Only Good Devs Catch AI's Mistakes, and AI Makes You Worse - Lars Faye | Ep. 97

Summary

The episode explores how AI coding tools can speed up development while potentially weakening developers’ skills and making projects harder to maintain. Guest Lars Faye argues that AI is useful, but its benefits depend heavily on the user’s experience and judgment.

Key points

• AI can accelerate experienced developers, but expertise matters. Faye says he first used ChatGPT to generate WordPress blocks and later adopted coding agents to reduce repetitive work. He uses tools such as Cursor and OpenCode, and has tried Claude Code and models including GLM and Kimi. He emphasizes that capable developers can assess and correct AI output; less experienced users may not recognize faulty solutions.

• The main risk may be project composition, not bad-looking code. Faye distinguishes syntactically sound code from a coherent, maintainable codebase. Because a language model generates patterns without an overarching understanding of a project, disconnected decisions can accumulate into a “rat’s nest” that becomes difficult to extend or hand off. He warns that teams may become dependent on AI subscriptions to understand or maintain such projects.

• AI-assisted work can impede learning. Faye describes relying on an agent for a project in a framework he did not know well, then finding it difficult to debug the code himself. He argues that observing an AI solve problems is not a substitute for the practice and trial and error needed to build expertise. The host counters that AI-assisted auditing and problem-solving can still create useful forms of learning. Faye agrees there is some friction, but says it may not be enough to develop fundamentals on its own.

• His practical approach is selective assistance, not full delegation. Faye uses AI to save typing and handle repetitive tasks while staying close to the code and reviewing its output. He describes using prior templates, prompts, and specifications to delegate portions of work rather than simply asking an agent to build an entire project. He also says he avoids some workflow automations because of privacy and data-access concerns.

• AI can be useful for low-stakes personal tools. Faye built a private time-tracking tool with AI assistance and says he was less concerned about code quality because it was internal and only needed to work. He and the host also discuss AI-generated workflows for tasks such as handling video uploads and preparing descriptions, titles, and transcripts.

• WordPress and custom block development. Faye says he chooses technologies based on project needs rather than sticking rigidly to one stack. He builds custom WordPress themes and, with React, creates Gutenberg blocks that give clients flexible editing within design and brand guardrails. This can preserve a site’s intended design while still letting clients create and edit pages. He notes that some sites may not need a full CMS if they are rarely updated.

• AI’s wider effects remain uncertain. Faye sees parallels with earlier technologies that were promised as labor-saving but often made jobs more complex. He expects AI to become more integrated into everyday workflows, while cautioning against claims that programming is already solved or that developers are about to disappear. He also predicts AI could eventually encourage more people to become programmers, as users who want dependable results will have to learn the underlying skills.

Reviews, guides, and tutorials

This is a discussion rather than a hands-on product review or step-by-step tutorial. The host introduces Faye’s article “Agentic Coding Is a Trap,” which was discussed by Theo (T3G) and covered by Cal Newport. Faye says it reached the front page of Hacker News. The conversation examines the article’s arguments about dependence on coding agents, learning, and maintainability.

Main speakers and sources

• Mark Samsky — host of Fueled by Progress
• Lars Faye — self-taught developer and agency co-founder; guest and author of “Agentic Coding Is a Trap”
• Referenced commentators: Theo (T3G) and Cal Newport