The Bitter Layout or: How I Learned to Love the Model Picker — Maximillian Piras, Yutori

Key Concepts

• Bitter Layout: A common, uninspired AI app layout with an input field, turn-by-turn UX, and a model picker.
• Schrödinger's Chat: The paradoxical state where chatbots have usability issues but are still widely used.
• Model Picker: A UI element (usually a dropdown) that allows users to select from various AI models.
• Modes: Settings in a UI that drastically change input-output mappings (e.g., Caps Lock).
• Flexibility-Usability Trade-off: The inverse relationship between a system's flexibility and its usability.
• Zeitgeist: The defining spirit or mood of a particular period of history as shown by the ideas and beliefs of the time.
• Product Architecture (Integrated vs. Modular): The structure of a system and how its components interact, ranging from tightly coupled (integrated) to loosely coupled (modular).
• Bitter Lesson: The idea that relying on computation and scaling is more effective than hand-engineering AI systems.
• Bitter Design Lesson: If models aren't commoditized, the UI must conform to the next model's capabilities.
• Goals and Constraints: A programming paradigm focused on defining desired outcomes and limitations rather than step-by-step procedures.
• Generative UI: User interfaces created or assisted by AI models.

Schrödinger's Chat: The Dualistic Future of AI UX

The talk begins by observing the consistent layout emerging across AI-first applications, characterized by an input field, turn-by-turn UX, and a model picker. This layout feels retrofitted into a chatbot UX. The speaker introduces the concept of "Schrödinger's Chat," highlighting the paradox that chatbots, despite their usability issues, are widely used.

• The Chatbot Debate: The speaker traces the debate around chatbots back to 2022, citing Lionus Lee's argument against exposing raw text completion. Despite criticisms, ChatGPT's success has led to widespread adoption of chat-based interfaces.
• Arguments Against Chat: The speaker references posts from Amelia Wadenberger and Maggie Appleton in May and June 2023, respectively, who argued against chat as the future UI. Julian Lear argued in March of the current year that chat is a bottleneck.
• Arguments For Chat: The speaker notes the intuitive appeal of chat, evidenced by the meme-based defense of its widespread use.

Models and Modes: The Model Picker Problem

This section focuses on the model picker UI element and its potential usability issues.

• Larry Tesler and Modes: The speaker invokes Larry Tesler, known for inventing copy and paste and for disliking modes in UIs. Modes are settings that drastically alter input-output mappings (e.g., Caps Lock).
• Model Picker as a Mode Selector: The speaker argues that the model picker functions as a mode selector because switching models leads to significant changes in output.
• Example: An older version of ChatGPT is shown where the user has to find the specific model that supports the mode they want to use.
• Flexibility-Usability Trade-off: The speaker emphasizes the trade-off between flexibility and usability. More flexible systems, while accommodating more edge cases, often become less usable due to increased complexity.

The Context of All: Product Architecture and the Zeitgeist

This section discusses the importance of considering the current technological landscape (zeitgeist) when designing interfaces.

• Innovator's Solution: The speaker introduces the theory of product architecture from "The Innovator's Solution," which distinguishes between integrated and modular architectures.
  • Integrated Architectures: Common in early-stage disruption, proprietary, optimized, interdependent, and allow vertical scaling.
  • Modular Architectures: Common when technologies commoditize, interdependent, and allow horizontal scaling.
• IBM Example: IBM's evolution from integrated mainframe computers to modular personal computers illustrates the shift between architectures.
• Commoditization of Models: The speaker poses the question of whether AI models are commoditizing.

The Bitter Lesson and the Bitter Design Lesson

This section delves into the implications of the "bitter lesson" for UI design.

• Bitter Lesson (Rich Sutton): The speaker summarizes Rich Sutton's "bitter lesson," which argues that relying on computation and scaling is more effective than hand-engineering AI systems. As long as scaling laws are in effect, models are not commoditized.
• Bitter Design Lesson: If models aren't commoditized, the UI must primarily focus on conforming to the next model's capabilities. This leads to the "bitter layout," which is uninspired but easily adaptable to new models.
• ROI of the Bitter Layout: The speaker acknowledges the high ROI of the bitter layout, as it allows for quick integration of new models with minimal effort.

From Bitter to Sweet: The Future of AI UX

This section explores potential future directions for AI UX design.

• Brett Victor's "Future of Programming": The speaker references Brett Victor's talk, which advocates for shifting from procedural thinking to focusing on goals and constraints in programming.
• Designers and Goals/Constraints: The speaker suggests that designers are well-suited to this shift, as they can use tools like design systems and user stories to guide AI models.
• Design Systems for Generative UI: The speaker proposes using design systems to constrain and guide generative UIs created by AI models.
• User Stories as System Prompts: The speaker suggests translating user stories into system prompts to help models understand user goals.
• Dario Amadai Quote: The speaker concludes with a quote from Dario Amadai, who says that generative AI systems are more grown than built, suggesting a shift from construction to gardening in UX design.

Synthesis/Conclusion

The talk argues that the current "bitter layout" in AI applications is a consequence of prioritizing adaptability to rapidly evolving AI models over usability. As long as models continue to improve significantly, UIs will need to conform to their capabilities. However, the speaker suggests that the future of AI UX lies in shifting from procedural design to a focus on goals and constraints, leveraging designers' skills in creating design systems and user stories to guide AI models in generating user interfaces. The ultimate goal is to move beyond the "bitter layout" and create more intuitive and user-friendly AI experiences.