Eric Jang is an AI researcher working on learning, search and autonomous research systems. He was vice president of AI at 1X Technologies and a robotics researcher at Google Brain, and coauthored the Gumbel-Softmax paper.[2][7]
Start with Building AlphaGo from scratch →
Education
Links
Personal writing and projects Owned education and career
X · ericjang11 
GitHub · ericjang Career
Jang’s Brown education combined a computer-science master’s with an undergraduate degree in applied mathematics and computer science, both listed for 2012–2016. From 2016 to 2022 he researched robotic learning at Google, including manipulation, shared autonomy and the SayCan project. He joined the company now called 1X in 2022, built its AI team and worked on learning systems for humanoid robots. His January 2026 departure announcement describes leaving 1X.[2][3]
Contributions
Use search to improve the next learning target
In his AlphaGo walkthrough, Jang explains how a policy proposes promising moves, a value model estimates outcomes and tree search examines selected continuations. The search can produce a stronger action target at each move, which the learning system can then imitate. He uses this worked example to examine self-play and credit assignment in language-model reinforcement learning, where a successful final answer does not immediately reveal which earlier tokens helped.[5]
Make research iteration observable in a game environment
AutoGo combines neural-network training, self-play, evaluation and distributed execution in a Go research codebase. Jang presents it as an environment for studying an automated researcher, with cheap generated experience and faster feedback than physical robotics. The repository documents a GPU-enabled development container and worker infrastructure. It also reports weaknesses in the current model’s understanding of life and death, so the project is an experimental research system rather than a claim of solved competitive play.[6]
Differentiate through a relaxed categorical choice
The Gumbel-Softmax paper replaces a nondifferentiable categorical sample with a continuous approximation that can carry gradients. Changing its temperature moves the sample between a softer mixture and a more discrete choice. This gives neural systems a way to learn components involving categorical decisions while retaining gradient-based optimization; the approximation and temperature remain part of the modeling choice.[7]
Tools, writing & talks
Writing, research & talks
Start here
Building AlphaGo from scratchdwarkesh.comA technical conversation connecting policy/value models, tree search, self-play and reinforcement learning.
Trace one move through the policy, value estimate and search before comparing the language-model discussion. Basic Go rules help, and the conversation introduces them; running the separate AutoGo code requires Python, GPU tooling and container familiarity.
Guest and technical demonstrator[5]Eric Jang explains rebuilding AlphaGo on sabbatical, how a deep network amortizes game-tree search, and what self-play and RL results suggest for LLM research.
Catalog workOriginal sourceA differentiable relaxation for learning with categorical variables.
Coauthor with Shixiang Gu and Ben Poole[7]