← Prometheans 100+ Lilian Weng

OpenAI safety lead, legendary technical blogger

Lilian Weng

VP of Research — OpenAI

Profile

Lilian Weng is one of those rare researchers whose blog is more influential than most academic papers. For nearly a decade, her site Lil’Log has been the place engineers and researchers go when they actually want to understand something — diffusion models, attention, RLHF, agents, reward hacking. Where papers obscure with notation and jargon, Weng writes clean, exhaustive explainers that walk you from first principles to the current frontier. Ask any working ML engineer where they learned how transformers or diffusion really work, and Lil’Log comes up a lot.

She spent nearly seven years at OpenAI, where her career mirrored the company’s own pivot. She joined in 2017 on the robotics team, leading the legendary project that taught a single robotic hand to solve a Rubik’s Cube via domain-randomized reinforcement learning. When OpenAI shifted to language models, she shifted with it — founding and running the Applied AI Research team, shipping the fine-tuning API, embeddings API, and moderation endpoints, and eventually becoming VP of Research, Safety in August 2024. Three months later, she left.

In February 2025 she co-founded Thinking Machines Lab alongside Mira Murati, John Schulman, Barrett Zoph, Andrew Tulloch, and Luke Metz — an OpenAI alumni reunion that raised a $2B seed round. She’s also a Distinguished Fellow at Fellows Fund. The mission is building more understandable and customizable AI; given the collective pedigree, it’s one of the most-watched labs on the planet.

For developers learning AI, Weng matters for a simple reason: she’s the best technical writer in the field. Read her posts before the papers they cite. You’ll save hours and understand more.

Key Articles & Papers

What are Diffusion Models? — The canonical explainer on diffusion models — if you've ever tried to understand DDPM from the paper, read this first. LLM Powered Autonomous Agents — Widely cited as the definitive overview of the agent paradigm — planning, memory, tool use, and reflection all in one place. Prompt Engineering — A rigorous taxonomy of prompting techniques, from zero-shot through chain-of-thought and beyond. Attention? Attention! — The post that introduced a generation of engineers to attention mechanisms — still holds up. The Transformer Family Version 2.0 — A 45-minute tour of transformer variants, efficiency tricks, and long-context architectures. Reward Hacking in Reinforcement Learning — A deep dive on how RL agents exploit reward functions — essential reading for anyone doing RLHF. Extrinsic Hallucinations in LLMs — Survey of why LLMs hallucinate and what actually works to reduce it. Why We Think — Her take on test-time compute and the wave of reasoning models — why thinking longer works. Policy Gradient Algorithms — The reference post for REINFORCE, A2C, PPO, and the RL algorithm family tree. Thinking about High-Quality Human Data — Why data annotation quality is the bottleneck most teams underestimate.

Spotify Podcasts

Episode 11: The Future of AI: Lessons from Fei-Fei Li and Lilian Weng
Episode 11: The Future of AI: Lessons from Fei-Fei Li and Lilian Weng
Lilian Thuram à propos d'Arsène Wenger : "Sa rencontre a changé ma vie"
Lilian Thuram à propos d'Arsène Wenger : "Sa rencontre a changé ma vie"
[download PDF] Gone to Dust (Gravediggers #2) by Liliana Hart
[download PDF] Gone to Dust (Gravediggers #2) by Liliana Hart

Related People

pioneer Sam Altman