Shaping Model Behavior in GPT-5.1
This episode of the OpenAI Podcast focuses on GPT-5.1, with research lead Christina Kim and product manager Laurentia Romaniuk explaining how the team approached shaping the model's behavior and personality. A major milestone is that, for the first time, all models in ChatGPT are reasoning models that can dynamically decide how much to think based on the prompt, drawing on Kahneman's System 1 / System 2 framing. The release is framed as a direct response to community feedback that GPT-5 felt less intuitive and warmer, with improvements to context retention, instruction following, and the auto-switcher that previously produced jarring transitions between chat and reasoning modes. A central theme is the distinction between "personality" as a narrow response-style feature (concise vs. lengthy, emoji usage, etc.) and personality as the entire user experience, which includes the harness: context window, memory, latency, rate limits, and the model chosen by the switcher. The team discusses balancing steerability with freedom via the Model Spec, the role of safe completions in replacing blunt refusals, and progress on bias and creativity. They also describe how user-signal research, memory, and priming can let models infer user expertise and tailor responses proactively while keeping users in control. Looking ahead, the guests argue that one model personality cannot serve 800 million weekly users, so customization and steerability are the path forward. They encourage users to keep pressure-testing models on topics they know well, since capabilities shift quickly between releases, and to use memory and personality settings to shape the experience they want.
Original sources
Podcast OpenAI Podcast
YouTube OpenAI Podcast
Video:Shaping model behavior in GPT-5.1— the OpenAI Podcast Ep. 11 ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.