Daiyi Peng

I retired from Google in 2026 to start an experiment for post-AGI living.

Former research engineer at Google DeepMind, working on long-term LLM research, especially agents. Before that, at Google Brain, working on AutoML. Before that, at Microsoft, working on distributed systems and search engines.

GitHub Google Scholar

Open source

PyGlove — manipulating Python programs for ML, AutoML and beyond Langfun — object-oriented programming with LLMs Napa.js — a multi-threaded JavaScript runtime

Selected publications

Type-Compliant Adaptation Cascades: Adapting Programmatic LM Workflows to DataICLR 2026 Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic CapabilitiesTechnical report, 2025 MoLA: MoE LoRA with Layer-wise Expert AllocationFindings of NAACL 2025 Long-form factuality in large language modelsNeurIPS 2024 Gemini: A Family of Highly Capable Multimodal ModelsTechnical report, 2023

All publications →

Earlier writing

Napa.js: A multi-threaded JavaScript runtime Winery.js: A framework for service experimentation Continuous modification: a process to build services that constantly evolve