Why AI Labs Struggle to Build Superior User Applications
Technical proficiency in code generation does not equate to high-quality product design decisions.
1The Paradox of AI Apps
Despite frontier models demonstrating exceptional coding capabilities, applications released by major AI labs often suffer from poor user experience. Ben Davis highlights the ChatGPT desktop app as a prime example of this phenomenon, noting its cluttered interface and disjointed menus. This raises questions about why powerful models produce such fragmented software products.
2Vibe Coding Is Misunderstood
Critics often dismiss modern AI-generated software as low-quality 'vibe-coded slop,' but this assessment is fundamentally flawed. Modern models like Opus 5.5 generate code that is significantly more stable than amateur efforts from 2009. The issue is not the model's ability to write functional code, but rather the architectural and design choices made.
3Programmatic vs. Subjective Tasks
Models excel at hill-climbing tasks where a clear programmatic verification function exists, such as optimizing code speed. However, they struggle with subjective decisions like color palettes or dashboard layouts. These areas require human taste, which currently lacks a quantifiable metric that models can effectively interpret or replicate during the development process.
4The Limits of Automated Testing
A recurring frustration for developers is that AI-generated code often passes all automated tests while still feeling unusable. The implementation may be technically correct, yet the overall product design fails to meet user needs. This gap between passing tests and creating a cohesive, intuitive experience remains a significant hurdle for teams.
5Human Discernment Remains Essential
Human taste is defined as the ability to make sound decisions in the absence of a programmatic verifier. As demonstrated by MacStories' development process, human input is still required for usability, onboarding flows, and iOS best practices. Models have not yet reached the level of abstraction necessary to handle these nuanced design requirements.
6Feature Bloat and Design
Big labs may be prioritizing infinite token usage and rapid feature deployment over refined software design. By cramming multiple modalities and sidebars into a single interface, they ignore the importance of simplicity. This 'kitchen-sink' approach to product development often results in a disjointed user experience that prioritizes technical capability over actual utility.
7Successful Agent-Driven Products
Despite the struggles of large labs, high-quality products can be built using agents when human oversight is properly applied. Tools like T3 Code and Grok Bot demonstrate that agents can be effective when guided by strong product vision. The key is using models as tools rather than relying on them for all decisions.
8The Future of AI Development
The current state of AI-driven software design is rapidly evolving, making today's observations potentially obsolete within weeks. While models will likely improve their design capabilities, human judgment remains the critical differentiator in creating great software. For now, human discernment is the only force capable of turning functional code into excellent products.
Original article: MacStories — Why AI Labs Struggle to Build Superior User Applications