barry@here: /home/barry — zsh

barry@here:~$ about

Barry (Xuanyi) Dong

Barry (Xuanyi) Dong

AI Research Scientist since 2015 · ACM-ICPC Gold Medalist · 2nd @ ImageNet 2015

Senior Staff AI Researcher @ Meta FAIR Coding ← OpenAI ← Augment Code ← DeepMind

I have been working on AI since 2015. Among all the AIs I have trained, model size has grown 10^6 times, from 1MB to 2TB; dataset size has grown 10^6 times, from 0.2GB to 200TB; and the number of GPUs I use has grown 10^5 times, from 1 GPU to 100K GPUs.

AIs I built after ChatGPT: Gemini, o3, o4-mini, Augment RLDB

AIs I built before ChatGPT: Lion Optimizer, DoReMi Pretraining Recipe, Automated Search for Architectures/Hyperparameters

type help to explore.

Selected work

  • 2025 · Meta FAIR CodingPush the boundary of single prompt optimization, agentic coding, and kernel generation.
  • 2024 · OpenAIContributed to infra, ChatGPT o3 / o4-mini, and the pretraining data wall.
  • 2023 · Augment CodeLed RL for coding models from real developer IDE behavior.
  • 2020 · Google Brain/DeepMindResearch on LLMs[1,2] and AutoML[3,4,5] at Brain; later a founding member of Bard (now Gemini); contributed to Gemini 1.0 and 2.5.

Recent writing

January 31, 2026

Being an AGI Believer, Not Skeptic - Lessons from Augment Code

What it takes for AI research to compound—and what happens when it doesn't. Lessons from Augment Code on the gap between being right and winning.

November 26, 2024

Reinforcement Learning from Developer Behavior and Real IDE Environment

How we achieved a breakthrough in AI code completion by learning directly from natural developer IDE interactions, matching the gains of doubling model size and 10x more finetuning data.