2026W23
- Software Is Made Between Commits — The claim that stuck with me: “the conversation that generates the code is becoming the true source of our software.” I’ve felt this shift myself. The PR description is always a lossy reconstruction of what actually happened. DeltaDB anchors references to deltas instead of line numbers, so they survive as code moves underneath. That’s the kind of detail that separates a real rethinking of version control from just adding chat to Git.
- If Claude Fable stops helping you, you’ll never know — The real problem isn’t the restriction itself, it’s that it’s invisible. When a tool silently degrades instead of refusing, you can’t tell the difference between “this problem is hard” and “the model is deliberately underperforming.” The author makes a good point that fine-tuning CLIP models for a travel app would have been frontier research five years ago. The boundary between normal product work and “frontier AI development” keeps shifting, and Anthropic gets to draw the line.
- Replace your CI with a merge queue — “The only thing worse than cleaning up after someone else is cleaning up after someone else’s robot.” That line captures what’s broken about agents and CI right now. By the time CI fails, the agent’s context window is gone, so nobody is left to fix the mess. Running tests before merge instead of after is such an obvious fix that it makes you wonder why we ever settled for the alternative, even for human workflows.
- Changing How We Develop Ladybird — The key insight here is that effort is no longer a proxy for trust. A substantial patch used to signal that someone invested real time and understood the codebase. With AI, that signal is gone. For a browser running untrusted input from the entire internet, the stakes are too high to rely on code review alone. I’m curious how this plays out. Closing PRs is a clear answer, but it also means losing the serendipity of outside contributors who spot things maintainers don’t.
- the mathematics of multi-tenancy — The “heat ratio” framing (peak divided by average workload) gave me a much sharper way to think about when multi-tenancy actually works. Two things kill it: correlated workloads and size skew. Just 25% correlation in workload patterns is enough to wipe out the cost advantage, even with hundreds of tenants. The detail about S3 splitting at the file fragment level, not even per object, explains a lot about why it works so well as everyone’s default storage layer.