This is a batch report from the Recurse Center — a summer's worth of projects, study groups, and pairing sessions, written up as a personal retrospective. It's not a product, a paper, or a polished artifact. It's a working journal of someone who showed up to RC wanting to understand LLMs and left having reimplemented DEFLATE in Rust, reverse-engineered a board game neural network, written a mini programming language, vibecoded a rhythm game, and built a constructed language. The sheer surface area of the work is the point. The strongest section is the DEFLATE implementation. The author and Kevan worked from the spec in Rust — a first language for the author — and the description of bit-packing debugging ('we would frequently have to stop and write down the bit sequences we expected to compare them to what we got') is the kind of concrete detail that makes technical writing land. This is the moment in the piece where you believe the author actually learned something hard, not just assembled components. The LLM/agent work is more diffuse but honest about its limits. The 'Just One' cooperative word game revealed that agents at temperature 1 still converge on identical hints — a genuinely interesting observation about the narrowness of LLM sampling. The diplomacy game failed outright: telling agents to negotiate produced 'stiff and combative characters.' The author doesn't dress this up. They tried, it didn't work, they moved on. That candor is worth more than a polished demo. The study group descriptions — Agentic Adventures, Practical Deep Learning, Math Monday — read as community documentation more than personal essay. They're useful as a map of what RC study groups actually look like in practice, but they're the thinnest sections in terms of voice. You get names and topics but not much about what was hard or surprising. The Math Monday section is the exception: Voronoi diagram games, Hilbert curves on a pen plotter, the chaos game in Desmos — these are specific enough to spark curiosity. The mini-language dodo and its match-statement implementation get only a paragraph, which is a shame — recursive pattern matching is genuinely tricky, and the author's note about using LLMs for specs and test cases but not code is the most interesting methodological observation in the piece. It suggests a mode of LLM use that's collaborative without being generative, and it deserved more space. The constructed language and cross-language phonological overlap projects at the end feel rushed, partly because the article itself appears truncated. The Indonesian-Farsi phonological search is a wild idea — using PHOIBLE data to find cross-language homophones — and the admission that the result was 'a very mediocre poem' is perfect. But you want more: what did the search space look like? How bad is 'very mediocre'? As a genre piece — the RC batch retrospective — this is above average. It covers genuine technical range, admits failures, names collaborators generously, and doesn't oversell. It's not trying to be a tutorial or a thought-leadership post. It's someone showing their work, and the work is varied and real.
