Recurse Center Hand-Coding Retreat Addresses AI-Induced Knowledge Gaps
Conner's six-week Recurse Center project manually training an LLM reveals reduced codebase intuition and creativity retention from AI coding reliance, synthesizing Newport, Anthropic, Meta Llama 3, and Karpathy materials.
Miguel Conner, previously at Aily Labs, is spending six weeks at the Recurse Center in Brooklyn training a Transformer-based LLM from scratch without AI coding tools while improving Python proficiency by hand (Miguel Conner, "Spending 3 months coding by hand," Substack, March 2026). Conner states that AI agents deliver specified output but limit codebase learning compared to manual coding, citing his internal web search agent built in early 2024, six months before Anthropic's "Building Effective AI Agents" (Anthropic, October 2024). Cal Newport's analogy equates mental strain in craft to physical exercise, directly referenced by Conner as applicable to coding fundamentals (Cal Newport, "The AI Revolution and the End of 'Busywork'", New Yorker, January 2025).
Coverage of AI coding tools such as Cursor and OpenAI's DeepResearch frequently emphasized iteration speed and tutoring benefits but omitted Conner's documented tradeoff in reduced conceptual retention and assumption handling by agents (Conner, 2026; Anthropic, 2024). Meta's Llama 3 technical report and DeepSeek R1 details, presented in Conner's former journal club, outline training steps he is now implementing manually rather than forking existing codebases (Meta AI, "Introducing Llama 3", 2024). Recurse Center's collaborative cohort model, active since 2010, supplies peer review absent in prompt-driven AI sessions.
Synthesis with Karpathy's 2023-2024 lectures on LLM internals shows manual pre-training and post-training replicates core concepts now automated, addressing the "growing list of coding concepts" Conner deferred during production work at Aily Labs (Andrej Karpathy, "Let's build GPT", YouTube, 2023; Conner, 2026). Original sources on AI agents often overlooked this deliberate return to fundamentals as a leverage multiplier rather than obsolescence countermeasure.
AXIOM: Hand-coded fundamentals will remain the differentiator enabling effective direction of AI coding agents rather than replacement by them.
Sources (3)
- [1]Spending 3 months coding by hand(https://miguelconner.substack.com/p/im-coding-by-hand)
- [2]Building Effective AI Agents(https://www.anthropic.com/research/building-effective-agents)
- [3]Introducing Meta Llama 3(https://ai.meta.com/blog/meta-llama-3/)