture firm Fellows Fund as a Distinguished Fellow, a role in which she said she would help support AI founders [14].
In February 2025 Weng became a co-founder of Thinking Machines Lab, an AI research and product company started by Mira Murati [3][4][15]. The founding group was drawn largely from former OpenAI staff and included John Schulman, who became chief scientist, along with Barret Zoph, Andrew Tulloch, and Luke Metz [3][4]. The company stated that it aimed to build multimodal AI systems that work collaboratively with people and that can be adapted and customized for different fields of work [4].
Thinking Machines Lab raised a large early funding round, reported at about 2 billion dollars at a valuation near 12 billion dollars, with investors that included Andreessen Horowitz and Nvidia [4][15]. In October 2025 the company released its first product, Tinker, an API for fine-tuning open-weight language models that lets developers run customization jobs without managing distributed training infrastructure directly [4][15].
Weng's interests span several areas of machine learning. Her doctoral work was in network science, and her later research and writing covered reinforcement learning, robotics, generative modeling, language model agents, and the safety of deployed AI systems [7][8][10]. Her published research at OpenAI includes the robotic manipulation work on Dactyl and contributions to applied and safety-oriented systems [9][10].
Much of her public influence comes through writing rather than individual papers. Her surveys on reinforcement learning, prompt engineering, diffusion models, autonomous agents, and adversarial attacks on language models are widely shared, and several have been treated as reference summaries of their respective subfields [2][5][6][11]. This educational writing, which sits alongside related efforts in the field to make work on alignment and safety more legible, is a large part of how she is known outside OpenAI [6].
Weng is regularly described in technology coverage as one of OpenAI's lead safety researchers and as a notable figure in the field's research leadership [1][10]. Lil'Log has been characterized as among the most influential technical blogs in AI, and it gave her a public profile that extends well beyond her institutional roles [6]. Her appointment as a Distinguished Fellow at Fellows Fund and her position as a co-founder of Thinking Machines Lab reflect her standing among researchers who moved from OpenAI into new ventures during 2024 and 2025 [3][14].
| Field | Detail |
|---|---|
| Name | Lilian Weng |
| Occupation | AI researcher |
| Nationality | Chinese |
| Undergraduate | Peking University |
| PhD | Indiana University Bloomington, 2014 |
| PhD field | Network science and complex systems |
| Doctoral advisor | Filippo Menczer |
| Earlier roles | Dropbox (2014 to 2016), Affirm (2016 to 2018) |
| OpenAI | 2018 to 2024; VP of Research and Safety; head of Safety Systems |
| Thinking Machines Lab | Co-founder, from 2025 |
| Known for | Lil'Log blog; OpenAI Safety Systems leadership |
| Blog | lilianweng.github.io (Lil'Log), since 2017 |