Vibeleaderboard
← All Intel
Intel / video

Rails World 2026 Opening Keynote - DHH

Source
youtube.com
Author
Ruby on Rails
Date

Context

The title promises the opening keynote at the Rails conference, but a large part of the talk explains why DHH's own company has stopped writing code by hand and is moving one product off its Rails web app. He says 37signals decided a couple of weeks earlier to go "pencils down" on writing code by hand: a hand-written change is now treated like an error report, a sign the setup needs fixing. When he asked the audience at Rails World 2026 how many still wrote meaningful amounts of code by hand each week, he counted about five hands.

The concrete consequence is Hey, 37signals' email service. DHH says Hey was only ever a web app because maintaining native apps on six platforms used to need a big team, and he now treats that cost as close to zero. He showed early builds of the six native apps, which he says were started about a week before the talk, and the backend is to become a mail server written in Rust. He calls Rust the ugliest language in decades for humans to read, but fine if only agents write it and he treats the result as a black box. He says the resulting backend needs 99 percent less CPU and 95 percent less memory, figures he states without showing how they were measured. He also says he wrote about 150,000 lines in August against a former average of roughly 30,000 lines of Ruby a year, and admits much of that is verbose Rust.

So where does Rails fit? DHH's answer is web apps that should not ask people to install anything, such as Basecamp, where outside collaborators drop in briefly, though he says he does not yet know exactly where the line between native and web will fall. He argues that Rails' long-standing habit of convention over configuration, meaning sensible defaults so you write less setup code, now also means fewer for an agent to read and write. He also says its aim of letting one developer build a whole app fits a world where single developers go further with agents. Earlier in the talk he replays a 2005 clip of himself demoing Rails and saying "look at all the things I'm not doing", and says the same delight now applies to code he no longer types at all.

That leaves Rails a narrower role, and the talk offers little on what comes next for the framework itself. A Rails developer's response to the keynote makes exactly that point. The shift in how DHH works with agents, describing problems rather than implementation, is set out at more length in his podcast conversation with Lex Fridman, which appears to be the five-hour podcast he mentions at the start of the talk.

Key takeaways · AI-distilled
  • DHH recounts that when 37signals had designers vibe-code final Basecamp 5 features in the spring, each PR looked reasonable but 20 or 30 together left the architecture 'like a Swiss cheese'; he now calls retreating to manual review the wrong conclusion.
  • For the next version of Hey, 37signals is replacing the web app with six native apps and a Rust mail backend; DHH says its estimates show 99% less CPU and 95% less memory, with 10 hosts kept only for redundancy.
  • He argues abstractions lose value with agents: they existed partly to avoid repetition, but repetition and keeping copies in sync are now nearly free, and abstractions become choke points when many agents change one codebase.
  • His one concrete ask: apps should ship CLIs so users' own agents can operate them, instead of chatbots; he cites an agent using the Hey CLI to find a five-year-old email by concept that Elasticsearch could not.
  • Agent evals run by Evil Martians for the Rails Foundation saturated quickly, with agents reaching about 95% completion on the first version, so a harder had to be built.
Terms in this piece · Glossary
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
  • token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
  • embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.
  • benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
Read the source www.youtube.com
Recommended reads
Comments

Checking sign-in…

Loading comments…