Jazmia Henry details the end-to-end process of building specialized foundation models for the energy industry. She covers the four key stages from data curation of unstructured, handwritten documents to optimizing inference, and introduces her Grounded Continuous Evaluation (GCE) framework to combat reward hacking in reinforcement learning.
Nicholas Kang and Michael Aaron from Google DeepMind's Kaggle team discuss the broken state of AI evaluations—scattered, non-transparent, and created by a homogenous group. They present their solutions: a community-driven benchmarks platform, a PvP Game Arena for non-saturating ELO ratings, standardized agent exams, and hackathons to crowdsource novel evals and address the limitations of current benchmarking practices.
Vincent Koc argues that static benchmarks are failing in the era of adaptive AI. He proposes a shift from static testing to 'malleable evals,' where agents self-optimize and curate their own test suites based on user intent and production data, treating evaluation as a living, evolving system.
Maggie Konstanty, AI Product Manager at Prosus, provides a candid look into the realities of LLM evaluation in production. She argues that standard metrics like accuracy are misleading and advocates for a culture of continuous, goal-oriented evaluation focused on deep failure analysis and understanding real user behavior, asserting that mature teams inevitably build custom tooling to meet their specific needs.
Despite benchmarks showing relentless progress, many users remain dissatisfied with LLM responses in real-world scenarios. This summary explores two key analyses—a custom 'nonsense question' benchmark and trends from Chatbot Arena's 'dislike both' data—to reveal the persistent gaps in model reasoning, reliability, and domain-specific understanding.
This hands-on workshop details the construction of a sophisticated, dual-part AI system for producing high-quality technical content. It begins with an MCP-powered deep research agent that autonomously plans, searches the web, and analyzes sources like YouTube to synthesize a grounded research artifact. The second part is a constrained, deterministic writing workflow that transforms this research into polished, non-sloppy content using an innovative "Evaluator-Optimizer" pattern for iterative refinement. The session emphasizes crucial AI engineering principles, such as choosing between agentic and workflow-based architectures, and concludes with a deep dive into implementing practical observability and evaluation pipelines to ensure the system is both measurable and improvable.