Thorsten Ball's 16 Predictions: Code Review, Terminal, Unit Tests Face Extinction in AI Era

Introduction

Thorsten Ball, author of Writing An Interpreter In Go and engineer at Sourcegraph/Amp, posted 16 predictions about the future of software development on X (September 19, 2026). Drawing from daily experience using AI to write hundreds of lines of code — including a 900-line Arduino C program that compiled and ran flawlessly on first try — he argues that many established engineering practices are becoming obsolete.

First Shockwave: Death of Code Review, Unit Tests, and Terminal

Code Review Is Dead

Ball states code review is already dead: humans will no longer find bugs in model-generated code within reasonable time. Review shifts from line-by-line PR inspection to evaluating system architecture and composition.

Unit Tests May Die

He cites the Arduino example: 900 lines of C code compiled with zero errors and ran perfectly on hardware. His reasoning: if models rarely produce broken code, the safety net of unit tests becomes unnecessary — like wearing knee pads when you don't fall.

The Craft of Writing Code Will Disappear

Ball compares it to Italian handmade shoemakers: they still exist but are not mainstream. The craft of writing code will become a niche artisanal activity.

Unified logic: When model output reliability reaches a threshold, the entire human quality-gate apparatus designed to catch human errors loses its necessity.

Value Migration: From "Writing Code" to "Building Software"

If coding craft devalues, where is engineering value? Ball answers: the craft of building software becomes more critical than ever. This means knowing how to solve business problems with software, understanding existing systems and their rationale, deciding when to ship, and obtaining real feedback.

Corollary: Most bugs will no longer be code bugs but "you specified the wrong requirement" bugs. Once code correctness is largely handled by models, the hard problems shift upstream to defining what to build.

This explains current trends: emphasis on clear requirements over code style, and prompts/specs emerging as a new engineering language.

Edge case: Performance optimization remains an edge-case human contribution. Ball notes 99% of software doesn't need extreme optimization; when it does, models can handle it. Don't compare the top 1% of developers to the other 99%.

Cognitive Upgrade: Token as the New Compute Paradigm

Ball declares: Token is the new compute paradigm; everything will be rebuilt on top of it. After 80 years of deterministic computing, we confuse "how computers have worked" with "how computers must work." We are entering a "post-binary era."

Terminal Is Dead

Despite being a terminal enthusiast, Ball predicts shells, text editors, and CLI tools will be washed away by tokens. Humans will no longer use them directly; command-line flags and jq syntax will become archaic crafts like Perl one-liners today.

Combined implication: When communication with computers shifts from deterministic instructions to natural-language tokens, the entire toolchain built on the deterministic paradigm must be re-evaluated.

Organizational Earthquake: Triad Dissolution and Big-vs-Small Divergence

PM-Designer-Engineer Triad Collapses

Traditional role division loses meaning; Agile and Scrum will "die." Engineers who merely act as human proxies — feeding requirements to AI agents and reporting results — add no value.

Gap Between Large and Small Companies Widens

Startups copying Google's engineering practices will look more absurd than before, because small teams can now build software faster with fewer constraints.

The article includes a derived diagram (not from Ball) suggesting organizations will restructure around who holds judgment over the problem, not functional silos.

Most Disruptive Claim: "Good Code" Definition May Be Collapsing

Ball asserts: There is no evidence that "good code" will remain important in the future. The concept of good code rests on the premise that code must be maintainable by humans. If humans rarely modify most code, that premise fails. He mocks the idea that line-length limits or brace placement matter to agents — if such details lose meaning, all attributes of "good code" deserve re-examination.

Related Judgments on Resources and Open Source

Open source in its current form becomes less rational. "Given enough eyeballs, all bugs are shallow" still holds, but we now have "artificial eyeballs."

Some people will be excluded from software production. Just as not everyone could afford a PC in the 80s/90s, in coming years those without sufficient token access will play in a "minor league." You need tokens first.

Ball further questions whether cheaper models will see wide adoption: smarter models make fewer errors and require fewer interaction rounds; would you accept more errors to save money?

UI Will Be Generated in Real Time

When models are fast enough, UI will be generated on the fly. Menus, settings pages, dashboards, and filters exist because software couldn't understand user intent; with sufficient intelligence, required UI shrinks dramatically.

Reality Check: A Generation-Long Transition

Ball explicitly states: This transition takes a generation. Just as ASP developers still exist today, people will still earn a living writing code in 10 years — but he asks: "Do you really want that job?"

Open Questions for Further Reflection

Responsibility allocation between "code review is dead" and "reviewing systems": If humans no longer review code line by line, who bears responsibility when systems fail — the specifier, the system reviewer, or the model?

Does the collapse of "good code" standards underestimate long-term evolution costs? Even if humans don't modify code short-term, long-term iteration, debugging, and security audits may still require readable, structured code — an open question needing more time.

Will token pricing as the gatekeeper of software production worsen the existing digital divide? If the barrier shifts from "can you code?" to "can you afford tokens?", is that liberation or a new wall for individual developers and small teams?

Conclusion

Ball's 16 predictions describe a visceral "felt sense" from watching models produce flawless 900-line Arduino code. Whether the changes unfold exactly as predicted remains to be seen. However, the discussion has moved beyond the binary "will AI replace programmers?" to a deeper question: When the cost of writing code approaches zero, what remains irreplaceable in the discipline of software engineering?

Reference: https://x.com/thorstenball/status/2101305394190557466