AI Engineer World's Fair 2025

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

Loading slides…

Read the talk

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

A prompt box and a crowded model picker can be frustrating to use, yet remarkably good at absorbing new capabilities. That trade-off changes what AI interface design must optimize for.

From a talk by Maximillian Piras

Different tools, the same layout

Open a chatbot, an answer engine, or a creative tool such as v0 or Canva, and the interface increasingly looks familiar: an input field, a sequence of conversational turns, and a dropdown containing more models than you know how to choose between. Products with different purposes appear to be fitting their capabilities into the same chatbot structure. That convergence is the starting problem in Maximillian Piras’s The Bitter Layout.

Four screenshots arranged in a grid show the prompt interfaces of v0, Bolt, Lovable, and Canva.Open full source frame
v0, Bolt, Lovable, and Canva place prompts at the center of their interfaces.

The obvious response is to propose a better interface. But there is an awkward gap between demonstrating an alternative and replacing the tool people actually use. Piras recalls Swyx’s criticism of designers who declare that chat is not the future, show an impressive demo, and then return to ChatGPT. The criticism lands because designers themselves keep using an interface whose limitations they can readily enumerate.

Piras calls this tension Schrödinger’s Chat: chat seems obviously central to the future because people use it every day, and obviously inadequate because its usability problems remain. The useful question is not yet which side wins. It is how those two observations can both be true.

0:230:31
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:23 · section reference included

Usability critiques meet adoption

The debate predates ChatGPT. Piras traces it to a May 2022 post by Linus Lee questioning whether exposing raw text completion was the right long-term interface paradigm. ChatGPT’s subsequent success did not settle that design question, but it challenged predictions about what users would accept. An interface could violate a designer’s expectations and still become the tool people reached for.

The next wave of criticism included Amelia Wattenberger’s Why Chatbots Are Not the Future of Interfaces and Maggie Appleton’s Language Model Sketchbook, or Why I Hate Chatbots, which Piras places in May and June 2023 respectively. These critiques need not imply that chat has no useful properties: Appleton also acknowledges its flexibility, familiarity, and ease of implementation. Meanwhile, defenders pointed to chat’s intuitiveness and growing adoption.

That defense eventually became a meme: objections to chat were answered by pointing at an adoption chart. The slide featuring Joe Kennedy captures the exchange. Adoption makes the defense immediately understandable, but it does not explain away the usability objections; it shows that those objections have not prevented people from using the product.

Slide labeled Joe Kennedy, April 2024, with the quote “Sorry guys chat is not the right interface. Just ignore this chart you gotta trust me” beside a chart comparing technology adoption.Open full source frame
Joe Kennedy’s chat-interface meme pairs a skeptical quote with an adoption chart.

By March 2025, Julian Lehr’s The case against conversational interfaces was examining communication speed across interface paradigms. Piras uses it to frame conversation as a potential bottleneck. Lehr’s position favors conversational augmentation rather than wholesale replacement of existing interfaces, and his speed comparisons are illustrative rather than a controlled interface benchmark. The tension remains: a persuasive account of chat’s limits can coexist with another day spent using it.

1:512:00
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

1:51 · section reference included

The model picker adds another mode

Alongside chat, another interface convention has become standard: the model picker. To understand its cost, Piras turns to Larry Tesler’s opposition to modes. A mode is a setting that changes how an input maps to an output. Caps Lock is the familiar example: enable it, and the same letter key produces a different result. The user must track not only what action to take, but also the state in which that action will occur.

Piras treats model selection as a related kind of mode change. Generative output is already stochastic, so a model does not provide one fixed response for each input. Nevertheless, switching models can produce a substantial change in behavior. The prompt field looks the same while the system behind it behaves differently.

An older ChatGPT demonstration makes the interaction cost concrete. Piras tries to use a feature that the selected model does not support, then searches the model menu for one that does. Instead of simply choosing the desired operation, the user must also find a compatible model. Model choice and feature choice become interacting settings. Piras explicitly notes that the interface had already been improved by the time of the talk; the demonstration captures an earlier frustration, not a claim about current ChatGPT behavior.

One tempting escape is to describe the use case and ask the model which model and mode to use. Piras raises that possibility, then jokes that the model might select itself. It remains a suggestion, not a demonstrated routing solution.

3:313:43
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:31 · section reference included

An interface has a useful lifetime

The picker exposes the flexibility–usability trade-off. When a team understands a narrow set of user needs, it can optimize an interface around them. As the product supports more possibilities, it accumulates requirements, edge cases, and choices. Flexibility gives users more ways to achieve an outcome, but it also makes the system harder to understand and operate.

That changes the question from whether chat is the future to which trade-offs are acceptable over a particular timeframe. An interface should be evaluated against the capabilities, constraints, and uncertainty of its technological moment. A specialized interaction may be excellent for a stable task; a flexible interaction may remain useful across changes that would force the specialized one to be redesigned. The design question includes how long the interface can remain appropriate.

