OpenAI

Shaping model behavior in GPT-5.1— the OpenAI Podcast Ep. 11

What does it mean for an AI model to have "personality"? Researcher Christina Kim and product manager Laurentia Romaniuk talk about how OpenAI set out to build a model that delivers on both IQ and EQ, while giving people more flexibility in how ChatGPT responds. They break down what goes into model behavior and why it's an important, but still imperfect blend of art and science. Chapters - 00:00:43 — GPT-5.1 goals and the shift to reasoning models - 00:02:18 — Differences between GPT-5 and GPT-5.1 - 00:04:55 — Unpacking the model switcher - 00:07:24 — Understanding user feedback - 00:08:27 — Measuring progress on emotional intelligence - 00:10:02 — What is model personality? - 00:14:25 — Model steerability, bias, and uncertainty - 00:21:59 — Advantages of memory in ChatGPT - 00:25:27 — Looking ahead and advice for getting the most out of models

AMAndrew MaynehostLRLaurentia RomaniukguestCKChristina Kimguest

Dec 2, 2025·28m·Watch on YouTube ↗

📖 CHAPTERS

1. 1

   What’s new in GPT‑5.1: reasoning as the default in ChatGPT

   The hosts frame the episode around GPT‑5.1 improvements, especially behavior and “personality” steerability. Christina explains the core shift: all chat models are now reasoning-capable, letting the system decide when deeper thinking is needed.

2. 2

   From GPT‑5 to 5.1: addressing “coldness,” intuition, and missed context

   Laurentia describes community feedback that GPT‑5 felt less warm and had weaker intuition. The team found causes beyond the base model—like context carryover and switching behavior—then tuned the overall experience to feel more consistent and human-friendly.

3. 3

   Custom instructions and user control: making quirks steerable

   A major practical improvement is stronger adherence to custom instructions across turns. The team emphasizes that users tolerate model quirks better when they can reliably correct or steer them and have those preferences persist.

4. 4

   Why model switching exists—and why it confuses people

   Andrew probes the “multiple models” reality behind a single ChatGPT experience. Laurentia explains the product challenge: mapping different model capabilities to user needs via UI and an auto-switcher that predicts which mode best serves a prompt.

5. 5

   Impact of giving everyone a reasoning model: toward a ‘system of models’

   Christina expands on what it means for the free/base experience to be reasoning-enabled. She argues the future isn’t a single set of weights but an ecosystem: multiple reasoning tiers, a switcher model, and tool-backed models working together.

6. 6

   Turning user feedback into engineering work: conversation links, signals, and tradeoffs

   With massive usage, feedback triage depends on seeing the full conversation context. Laurentia explains how share links enable debugging, and how switching decisions balance factuality, helpfulness, and latency preferences.

7. 7

   Measuring ‘EQ’ in models: user-signal research, memory, and context

   Andrew asks how to evaluate emotional intelligence compared to IQ-style benchmarks. Christina highlights “user signals research” and reward modeling, while Laurentia ties EQ to listening, remembering, and matching a user’s preferred style.

8. 8

   What ‘personality’ really means: tone settings vs the whole product harness

   Laurentia disentangles “personality” as a feature (response style/traits) from personality as the entire ChatGPT experience. She argues fonts, latency, rate limits, and model routing all shape what users interpret as personality.

9. 9

   Post-training as an art: reward design, warmth, and steerability vs safety

   Christina explains the balancing act in post-training: optimizing many goals simultaneously through subtle reward configuration choices. Laurentia connects this to the Model Spec’s principle of maximizing freedom while minimizing harm—without hard-coding away user-desired behaviors.

10. 10

    Evolving safety behavior: from refusals to ‘safe completions’ and nuanced context

    The discussion recalls early ChatGPT’s heavy refusals and prompt-hack era. Laurentia describes progress toward “safe completions,” where the model tries to help within boundaries, plus the difficulty of handling sensitive domains like legal work without harming legitimate use.

11. 11

    Bias, uncertainty, and creativity: expanding expressive range in 5.1

    Laurentia points to improvements in subjective domains: expressing uncertainty, engaging open-ended questions, and staying grounded in objective truths. She also calls creativity a “sleeper feature,” with broader expressive range when users push the model stylistically.

12. 12

    The future of customization: inferred expertise, user control, and transparency

    Looking forward, both guests argue one default personality can’t serve hundreds of millions of users. They predict more inferred adaptation (expertise-aware responses) alongside explicit controls, with transparency so users can see and adjust what the system infers.

13. 13

    Memory in practice: what it stores, why it feels warmer, and proactive experiences

    Christina explains memory as stored user facts/preferences used across future chats to reduce repetition and improve grounding. Andrew shares how memory enables proactive, personalized updates, while Laurentia stresses that better models enable downstream product features.

14. 14

    Getting the most out of GPT‑5.1: pressure-testing, iteration, and prompt help

    The episode closes with practical advice: keep experimenting as models update continuously, and test the model on domains you know well. Christina recommends asking the model to help generate better prompts; Laurentia shares her own frequent style-switching habits.