Parallel Web Systems engineer explains how the company moved beyond academic web-research benchmarks to a fully automated evaluation flywheel that samples production questions, generates reference answers with an oracle agent, judges with tool-equipped agents, and masks components to diagnose failures, along with the open challenges of evaluating subjective and infrastructure-dependent quality.
Halluminate, a Y Combinator S25 company, builds RL environments and benchmarks that teach AI models to do non-coding knowledge work, starting with financial services. In this fireside with YC Partner Jon Xu, co-founders Jerry Wu and Wyatt Marshall trace their pivot from evals to training environments, explain why verification quality matters more than task volume, and outline a roadmap toward multi-agent simulated companies. They also address reward hacking, alignment as an industry-wide responsibility, and what they look for in new hires.
Benjamin Clavié argues that the industry has mistakenly generalized from the "easy case" of coding agents to the much harder domain of general knowledge work. He highlights that code offers durable, explicit cues and narrow tasks, while knowledge work involves implicit meaning, contextual understanding, and agents needing to reconstruct intent. Clavié traces human solutions through "tool loops" and "organization loops," demonstrating how optimized tools and hierarchical agentic structures (like searcher agents generating memos) are essential to overcome performance ceilings and significantly bridge the "oracle gap" between AI and human knowledge workers.
David Brumley discusses the challenges and solutions for teaching AI models to hack, drawing parallels with human learning. He introduces a 'ladder of tasks' approach for reinforcement learning and addresses the critical flaw of traditional benchmarks: measurement difficulties with multiple vulnerabilities and 'reward hacking.' His team's 'Audit Task' uses deterministic graders and precision/recall metrics for open-world assessment. He demonstrates this with an in-depth case study on attacking Chrome's V8 engine, showcasing how advanced models achieve real zero-day exploits, and warns against 'benchmaxxing security' without robust, honest grading.
Ali Khial exposes critical flaws in popular coding benchmarks, revealing how ambiguous instructions, weak verifiers, and model 'reward hacking' create a disconnect between reported performance and real-world utility. He argues that this leads to a "trust gap" where engineers disregard leaderboards. Khial then outlines five principles for building trustworthy, production-grade benchmarks, emphasizing human-authored instructions, holistic grading, economic value, contamination-free design, and informative leaderboards, urging software engineers to contribute to their improvement.
Sean Cai discusses the evolving landscape of AI data markets, highlighting the shift from raw annotation to high-quality, process-based data. He introduces "Verifier's Law" and its three axes to explain application layer maturity, critiques the shortcomings of current benchmarks, and predicts future AI trends by analyzing data market signals. Cai concludes by envisioning data companies transforming into "neo-labs" that provide enterprise-level reinforcement learning as a service and "Antikythera mechanisms" to manage and monetize real-world data assets.