5:465:57
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:46 · section reference included

Where to integrate, where to stay modular

Piras uses a framework from The Innovator’s Solution, the follow-up to The Innovator’s Dilemma, to connect that interface question to product architecture. Architecture describes how a system’s components interact and where their interfaces sit. The framework places architectures along a spectrum between integration and modularity.

AttributeIntegratedModular
Typical settingEarly disruptionCommoditizing technology
ComponentsProprietary, interdependentMore independent
EmphasisJoint optimizationInterchangeable parts
Scaling in Piras’s framingVerticalHorizontal

Integration allows interdependent components to be optimized together. Modularity becomes more attractive when components can operate independently and their boundaries are sufficiently established.

These are not permanent destinations or a one-way sequence. Different parts of an industry’s technology stack can commoditize and become differentiated again at the same time. Builders must identify the particular boundaries where integration creates an advantage and those where modularity is more useful. A single product can require both.

Blue arrows loop between grouped and separated shape icons around a headline about continuous commoditization and de-commoditization, with a quotation about shifting points of integration.Open full source frame
Commoditization and de-commoditization form a continuous cycle.

Piras illustrates the shift with IBM: integrated mainframes, a more modular personal-computer architecture, and eventually a commoditized PC business that IBM exited. The example gives the design problem a strategic dimension. As the basis of competition changes, the valuable points of integration can move.

6:507:02
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

6:50 · section reference included

The interface must accommodate the next model

Apply that framework to AI and the immediate question is whether models themselves are becoming commodities. Piras introduces Rich Sutton’s The Bitter Lesson with a deliberately compressed interpretation: do not assume computation remains constant while scaling continues to matter. The attention paid to each new model release suggests, in his account, that model capability still differentiates products.

From there, Piras makes his own design inference. If models remain differentiated, and inference performance is the basis of competition, then the interface’s primary job is to accommodate the next model. He calls this the bitter design lesson: under those conditions, the interface becomes the commodity. The balance could change when models exceed what a particular use case requires. Once additional capability stops mattering to that user, there is more room to compete through specialized interface integration.

This explains why the generic layout survives despite its shortcomings. A prompt field, conversation history, and model picker can absorb new capabilities without requiring an entirely new interaction structure. Release a stronger model, make it available in the picker, and users can access it through the pattern they already know. The layout may be uninspiring or awkward, but its adaptability has practical value.

Piras describes the return as adding one line item and making the app “N× more intelligent”—rhetoric about the appeal of the upgrade, not a measured improvement. The durable point is that substantial capability changes can arrive with little visible interface change. He remains unconvinced that chat is the ultimate interface, but identifies its ability to accommodate the next model as an attribute future AI interfaces need to preserve until models commoditize.

8:318:49
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:31 · section reference included

Design above the level of individual procedures

What would moving beyond that layout require? Piras turns to Bret Victor’s The Future of Programming and its account of resistance to moving from binary programming to SOAP, the historical IBM 650 assembler, not the web-services protocol. Programmers accustomed to hand coding were reluctant to surrender control to a machine and trust a higher level of abstraction. In retrospect, accepting that abstraction was an important shift.

The corresponding design shift is from prescribing procedures to expressing goals and constraints. Detailed user flows and edge-case analysis still matter. But as applications become more dynamic and probabilistic, a designer cannot anticipate every possible path. Instead, part of the work moves up a level: define the outcome users should reach and the boundaries within which the system should help them reach it.

11:1311:24
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

11:13 · section reference included

From design systems to gardening

Existing design practices offer starting points for working at that higher level:

  • Design systems → generative UI. Design systems already guide designers and developers by setting constraints. They could also bound the interfaces a model generates, providing a shared basis for collaboration.
  • Quality assurance → feedback loops. Critiquing a model’s design could become a loop that guides subsequent output. Piras tentatively compares this to reinforcement learning; he does not establish that ordinary design critique constitutes an RL training process.
  • User stories → system prompts. User stories express what a person is trying to accomplish. Translating them into system prompts could help communicate that intent to a model, which might also participate in clarifying the user’s goal.

These mappings extend familiar design work rather than requiring designers to imagine every possible generated interaction in advance.

Slide maps procedures to goals and constraints, design systems to generative UI, quality assurance to reinforcement learning, and user stories to system prompts.Open full source frame
From procedures to goals and constraints: design systems, quality assurance, and user stories map to new approaches.

Piras explicitly presents these possibilities as speculative. He closes with a remark he attributes to Dario Amodei: generative AI systems are more grown than built. The metaphor suggests a different relationship to the interface. Instead of constructing every interaction directly, designers shape the conditions under which useful interactions can emerge.

Gardening still requires deliberate choices about goals, boundaries, and intervention. It changes where that effort is applied. For Piras, learning to guide a probabilistic system at this higher level is a possible route beyond the bitter layout—while retaining the adaptability that made the layout useful in the first place.

12:1812:44
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:44 · section reference included

Resources

From the talk

Transcript download
Full transcript

Loading transcript…