# The Bitter Layout, or How I Learned to Love the Model Picker

Maximillian Piras, Yutori | AI Engineer World's Fair 2025 | 14:24

Source: https://www.youtube.com/watch?v=BZtD0yYAgCQ
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/the-bitter-layout-or-how-i-learned-to-love-the-model-picker
Published: 2025-07-21
Tags: design, product-strategy

## TL;DR
- AI applications keep converging on an input field, turn-by-turn interaction, and a model picker because this layout can absorb new model capabilities quickly.
- As long as scaling laws keep improving inference performance, AI interfaces need to conform to changing models rather than depend on tightly integrated features.
- Designers will need to work at a higher level, using goals and constraints to guide dynamic interfaces instead of specifying every possible user path.

## Summary
Maximillian Piras examines why AI products keep adopting what he calls the "bitter layout": an input field, turn-by-turn interaction, and a long model picker. He connects this pattern to Clay Christensen's theory of integrated and modular product architectures, then applies Rich Sutton's bitter lesson. If inference performance remains the basis of competition and new models keep mattering, interfaces need to stay modular so they can absorb the next model's capabilities. The result is often awkward for users, but it gives teams a simple way to make an app more capable. Piras then looks beyond this compromise. Drawing on Brett Victor, he argues that designers should move from specifying procedures toward setting goals and constraints. Design systems, user stories, critiques, and system prompts may become ways to guide generative interfaces. He ends with Dario Amodei's idea that generative AI systems are more grown than built.

## Key ideas
### AI products are converging on the same bitter layout
[00:03](https://www.youtube.com/watch?v=BZtD0yYAgCQ&t=3s)
Piras sees the same structure across chatbots, answer engines, creative tools, coding assistants such as v0, and Canva. They use an input field, a turn-by-turn flow, and a dropdown containing many models. He calls this convergence the "bitter layout" because it feels like products are retrofitting different tasks into a chatbot interface. The design has well-known usability problems, but it keeps appearing across categories. Piras does not try to settle whether chat is the future. He frames the situation as a tension: designers can explain why chat is a poor interface and still return to using ChatGPT every day.

### The model picker behaves like a mode selector
[02:48](https://www.youtube.com/watch?v=BZtD0yYAgCQ&t=168s)
Piras compares model selection with UI modes, where changing one setting makes the same input produce a very different output. Caps Lock is his simple example. A model picker creates a similar shift because changing models can produce a step change in the result, even though the setting is not a conventional mode. In an older version of ChatGPT, a requested capability might only work with a particular model, forcing the user to find that model in a menu. This creates modes on top of modes: users must match a model to the mode they want. Piras says the model picker is therefore flexible, but difficult to understand.

### Flexibility reduces usability when user needs are unclear
[05:46](https://www.youtube.com/watch?v=BZtD0yYAgCQ&t=346s)
The model-picker problem illustrates the flexibility-usability trade-off. When a product team understands a user's needs well, it can create a focused flow with fewer choices. A more flexible system covers more cases, but it also adds edge cases, complexity, and requirements. Piras says the chat debate is often framed as a yes-or-no question when the real issue depends on timing and trade-offs. An interface can be appropriate under one set of technical constraints and less appropriate later. Designers need to ask what the current constraints are, what users need now, and how long a particular interface can remain useful.

### AI product architecture shifts between integration and modularity
[06:58](https://www.youtube.com/watch?v=BZtD0yYAgCQ&t=418s)
Using ideas from The Innovator's Solution, Piras describes product architecture as the way components in a system interact. Integrated architectures are common in early disruption. They use proprietary, optimized parts that support vertical scaling. Modular architectures become more useful as technologies commoditize, because interchangeable components allow horizontal scaling. Industries move between these conditions, and different parts of a technology stack can commoditize or become differentiated at the same time. Piras uses IBM's history as an example, from integrated mainframes to more modular personal computers, followed by the eventual commoditization of the computer itself. Designers should identify which parts of AI are changing before deciding how tightly to integrate an interface.

