A software engineer is spending three months at the Recurse Center in Brooklyn coding without AI assistance, seeking to deepen his understanding of computer science fundamentals. The developer, who previously built AI agents at Aily Labs in Barcelona, noticed that while coding agents enabled rapid iteration, they often prevented him from truly learning codebases and core concepts. He observed that top programmers who effectively use AI tools typically possess deeper foundational knowledge that gives them more leverage.

His goals include training an LLM from scratch, improving Python proficiency without relying on documentation or AI assistance, and building a better mental model of computer systems. He's already completed Stanford's CS336 Language Modeling from Scratch assignment, writing a tokenizer and GPT-2 style architecture in PyTorch, training on datasets like Tiny Stories and OpenWebText. He's also working through computer security challenges in CTF Fridays and experimenting with vintage computing, including writing fizzbuzz in BASIC on an Apple IIe from 1983.

Through pair programming with experienced developers and hands-on projects, he's developing techniques like rapid terminal testing to work through syntax questions independently. The retreat represents a deliberate step back from AI-assisted development to strengthen fundamental programming skills that he believes will ultimately make him more effective when using AI tools.