Ted Moskovitz of Anthropic at RAAIS 2026

Key Takeaways

• Ted Moskovitz will speak at RAAIS 2026 in London.
• He leads Anthropic’s Science of Scaling team focusing on RL and optimization.
• His ICLR 2024 paper shows reward model overoptimization harms human-aligned performance.
• Research highlights need for constrained RLHF to keep models reliable at scale.
• Background spans neuroscience PhD, DeepMind and Uber AI internships.

Pulse Analysis

The Research and Applied AI Summit (RAAIS) has become a pivotal gathering for AI entrepreneurs and scholars seeking to translate breakthrough research into market‑ready solutions. The 2026 edition, held in London on June 12, marks the summit’s tenth anniversary and features a lineup of leading technologists, with Anthropic’s Ted Moskovitz headlining the program. Anthropic, a fast‑growing contender in safety‑focused large‑model development, has positioned its Science of Scaling group at the forefront of the debate on how massive neural networks can be made reliable and controllable. Attendees expect practical takeaways that bridge theory and product deployment.

Moskovitz’s research tackles a subtle but consequential failure mode: reward model overoptimization. In his ICLR 2024 spotlight paper, he demonstrated that as training progresses, models can continue to improve on a composite score while human evaluators perceive a decline in output quality. By reframing each objective as a hard constraint rather than a scalar to maximize, his constrained RLHF framework preserves alignment even as model size and data volume increase. Earlier work on policy drift (ReLOAD) and default policies further clarifies how optimization pathways affect generalization across tasks, offering a roadmap for more robust model training pipelines.

For enterprises building AI‑driven products, these findings have immediate commercial relevance. Reliable scaling reduces the risk of costly post‑deployment failures and regulatory scrutiny, especially in high‑stakes domains such as finance, healthcare, and autonomous systems. Moskovitz’s emphasis on constraint‑based optimization equips engineers with tools to embed safety checks directly into the training loop, shortening the gap between research prototypes and production‑grade deployments. As the industry pushes toward ever larger models, the ability to predict and control emergent behavior will become a competitive differentiator, making the insights shared at RAAIS 2026 essential for forward‑looking AI strategies.