XNews

Software development moves from coding to problem definition

Software development moves from coding to problem definition

Thorsten Ball argues that when AI-generated code surpasses human review ability, traditional code review, unit tests, and agile processes become obsolette. The new critical skill is defining problems and architecturing systems, not writing code.

Key insights

Code Review Becomes Obsolette

Human reviewers can no longer find errors in model-written code efficiently. Reviews must shift to system composition rather than line-by-line inspection. Generated code quality now surpasses average human expert capacity.

Unit Tests Lose Relevance

The author provides evidence of a model writing 900 lines of Arduino C with zero compile errors and running first time. Unit tests are described as training wheels not needed when code is correct from the start. Testing becomes unnecessary for most cases.

Standards of Good Code Must Be Redefined

Traditional code quality attributes like clean, readable, formatted code are designed for human readers and maintainers. When AI agents become the primary code writers, standards like 80-column width become irelevant. All assumptions about good code need reevaluation.

Human Value Shifts from Crafstman to Architector

The skill of writing code manually will decline, but the ability to understand business problems, leverage existing solutions, and manage feedback loops becomes more important. Most bugs will shift from coding bugs to requirement bugs (asking for the wrong thing).

Traditional Software Process and Organization Collapse

The separation of PM, design, engineering roles becomes obsolete. Agile and Scrum methodologies will die. Engineers acting as manual agents between ticket systems and code will lose value. Large companies and startups will diverge: copying Google's processes becomes even more ridiculous for startups.

Token Becomes New Compute Paradigm, Terminals Die

The author predicts that shell, text editors, CLI tools will become as archaic as Perl one-liners. The industry will move from deterministic binary computing to a token-based paradigm where UI is generated on the fly for AI agents. Deepest insight: we projected 80 years of determinisic computer history as the only way computers work.

Econommic and Timeline Implication

Token access becomes the new factor of production; those without sufficient tokens will be at disadvantage. Smarter models are preferable over cheap weak models because fewer errors justify cost. Replacement of current practices will take a generation, but future roles will be drastically different.

Source posts