Canonization vs. Tech Debt: Why AI Coding Eats Its Own Seed Corn

Doctorow distinguishes between 'today's task' software, which is personal and disposable, and 'accretive work' that builds long-term infrastructure through canonization. He argues that while AI can assist with the former, the industry's focus on replacing skilled labor creates 'reverse centaurs' and massive technical debt. Ultimately, the AI industry is consuming the social production of knowledge without contributing maintainable work back to the canon.

Key Points

• The difference between AI success and failure depends on whether the human is a 'centaur' directing the tool or a 'reverse centaur' serving it.
• Vibe coding is a valuable form of vernacular programming for personal use but is dangerous when applied to production-grade software.
• Canonization is the essential process of making code legible and maintainable for future human and non-human systems.
• The AI industry's business model necessitates firing workers and replacing them with 'homework markers,' which destroys the social production of knowledge.
• AI-generated code often lacks the 'accretive' quality needed to build long-term infrastructure, leading to unsustainable technical debt.

Sentiment

The overall sentiment is cautiously supportive of the article's core distinction. The community generally agrees that disposable code and durable software need different standards, and many commenters accept that AI raises real risks when generated work becomes shared infrastructure without skilled stewardship. The mood is not anti-AI across the board: commenters repeatedly defend pragmatic personal automation and skilled AI use, while reserving their strongest concern for management incentives, misplaced permanence, and loss of ownership.

In Agreement

• Single-use scripts, spreadsheet-like automations, and personal tools can be worthwhile even when they are not good software by professional engineering standards.
• The main risk is that temporary code often becomes load-bearing infrastructure before anyone deliberately hardens, deletes, or replaces it.
• AI-assisted coding is most defensible when skilled users remain in control and understand the systems they are producing.
• The centaur versus reverse-centaur framing usefully highlights who controls the tool and who bears the cost of bad automated output.
• Canonization requires social and organizational commitment: someone must be responsible for tuning rough fragments into a coherent, maintainable system.

Opposed

• Some commenters argue that AI is simply a tool, and that the article overcomplicates the issue by treating it as categorically different from other productivity tools.
• A few readers are tired of AI-centered essays and see the framing as distracted from otherwise useful ideas about software lifecycle and technical debt.
• Some skepticism focuses on the repeated use of terms like reverse centaur, which a few readers view as forced or self-promotional rather than clarifying.
• One counterpoint is that rough, accreted systems are not always disposable failures; with an intentional tuning phase, they can become highly successful products.
• Some commenters resist strong claims about AI dependency, saying experienced users can use the tools heavily without surrendering responsibility or judgment.