Scott and Mark Learn to Vibe Check

Key Concepts

• AI Slop: Low-quality, unrefined, or purposeless AI-generated content/code.
• Vibes (Vibe Coding): An iterative, experimental approach to software development where the developer relies on AI to generate code, often prioritizing speed and intuition over traditional architectural rigor.
• AI-Augmented Software Engineering: The professional application of AI tools (Copilot, Claude, etc.) to build robust, maintainable, and high-precision software, characterized by human oversight, testing, and architectural control.
• Hallucinated Operating System: A conceptual UI/UX simulation where the interface is generated in real-time by an AI model without underlying traditional logic or event handlers.
• Agentic Workflows: Systems where AI agents perform tasks (like scheduling or data querying) with human-in-the-loop verification.

---

Overview

Scott Hanselman and Mark Russinovich hosted "Scott and Mark Learn to Vibe Check" at Microsoft Build, a session designed to evaluate whether community-built AI projects fall into the categories of "slop," "vibes," or "AI-augmented software engineering." The session featured four developers demonstrating unique applications built with AI assistance.

---

Featured Projects and Methodologies

1. Vibe OS (Steve Sanderson)

• Concept: A "hallucinated" operating system where applications (Notepad, Calculator, Internet Explorer) have no underlying code or logic.
• Methodology: Each window is an iframe running a separate Copilot session. The AI generates HTML/UI in real-time. When a user interacts with an element (e.g., clicking a button), the system sends the element's ID back to the AI, which then returns a "diff" to update the UI.
• Key Insight: While technically "slop" by traditional standards, it represents a high-level engineering feat in prompt orchestration and state management.

2. CSS-Based Database (Cassidy Williams)

• Concept: A functional query shell and database interface built entirely within the browser using CSS.
• Methodology: While JavaScript and WebAssembly load the SQLite database, the actual querying, filtering, and sorting logic are handled by CSS selectors, variables, and Flexbox.
• Key Insight: A "cursed" but brilliant application that requires deep domain knowledge of CSS to manipulate the browser engine into performing non-standard tasks.

3. AI E-Bot Scheduler (Swix)

• Concept: An internal tool for managing conference logistics (scheduling 6,000+ attendees and speakers) that replaces expensive SaaS products.
• Methodology: Built on Cloudflare using a "wild loop" agentic harness. It integrates with Slack and email to allow the user to assign speakers to tracks via natural language.
• Key Insight: Emphasizes the importance of "human-in-the-loop" verification for high-precision tasks to prevent AI hallucinations from causing real-world logistical errors.

4. Datasette Agent (Simon Willison)

• Concept: A plugin-based agent for the "Datasette" tool that allows users to query and visualize personal data (e.g., 24 years of blog posts) using natural language.
• Methodology: Uses a Python-based plugin architecture. It includes a WebAssembly sandbox to safely execute Python code generated by the AI.
• Key Insight: Demonstrates "adversarial vibe coding," where multiple models are used to critique each other’s code to ensure security and correctness.

---

Key Arguments and Perspectives

• The "Expertise" Requirement: All presenters agreed that while AI accelerates development, deep domain expertise (CSS, Python, OS internals) is required to "steer" the AI effectively. You cannot "vibe" your way into production-grade software without understanding the underlying principles.
• The Role of Trust: Simon Willison argued that trust in AI-generated code is built over time through usage and testing, rather than just by reading the code itself.
• Human-in-the-Loop: The consensus was that unbounded, fully autonomous AI is dangerous. High-stakes applications require human oversight, audit logs, and rollback capabilities.
• The Future of Voice: The hosts and guests suggested that voice interaction should be an OS-level feature rather than being redundantly implemented in every individual application.

---

Synthesis and Conclusion

The session concluded that the line between "vibes" and "AI-augmented software engineering" is defined by the developer's intent and their ability to maintain control. While "vibe coding" is an excellent tool for rapid prototyping and exploration, professional software engineering remains essential for production environments. The most successful projects were those where the developer used AI to handle the "typing" and boilerplate, while they maintained strict control over the architecture, security, and logic.