Videos

This Build Hour session, led by OpenAI's Michaela and Emre Okcular, focuses on Context Engineering as a critical discipline for building robust, long-running AI Agents. Emre defines context engineering as both an art and a science, emphasizing its role in optimizing LLM performance beyond just model quality or prompt engineering. The session identifies four common failure modes for long-running agents: context bursting, conflict, poisoning, and noise. To address these, three core strategies are introduced: Reshape + Fit (adapting context window through trimming, compaction, summarization), Isolate + Route (directing context to specific sub-agents), and Extract + Retrieve (creating high-quality, retrievable memories). Emre demonstrates these techniques using a Next.js and OpenAI Agents SDK-based IT troubleshooting agent, showcasing how summarization can create persistent, cross-session memory for personalized interactions. The session concludes with best practices for Agent memory design, including understanding typical context, deciding when to remember and forget, and implementing memory-specific evaluations, offering valuable insights for developers aiming to build stateful and intelligent AI applications.

This OpenAI podcast episode features Christina Kim (Post-Training Research Lead) and Laurentia Romaniuk (Product Manager for Model Behavior) discussing the development and goals of GPT-5.1. A key highlight is the transition to making reasoning models the default for all users in ChatGPT, enabling models to 'think' before responding, similar to Daniel Kahneman's System 1 and System 2 thinking. The conversation delves into the challenge of defining and shaping 'model personality,' emphasizing that it encompasses the entire user experience, including context retention and the new memory feature, which significantly enhances personalization. They also address the delicate balance between maximizing user freedom and ensuring model safety and steerability, noting the evolution of safety systems to provide 'safe completions' rather than outright rejections. The discussion touches on measuring 'EQ' in models, the importance of user feedback in driving product updates, and offers advice for users to get the most out of these evolving AI capabilities, such as stress-testing models with complex queries and utilizing custom instructions or memory.