Titus Winters says Google's Abseil refactoring of roughly 250m lines of C++ was probably the largest refactoring in human history, done with deterministic compiler-based tools.
Winters argues AI is best used to write deterministic rewrite recipes and lint rules rather than relying on constant inference, echoing Kelsey Hightower's zero token architecture idea.
Winters says most of the undergraduate CS curriculum has barely changed since about 2000 and should be partly burned down, with more design, testing and an AI-free first semester.
Lessons from Google's largest refactor for the AI era
Software engineering is the team discipline of changing code safely over time, and AI raises the stakes on that discipline. Winters draws on Google's testing culture, its record-scale Abseil refactor and its readability program to argue that tests, deterministic tooling and learning cultures make change sustainable. AI helps most when it writes those tools and least when it just generates more code.
Takeaways
A sustainable codebase makes valuable change safe. Winters argues fear is no basis for engineering. After Google search lost its regular release cadence around 2006 and 2007, release engineers refused to merge changes without tests. That built the safety net he found when he joined in 2010.
Google's Abseil cleanup proved deterministic refactoring scales. Google's C++ codebase was about 250m lines, edited by around 10,000 engineers a month. About 10 to 12 people ran compiler-based AST transformations as MapReduce jobs. Old and new APIs coexisted while edits landed directory by directory.
Use AI to build deterministic tools. Orosz says AI rewrites still feel janky next to Google's approach. Winters says AI is good at generating rewrite recipes for tools like OpenRewrite and BrontoSource, echoing Kelsey Hightower's zero token architecture. Both favour AI-written lint rules over constant inference.
Psychological safety predicts whether teams succeed. Winters calls software engineering a team sport that stays critical with AI. DORA research shows psychological culture scores explain 30 percent of the variance in hitting goals. He says layoffs carry a large hidden productivity cost and praises Adobe for avoiding sweeping cuts.
AI's 10x productivity claims look doubtful. Senior engineers report 70 percent more code and 30 percent more outages with the same features. Winters calls developer experience the business of productivity but hard to measure. A multi-indicator index misses the value of changes, so his develop deploy operate paper distinguishes four forms of impact.
Google's readability program was education at scale. It began with employee number three walking new hires through a first change and grew into a mentoring program with tens of thousands of participants. Winters says it was absolutely designed as education, and that culture fed the book Software Engineering at Google.
The CS curriculum has barely changed since 2000. Winters, the only working engineer on the ACM and IEEE curriculum steering committee, says it should be partly burned down. He sees the weakest students producing slop with AI and the middle losing resilience, and expects CS enrollment to fall as software engineering programs grow.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
How Winters would rebuild CS degrees
Spend less time on linked-list drill and more on design, specification, testing and validation.
Grade the process, including prompts and tests, rather than only the final code.
Keep the first semester free of AI so students learn to read code and control flow, and add a CS zero course.
Run projects with peer critique. Orosz adds that students should build software from day one, citing bootcamps and University of Waterloo co-op placements.