Will Hang is a machine-learning engineer, researcher, and startup cofounder whose work includes AlphaChip, enterprise machine-learning infrastructure, and agent reinforcement fine-tuning. He builds systems that learn under practical constraints, from chip-design optimization to tool-using agents operating within strict inference budgets.
Hang studied computer science at Stanford, cofounded include.ai, an accessibility-focused artificial-intelligence startup later acquired by Retool, and conducted research at DeepMind and Google Brain. He coauthored work combining reinforcement learning with heuristic optimization and the 2021 Nature paper on learning-based chip placement, associated with AlphaChip.
At Snorkel AI, he led work on interactive machine-learning infrastructure. His redesign of Snorkel Flow’s underlying systems used Ray, distributed task queues, out-of-core processing, and in-memory actors to accommodate customers with dramatically different hardware and deployment constraints.
His subsequent work at OpenAI included GPT-4o fine-tuning, vision fine-tuning, and reinforcement fine-tuning for agents. In a joint technical session with Cathy Zhou, Hang outlined several principles for improving tool-using models:
- Grade complete agent trajectories. Associate each rollout with its tool calls, intermediate results, and final answer so graders can evaluate the full behavior.
- Train against production constraints. Apply inference-budget penalties to excessive tool use while preserving accuracy, and align training and evaluation examples with real deployment conditions.
- Exhaust simpler improvements first. Establish representative evaluations and baselines, then refine prompts, task design, and tool interfaces before changing model weights.
- Design rewards that cannot be gamed. Define success precisely, verify that additional exploration can uncover better solutions, and use informative continuous rewards when they provide useful guidance.
