David Heinemeier Hansson opened Rails World 2026 by announcing that he has retired from being a professional programmer and that the next version of Hey is leaving the Rails stack — Rust on the server, native apps on every client. Jared Norman builds Rails applications for clients, so he sat through the keynote waiting for the plan. What he got instead was reassurance.

The Rails content in a keynote about Rails came down to a few lines:

  • Rails is still a fine fit for “web apps of necessity” — reframed from a preference into a workaround
  • Convention over configuration now sells itself as token efficiency
  • Only about 3% of his output this year was Ruby, down from roughly half over two decades
  • The one message aimed at Rails developers was that they are “the best of the best” — a line that would work word for word in a room of Django or Laravel developers

That last point is the real complaint. A framework’s steward moving a flagship product off the stack is a governance signal, and the community can only act on it if someone says what it means. Rails in maintenance mode, explicitly declared, is a legitimate answer. A new direction is a legitimate answer. Silence delivered as confidence is not.

The productivity numbers get the same treatment:

  • 150,000 lines written in August is compared against a 30,000-per-year pre-LLM average — after conceding that lines of code is a poor metric and that Rust he would tolerate from a model is Rust he would never accept from himself
  • Hey’s “99% less CPU, 95% less memory” compares a Rust backend with no web frontend against the existing full Rails application, so the share attributable to Rust versus dropping the web app is unknowable
  • The 10x-to-1000x programmer lore traces back to a study of developer tooling, and the “average of 10x” often attributed to it is not in the paper
  • A Basecamp 5 architecture described as “Swiss cheese” is blamed on the models, but unreviewed and uncoordinated contributions degrade architecture whether they arrive from agents or from people
  • The ATM analogy used to argue the transition is safe gets the decade, the economist, the teller counts and the conclusion wrong

None of this is an argument against writing software with AI assistance. It is an argument about what a keynote owes an audience: checkable claims that were actually checked, and a direction that the people maintaining the framework can plan against.

The transferable lesson is the pairing. “Never look at the code” and “security, something’s coming, get ready” sat fifteen minutes apart with nothing bridging them. When a claim about engineering productivity arrives with confidence but without a comparable baseline, that gap is the finding — and it is worth reading the study before quoting the number.