Miguel Conner spent two years building AI agents at Aily Labs in Barcelona, including an internal web search agent months before Anthropic published their widely-cited guide on the topic. His team adopted Cursor early, experimented with LLM-driven knowledge graphs, and ran a weekly journal club dissecting papers on Llama 3, DeepSeek R1, and Olmo 3. He knows these systems inside out.
That's exactly why he's now at the Recurse Center in Brooklyn, spending three months coding mostly without AI.
The core tension he identifies is real. When he coded by hand, he was doing two things: writing what he wanted and learning the codebase. Coding agents gave him exactly what he specified, but they skipped the learning part. If he didn't know precisely what he wanted, the agent would happily fill in assumptions. He could iterate faster and ship working software, but his grasp of the underlying code suffered. He quotes Cal Newport comparing mental strain to physical exercise: strain is not an annoyance to eliminate, it's the point.
His retreat goals are concrete. He's working through Stanford's CS336 course to train an LLM from scratch, writing his own Transformer implementation rather than forking existing code. He completed the first assignment, building an optimized tokenizer and a GPT-2 style architecture in PyTorch with another Recurser, tuning hyperparameters on Tiny Stories before scaling to OpenWebText's 9 billion tokens. He's also pair programming with veteran Python developers and tackling CTF challenges to strengthen his Unix skills.
The Hacker News crowd noticed one detail that undercut the premise: Conner turned to Claude for debugging help after 20 minutes. Some saw this as evidence that problem-solving stamina has already eroded among developers who use AI regularly. The pushback highlights a genuine tension. Conner isn't anti-AI. He's arguing that the best AI users he worked with at Aily were also the strongest programmers, because deeper knowledge gets more from the tool. The retreat is about rebuilding that depth, not rejecting the technology.