SUMM

Many AI-first applications, from chatbots to creative tools like coding assistants and Canva, have converged on a very similar layout: an input field, turn-by-turn UX, and a model picker dropdown.
There is a pervasive trend of retrofitting diverse AI tools into the chatbot user experience.
Despite ongoing debates about whether chat is the future interface for AI, actual usage patterns reinforce its dominance.
Designers and thought leaders have argued both against and in favor of chat interfaces, pointing out their usability issues versus their intuitive appeal and widespread practical adoption.
Even with convincing arguments against chat as a future paradigm, real-world usage often returns to chat-based interfaces.

The “model picker”—a dropdown to select among various AI models—has become a central UI element alongside chat.
The presenter references Larry Tesler’s disdain for “modes” in UI, arguing that the model picker itself acts as a mode selector, making user interactions less intuitive.
Example: Users have to match specific models to desired modes, which can be confusing and reflects a compromise in usability for flexibility.
This illustrates the “flexibility-usability trade-off”: As products aim to support more use cases (flexibility), overall usability for specific users tends to decrease.

Effective UI and product design depends on understanding the current context, including prevailing trade-offs and technological constraints.
Drawing on “The Innovator’s Solution,” the speaker introduces the spectrum from “integrated” (proprietary, optimized, vertically scaled) to “modular” (commoditized, horizontally scaled) product architectures.
The AI industry is continually shifting between these points; as components commoditize and decommoditize, designers must strategically decide which parts to integrate and which to modularize.
A key question: Are AI models themselves commoditizing, or does each new model still offer a significant leap?

“The Bitter Lesson,” from Rich Sutton, suggests we shouldn’t assume computational progress plateaus as long as scaling laws persist.
Each new AI model triggers intense interest, implying that models themselves are not yet commoditized.
This leads to the “bitter design lesson”: If model performance (inference) is the main competitive edge and models aren’t commoditized, then UI design must prioritize adapting to emerging models.
Current “bitter layouts”—unremarkable but flexible interfaces with model pickers—excel at integrating new model capabilities, making them pragmatic despite lacking inspiration or usability.
Until models fully commoditize and user needs are easily met, UI must prioritize absorption of new model capabilities.

Referencing Bret Victor’s “The Future of Programming,” the speaker advocates a paradigm shift: from thinking in procedural terms to focusing on higher-level goals and constraints.
This mindset shift is necessary as AI-assisted apps become more dynamic, stochastic, and probabilistic; designers can no longer anticipate every possible user journey.
The future may involve design systems that set goals and constraints, not just for developers and users but for collaboration with generative models.
Quality assurance and user stories may become analogous to reinforcement learning and system prompts when partnering with AI in design.
Dario Amodei’s notion that generative AI systems are “grown” rather than “built” provides inspiration to approach design more like gardening than construction.
Embracing these lessons can help UI/UX design move past the “bitter layout” and unlock new, adaptive possibilities in the era of AI.

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