Titus Winters: Burn Down Most U.S. CS Curriculum — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Titus Winters, the builder of Google’s open-source foundational C++ library Abseil, described his team’s 2010 refactoring of a nearly 250-million-line C++ codebase with 12 people; the code had been edited by 10,000 engineers.
- Google’s 2010 C++ refactor parsed every build target into an abstract syntax tree using MapReduce, matched exact call patterns, generated edits, ran old and new APIs side by side, sharded changes by directory, and then blocked and deleted the old API.
- Google mandated a test for every change around 2007 after Search’s release cadence fell from monthly to quarterly; over time, Search returned to its previous higher cadence.
- Google’s readability program required an owner with domain knowledge plus a certified readability reviewer for every code change; more than 10,000 engineers were enrolled at a time when Winters left.
- Titus Winters says engineers at other companies report code output up sharply and outages about 30% higher than a year ago, with feature shipments static, while he sees AI’s larger payoff in ambitious projects that previously made no sense.
- Titus Winters, the only working engineer on the 2023 ACM/IEEE/AAAI curriculum steering committee, says most of the U.S. undergraduate CS curriculum should be burned down because it has barely changed since around 2000 and omits testing and verification.
- Titus Winters says professors report that top students are better than ever, students below a B no longer understand code, and the weakest can catch up largely by producing AI “slop”; he proposes grading prompts and tests rather than only final code.
Why it matters: For students, professors and new hires, the practical stakes are assessment and training: Winters proposes grading prompts and tests instead of only final code, and his 2023 curriculum work identifies missing testing and verification. His separate observation—about 30% more outages despite static feature output—also makes raw code volume an incomplete measure of engineering productivity.
Ask SkimNews
