What I did at Recurse Center

Spent a summer at a programming retreat where the work centered on hands-on study groups, pairing, and building projects around modern large language models, tooling, and classic systems. Participated in Agentic Adventures (sandboxed local agents, a cooperative “Just One” agent game, training tiny models with minGPT and LoRA), a Practical Deep Learning study group that covered practitioner-focused ML and neural nets, and a Math Monday group that tackled Project Euler, fractals, Voronoi art, Rocq proofs, and Hilbert curves. Reverse-engineered the Keldon Race for the Galaxy AI - an old two-layer neural net with curated features - and found economic strategies dominate in the base game and many strong-weighted nodes correspond to individual cards. Wrapped up a mini language called dodo (including a tricky recursive match implementation) and implemented a DEFLATE decompressor in Rust (LZ77 backreferences plus Huffman coding), learning low-level bit packing, debugging, and enjoying Rust despite lifetime complexity.

Pairing underpinned much of the work: using SAT solvers for Sudoku, building Game of Life and Mastermind implementations, and attempting an ambitious Magit view. Vibecoded many LLM-driven minis: a rhythm game built from multiple prototypes, a fantasy map generator with climate modeling, and agent-driven games. Agent experiments showed a tendency to give identical hints unless given distinct “personalities,” while agent bargaining produced brittle, combative behavior. Built deployable tooling so others could use their own keys or GPUs, an Emacs session view, and explored cross-lingual phonetic searches and a constructed language using PHOIBLE and APiCS, which led naturally to an analytic, creole-like grammar. Overall, the summer emphasized practical learning, rapid prototyping with LLMs, and low-level systems hacking.