Aakanksha Chowdhery

Aakanksha Chowdhery is a Stanford adjunct professor and artificial-intelligence researcher who helped lead Google’s 540-billion-parameter PaLM model and now teaches researchers how to build self-improving AI agents. Her work spans the infrastructure needed to train frontier models and the reinforcement-learning techniques that help those models solve, verify, and learn from difficult software-engineering problems.

Chowdhery studied electrical engineering at IIT Delhi, graduating in 2007, and earned her master’s degree and doctorate at Stanford in 2009 and 2013. Her early research addressed communications networks, signal processing, distributed systems, and privacy; in 2012, she received the Paul Baran Marconi Young Scholar Award. She subsequently conducted research at Microsoft Research and Princeton before joining Google.

At Google, Chowdhery became a technical lead on PaLM, whose research paper lists her first among its authors and credits her with shared project leadership. Her contributions encompassed scaling validation, training efficiency, distributed infrastructure, and code evaluation. The 540-billion-parameter model demonstrated capabilities across mathematical reasoning, multilingual tasks, and code generation. She also contributed to Gemini, Pathways, PaLM-E, MedPaLM, and instruction-tuning research.

In 2025, Chowdhery worked on autonomous coding at Reflection AI, applying reinforcement learning to software-engineering tasks. Her research perspective on autonomous coding centers on several related challenges:

• Verification makes coding especially suitable for reinforcement learning. Unit tests, execution results, and compilers provide concrete signals for evaluating generated programs and constructing rewards, although poorly designed rewards still invite manipulation.
• Inference-time scaling cannot substitute for reliable training. Sampling multiple answers, majority voting, and iterative revision can improve results, but searching extensively for rare correct solutions becomes prohibitively slow and expensive. Training should increase the likelihood of generating useful solutions directly.
• Reinforcement learning is also an infrastructure problem. Large-scale training requires coordinating generation and optimization loops, placing multiple model copies across accelerator clusters, and maintaining efficient hardware utilization.
• Autonomous software engineering exceeds code generation. Effective agents must reason across extended workflows, use tools, verify intermediate results, and learn from incomplete real-world feedback.

At Stanford, Chowdhery co-teaches CS329A: Self-Improving AI Agents with Azalia Mirhoseini, covering test-time computation, verification, reinforcement learning, planning, memory, and agent evaluation. She also teaches in Stanford’s 2026 agentic AI program and served as a program chair of MLSys 2026.