Three Months Offline: A Programmer's Retreat From AI-Assisted Coding

A former Aily Labs engineer who spent two years building AI agents is deliberately stepping away from coding assistants for a 12-week stint at Brooklyn’s Recurse Center. The motivation isn’t Luddism but leverage: he observed that the strongest programmers at his previous job were also the strongest AI users because their deeper fundamentals gave them more control over the tool. Coding agents, he argues, produce exactly what you specify, which becomes a liability when you don’t fully understand the codebase or what you actually want.

His retreat agenda is concrete. He’s working through Stanford’s CS336 language modeling course without LLM assistance, having already written a tokenizer and GPT-2-style Transformer in PyTorch and run ablations on TinyStories and OpenWebText. He’s deliberately pair-programming with veteran Python developers, hand-coding exercises in Vim, and grinding through Bandit-style terminal CTFs to shore up Unix fundamentals he’d glossed over in day-to-day shipping work.

The piece echoes Cal Newport’s argument that the cognitive effort of writing is itself the craft, not friction to be automated away. For engineers weighing how much to outsource to Copilot and Cursor, it’s a pointed reminder that the skill ceiling on AI tooling is set by what the operator actually understands.