### Scaling laws keep models from being treated as commodities
[08:42](https://www.youtube.com/watch?v=BZtD0yYAgCQ&t=522s)
Piras applies Rich Sutton's bitter lesson to interface design. His version is that teams should not assume computation will remain constant while scaling laws continue to work. Each major model release still changes what people want to test, and new models continue to attract attention. From this, he concludes that models have not yet commoditized. If inference performance remains the basis of competition, the interface must be able to adapt to the next model's capabilities. This is why the model picker persists, even when it adds friction. The technical change underneath the interface has not settled, so a fixed, tightly integrated product would risk becoming outdated.

### The bitter layout earns its place by absorbing new models
[09:30](https://www.youtube.com/watch?v=BZtD0yYAgCQ&t=570s)
Piras calls the resulting design rule the "bitter design lesson": while models remain differentiated, the interface has to conform to the next model. The layout may be uninspiring and hard to use, but it can absorb a new model with limited product work. A team can add another model to the picker and make the application more intelligent without rebuilding the whole experience. Piras dislikes the design, yet he admits that the return on investment is hard to dismiss. The layout is modular at the interface level because model capability remains unstable. He expects more integrated interface features once models exceed what users need for ordinary tasks.

### Designers need to guide stochastic interfaces with goals and constraints
[11:20](https://www.youtube.com/watch?v=BZtD0yYAgCQ&t=680s)
Drawing on Brett Victor's The Future of Programming, Piras asks designers to move from procedural thinking toward goals and constraints. Earlier programmers had to shift from writing binary code to working through abstraction layers. AI interfaces create a related problem because designers cannot specify every possible path when the application is dynamic and probabilistic. Design systems could provide constraints for generative UI. Quality assurance might become part of a reinforcement-learning loop, though Piras presents that as speculation. User stories could describe user goals and become system prompts. His proposed shift is from designing every procedure to setting conditions that help a model produce an acceptable result.

### Generative systems may be grown through guidance rather than fully built
[13:38](https://www.youtube.com/watch?v=BZtD0yYAgCQ&t=818s)
Piras closes with Dario Amodei's view that generative AI systems are more grown than built. He uses the idea to suggest a different design mindset. Instead of treating an interface as a fixed construction with every path specified in advance, designers might guide its behavior through systems, goals, constraints, and feedback. This does not solve the usability problems of the bitter layout immediately. It gives designers a way to work while applications remain stochastic and models continue to change. Piras presents the move toward a higher level of abstraction as a possible route beyond today's model picker and chatbot-centered layouts.

## Notable quotes
- "The one thing this does really well is it can absorb the next model's capabilities." (10:07)
- "If the model's not commoditized, it's actually the interface that's the commodity now." (09:49)
- "You should stop thinking in terms of procedures and start trying to think of programming in terms of goals and constraints." (11:56)
- "Generative AI systems are more grown than they are built." (13:38)

## Tools & references mentioned
- Yutori
- v0
- Canva
- ChatGPT
- OpenAI
- Larry Tesler
- Clay Christensen
- The Innovator's Solution
- The Innovator's Dilemma
- IBM
- Rich Sutton
- The Bitter Lesson
- Brett Victor
- The Future of Programming
- Dario Amodei
- Amelia Wattenberger
- Maggie Appleton
- Julian Lehr
- Linus Lee

## Who should watch
- You are designing an AI product and need to decide whether to build tightly around one model or keep the interface adaptable.
- Your product has a model picker with many choices, and you want a framework for understanding the usability cost behind that flexibility.
- You are exploring generative UI and need practical design concepts for setting goals and constraints when user flows cannot be fully specified.

## Related talks

- [Form factors for your new AI coworkers](https://aietalks.com/talks/form-factors-for-your-new-ai-coworkers) (Craig Wattrus, Flatfile, 15:35)
- [The Missing Layer: Design Taste in AI Agents](https://aietalks.com/talks/the-missing-layer-design-taste-in-ai-agents) (Hassan El Mghari, Together AI, 14:10)
- [Designing AI To Scale Human Thought](https://aietalks.com/talks/designing-ai-to-scale-human-thought) (Jun Yu Tan, Tusk, 12:24)
- [Climbing the Ladder of Abstraction](https://aietalks.com/talks/climbing-the-ladder-of-abstraction) (Amelia Wattenberger, Adept.ai, 16:47)
- [The Intelligent Interface](https://aietalks.com/talks/the-intelligent-interface) (Samantha Whitmore & Jason Yuan, New Computer, 18:56)
