How do you understand code you didn't write?
GitHub Universe is coming up next week and will show new AI features, talk big numbers, and wow us with what’s coming.
Nobody will solve this: How do you understand code didn’t write?
I’ve been using Gen AI tools for 18 months. Claude, Cursor, GitHub Copilot—all excellent at generation.
But here’s what the keynote may miss — code generation speed ≠ development velocity.
The bottleneck isn’t how fast AI writes code.
It’s how fast YOU can understand, verify, and trust it.
Real numbers from recent studies:
- MIT/METR: AI makes experienced developers 19% SLOWER
- Stack Overflow: Developer trust in AI code dropped to 29%
- GitHub Research: 66% cite “almost right but not quite” as biggest frustration
Why? Verification overhead.
Your AI generates 200 lines in 30 seconds. You spend 20 minutes understanding if it’s actually what you wanted. Then 15+ messages asking for clarification.
The homework problem:
- Remember when someone gave you the answer but wouldn’t show their work? That’s AI-generated code.
- Your agent produces an auth endpoint, explains it in chat, but you’re hunting through files trying to verify the implementation matches the description.
Code creation is 5% of our time.
What’s actually slow:
- Understanding how things are built
- Planning alignment
- Root cause debugging
- Fire drills & co-ordination overhead
- All that stuff you learn in retrospectives …
If you optimize only for generation speed, you’re optimizing 5% of the problem.
The missing piece: AI tools excel at “write this code fast.” They fail at “help me understand this code.
We need tools that optimize for comprehension, not just generation. Interactive walkthroughs instead of chat reports.
Exact line references instead of descriptions. Mental model synchronization instead of context-switching.
That’s the real opportunity.
What’s your experience? Is your team faster with coding tools alone or are you finding it requires tooling across the SDLC?