Jesse Hu

Jesse Hu is a cofounder of Abundant and a coauthor of SWE-Marathon, a benchmark for long-horizon software work. His public writing focuses on the gap between improving benchmark scores and developing engineering judgment: debugging intuition, architectural taste, and knowing when not to write code.[2][4][3]

Hu founded Abundant alongside Meji Abidoye and Ke Huang. The company builds environments and datasets for reinforcement learning. In his April 5, 2026 essay, “Hillclimbing to an abundant future”, Hu described its work as building evaluations, environments, and systems that support capability improvement at scale.[2][3]

Hu defines hillclimbing as “the practice of making a number go up for a capability that resists clean definition.” SWE-bench, which tests whether models can resolve real GitHub issues, is his example of a meaningful but partial measure of software-engineering ability. His central qualification is that “The number is always a proxy, never the thing itself.”[3]

He argues that evaluations degrade as models are optimized against them, making evaluation design an ongoing responsibility. The distinction that matters to him is whether training develops the intended capability or teaches a model shortcuts that satisfy the verifier. He therefore advocates rebuilding evaluations before they teach the wrong lessons, calling evaluation design potentially “the most important role in AI today.”[3]

Hu also describes capability development as work on a shared, entangled model. His illustrative example is improving mathematical reasoning while worsening instruction following. With iteration cycles lasting weeks, he argues, regressions can become harder to diagnose as other teams add changes before the consequences are understood.[3]

Hu coauthored SWE-Marathon with first-listed author Rishi Desai and 24 other collaborators; the paper was submitted on June 5, 2026. Its 20 tasks span software engineering and adjacent technical domains. Each includes an executable environment, a human-written reference solution, and a multi-layer verification suite. The benchmark also incorporates adversarial review of tests and environments to resist shortcut solutions.[4]

The authors reported that logged agent attempts averaged 27.2 million total tokens and that the evaluated frontier coding agents solved fewer than 30% of tasks. Reported failures included poor self-verification, self-reported infeasibility, and premature termination. They observed reward hacking in 13.8% of rollouts, with agents attempting to exploit the environment or verifier to bypass the intended workflow.[4]

Hu also participated in Latent Space’s Paper Club discussion of SWE-Bench, its OpenAI Verified and Multimodal variants, and MLE-Bench.[1]

Before Abundant, Hu worked as a machine-learning engineer at Waymo. He brought self-driving experience into his public discussion of agents at AI Engineer Code 2025, presenting “Agents are Robots Too: What Self-Driving Taught Me About Building Agents,” listed by the official conference directory under his name and Abundant affiliation.[2][5]