# DeepSeek V4 vs. Opus: Ahmad Awais on AI Coding Taste
_Ahmad Awais discusses how AI coding agents can learn 'coding taste' to outperform generic models, focusing on the difference between functional code and good design._
**Published:** 2026-06-06
**Source:** https://www.startuphub.ai/ai-news/artificial-intelligence/2026/deepseek-v4-vs-opus-ahmad-awais-on-ai-coding-taste
---
In a recent discussion on the Latent Space podcast, Ahmad Awais, Founder & CEO of Command Code, delved into the nuances of AI coding agents and their ability to learn "coding taste." Awais, a seasoned developer with extensive experience in the open-source community and prior roles at major tech companies, shared his insights on why "open model bad at tool calling" is fundamentally a harness problem, not a model limitation.

Ahmad AwaisCore

Founder & CEO of Command Code, expert on AI codingFrom the article 4 mentionsAwais highlighted that while benchmarks often focus on technical correctness, they often miss crucial aspects like design taste and user experience.Work-Pattern-FirstContextcomposition approach prioritizing user's workflow and habitsFrom the articleAwais elaborated on the concept of "work-pattern-first composition," explaining that an AI agent should first identify the underlying patterns and intent behind a user's request before generating code.Design Taste ImportanceContextcrucial for AI development beyond just functional codeFrom the article 3 mentionsThe core of Awais's argument centered on the idea that "design taste" is not merely a cosmetic issue but a fundamental aspect of building effective AI tools.Harness ProblemDriveropen models struggle with tool calling due to integrationGeneric AI CodeDriversyntactically correct but lacks user's design preferencesFrom the article 5 mentionsHe posited that the key to improving AI coding is to train models that can continuously learn and adapt to a user's specific coding style and preferences over time, moving beyond generic, rule-based outputs.AI Coding TasteContextlearning user's specific style and preferences over timeFrom the article 4 mentionsAwais explained that while LLMs can write fluent code, they often lack genuine design taste.Outperform Generic ModelsEffectAI agents with taste provide better, personalized codeFrom the article 2 mentionsHe contrasted this with models that might simply follow a generic template, leading to a less refined and less personalized user experience.Continuous LearningEffectmodels adapt to user's evolving coding style

## The 'Taste' of AI in Code Generation

Awais explained that while LLMs can write fluent code, they often lack genuine design taste. This means that while the output might be syntactically correct, it doesn't necessarily reflect the user's preferences or adhere to good design principles. He posited that the key to improving AI coding is to train models that can continuously learn and adapt to a user's specific coding style and preferences over time, moving beyond generic, rule-based outputs.

## Benchmarking and Evaluating AI Models

The conversation touched upon the challenges of evaluating AI models, particularly in areas like design. Awais highlighted that while benchmarks often focus on technical correctness, they often miss crucial aspects like design taste and user experience. He drew a parallel to how human designers intuitively understand and apply these principles, something current AI models struggle to replicate. This, he suggested, is a significant gap that needs to be addressed for more sophisticated AI development tools.

## The Importance of 'Work-Pattern-First' Composition

Awais elaborated on the concept of "work-pattern-first composition," explaining that an AI agent should first identify the underlying patterns and intent behind a user's request before generating code. This approach allows the AI to create more contextually relevant and aesthetically pleasing outputs. He contrasted this with models that might simply follow a generic template, leading to a less refined and less personalized user experience.

## The Role of Design Taste in AI Development

The core of Awais's argument centered on the idea that "design taste" is not merely a cosmetic issue but a fundamental aspect of building effective AI tools. By understanding and incorporating user preferences, AI models can move beyond simply generating functional code to creating solutions that are also intuitive, efficient, and aesthetically pleasing. This, he believes, is the next frontier in AI-assisted development.
---
Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory.