Recurse Center's return(): The quiet counterweight to AI-era coding

The most contrarian artifact in tech this summer is a 17-minute memoir about a 12-week programming retreat. RecurseCenter.return(), published in August by a UK-based self-taught coder, documents their experience at The Recurse Center (RC), the Brooklyn-based “programming retreat” founded in 2011. It is a warm, granular account of pair programming, Zulip DMs, and body-doubling sessions. It contains zero mentions of large language models, AI assistants, or copilots. That absence is the story.

The author, who writes anonymously and describes themselves as “a council estate chav from England,” applied after twice encountering RC links at the bottom of Hacker News articles by Nicole, a 3D-printing blogger. The application process involved a written application with prose questions, an informal Zoom chat with an RC alum, and a pair programming session where they rebuilt a Space Invaders game from an existing engine. They were accepted for a 12-week batch, fully remote from the UK, five hours ahead of New York.

The piece is structured as a daily routine: morning coding from 07:00 to 11:15, body doubling at 11:30, food and a nap until 15:00, then the “main RC day” from 16:00 to 22:00 New York time. The author voluntarily committed to being present until 10pm daily, a discipline they mostly kept. They avoided Zoom calls in the final hours of the day, worried that tiredness would flatten their demeanor. They describe RC culture as one where “cold DM-ing is fine,” where faculty curate a calendar but never enforce attendance, and where the only real rule is that you get out what you put in.

What makes this worth reading in September 2026 is the timing. We are now two-plus years into the mainstream AI coding assistant era. GitHub Copilot, Claude Code, Cursor, and their successors have reshaped how professional programmers write code. The industry discourse has shifted from “will AI replace programmers?” to “what does a senior engineer even do when the boilerplate writes itself?” Recurse Center, which has always been about humans learning from humans in a self-directed environment, looks almost like a museum piece. The author’s account suggests otherwise.

RC’s model is built on assumptions that AI tools increasingly challenge. The retreat assumes that struggling with a problem is pedagogically valuable, that pair programming teaches communication as much as syntax, and that a diverse batch of strangers can teach each other more than any single curriculum. The author notes that their batch seemed unusually weighted toward “visual or graphically-creative things,” and wonders whether RC intentionally mixes cohorts by interest. That kind of curation is a human judgment call no algorithm currently makes well.

The memoir also quietly documents the social infrastructure that makes deep work sustainable. The author, who admits to a “code hermit” tendency, describes body doubling as both beneficial and a useful “separator between my morning coding and my RC day.” They praise the explicit norm that direct messaging strangers on Zulip is acceptable, which removed the “fraught with ambiguity” anxiety of cold outreach. These are not technical features. They are cultural protocols, and they are exactly the kind of thing that AI-assisted development environments have not yet replicated.

There is a deeper point here for the AI industry. The current frontier of coding tools is moving toward agents that act autonomously: given a ticket, they explore a codebase, write a patch, run tests, and open a pull request. The human’s role is shifting toward specification, review, and judgment. Recurse Center’s model trains precisely those skills, but in a way that assumes the human does the coding. The author spent 12 weeks generalizing a game engine into Space Invaders, working through bugs with a partner, and exploring “outside my comfort zone” projects. None of that is obsolete. But the retreat’s value proposition may need updating.

Consider what the author gained from the application process alone, before acceptance: they clarified their “story so far” as a programmer, normalized video calls with strangers, and did their first pair programming session. “Already, I’d gained from this endeavour. Even if I didn’t get invited to a batch I was still up on the deal,” they write. That framing, where the process itself is the reward, is antithetical to the efficiency-obsessed logic of AI tooling. The entire AI coding pitch is about removing friction, shortening feedback loops, and compressing time-to-output. RC is deliberately frictionful. It asks you to slow down, talk to people, and sit with problems.

The author’s remote experience also highlights a structural tension in AI-era education. They committed to six-hour core days starting at 4pm UK time, a schedule that required a nap at 13:00 to sustain. They note that remote participants must “put effort into making myself visible” because in-person Recursers naturally bump into each other. This is the same problem distributed AI teams face: the async, screen-mediated collaboration that tools enable is thinner than physical presence. The author’s solution, showing up for the full RC day even when exhausted, is a human choice that no assistant can make for you.

There is also a quiet economic argument buried in the memoir. The author is self-taught, self-employed, and describes a moment of career doubt that prompted the application. They were uncertain whether they were “computer-sciency enough” as a self-taught coder. RC accepted them anyway, and the batch included people younger and older, formally educated and not, currently in education and currently educating others. That heterogeneity is a feature. In an AI economy where coding bootcamps are struggling and entry-level roles are shrinking, RC’s model of peer-supported, judgment-free skill development is a plausible alternative pathway. It is expensive in time and attention, but it produces something AI tools cannot: a network of humans who know how you think.

The memoir ends mid-stream, truncated in the single-page view, with a list of general activities that “change all the time.” The author never reaches a grand conclusion, and none is needed. The piece’s power is in its ordinariness: a person who loves coding spent three months getting better at it, surrounded by kind people, and found the experience nourishing.

For AI builders, the takeaway is uncomfortable. The tools we are shipping optimize for individual throughput. They make one programmer faster, but they do nothing to build the social fabric that the author describes: the Zulip threads, the body-doubling sessions, the shared calendar of exploratory talks. If the future of programming is fewer humans directing more agents, then the humans who remain will need exactly the skills RC cultivates: communication, judgment, curiosity, and the ability to collaborate with strangers. Recurse Center’s return() is a reminder that those skills are not learned from a model. They are learned from other people, slowly, in a room that may be physical or virtual but is always social. The author’s 10pm bedtime, five hours ahead of New York, was the price of that lesson.