I’m a technical AI researcher, engineer, and former founder based in SF/NYC. What motivates my work is a desire to better appreciate the beauty of the world, either by understanding it more deeply or by helping reshape it so that it better embodies our sense of what is good and beautiful. My current research interests lie in building epistemically humble systems that know when they’re in situations that require deference to humans and developing scalable ways to align super-intelligent AI. These include directions like bootstrapped supervision (i.e. weak to strong supervision), scaling alignment compute to pre-training scale, and reward-model-free techniques for value alignment. If you’re interested in any of these topics, please don’t hesitate to reach out!

In my research endeavors, I was most recently a research fellow at both MATS and Haize Labs. A concise summary of my recent work history:

• Haize Labs | Research Fellow: ****Worked on techniques for improving model alignment to human preferences in non-verifiable domains, improving sample efficiency for training in such domains, and scaling weak supervision in LLMs (last one currently in progress).
• MATS | Research Fellow: ****Worked with Owain Evans on model evaluations and dangerous capabilities demonstrations. Developed black-box techniques to probe models for hidden beliefs and demonstrated that models can generalize novel backdoor triggers and behaviors through out-of-context reasoning. Currently working on causally understanding shared mechanisms that drive phenomena like chunky post-training, inoculation prompting, and backdoored models.
• OpenAI | Engineer: ****Worked on ChatGPT and related products. Part of the core team that developed 1-800-ChatGPT and OpenAI’s authentication platform.
• Sonnet | Founder: ****Founded a startup in 2022 for automating meeting note-taking. Handled all of the engineering and research required for such a product to work consistently at scale. Exited October 2024.

Writing and Publications

<aside>

🔗 Links

• 📬 [email protected]
• :github: Github (@Dylan102938)
• :linkedin: Linkedin (@dylan-feng)
• :google-scholar: Google Scholar
• :twitter: Twitter/X (@dylanfeng_) </aside>

<aside>

🚌 Education

University of California, Berkeley c/o 2024 Computer Science, Highest Distinction

Cal Dragon Boat Upsilon Pi Epsilon Honors Society ”Borrowed” a Calapalooza Banner (seeInstalling the Calapalooza Banner) one night with my roommate and kept it hung up in our apartment for 2 years

</aside>