Key Summary
- Among 14 chapters, only Chapter 14 lacks a "Key Takeaways" section
- Replacing judgment, not adding knowledge
- Quality standards and incident determination procedures
- One thing to check this week
AI Times' book review column 'AI Leader's Bookshelf' covered "Essential Practical Knowledge Every Junior AI Engineer Must Know" (by Kim Tae-heon, Hanbit Media), published in late July. The book consists of 5 parts and 14 chapters total. What the column noticed first in the table of contents was not 'what to learn' but 'what to unlearn'—the point being that the author emphasizes, as practical knowledge for new engineers, 'how to erase already-ingrained faulty judgment.'
The chapter titles the column cited are framed negatively. Chapter 3 is "A Data Pipeline Is Not Simple ETL," Chapter 6 is "Generative AI Is Not an API, It's a Probability Engine," and Chapter 12 is "Not Intuition, But Measurement: Evaluation-Driven Development." Chapter 1 is "Deterministic Systems vs. Probabilistic Systems," and Chapter 2 is "Why Are the Performance Metrics Normal While the Service Is Broken?" There's also one exception in the structure. Of the 14 chapters, 13 end with a "Key Takeaways" section, but Chapter 14, titled "What Must AI Engineers Judge Differently?", has no such section. The column read this as reflecting that the process of doubting and judging cannot be simply summarized.
Implications for AX at Korean Companies
Translating this table of contents into a workforce perspective makes it read more like a checklist than a training curriculum. If designing in-house AI training, it makes sense to place 'what assumptions to remove' alongside 'what more to teach.' If quality standards, incident determination, and release approval procedures built on the assumption of deterministic software are left unchanged while generative AI features are simply layered on top, the question posed in Chapter 2—'performance metrics are normal, yet the service is broken'—becomes a real question on the operational front line. Chapter 12's title, 'Not Intuition, But Measurement,' reads as a call to establish new evaluation standards at exactly that point.
The column recommended not reading the book cover to cover from the first chapter, but instead first searching the table of contents for traces of trial-and-error or faulty judgment experienced in the service one currently manages. Translated to the organizational level, there is one thing to try this week. Pick one recent incident or quality issue in an AI feature and check whether it was caught by a metric or first surfaced through a user complaint. If it was the latter, that already reveals where measurement standards need to be rebuilt from.
Source: [AI Leader's Bookshelf] Essential Practical Knowledge Every Junior AI Engineer Must Know