TL;DR
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A September 2026 blog post recounts a summer spent on collaborative programming projects at Recurse Center in Brooklyn. The writer describes study groups in machine learning and mathematics, experiments with language models, work on a small programming language, and a Rust implementation of a DEFLATE decompressor.
A programmer who spent the summer at Recurse Center, a programming retreat in Brooklyn, has published an account of collaborative study groups and software projects, including small language-model experiments and a DEFLATE decompressor written in Rust. The September 11, 2026 post offers a first-person record of what the writer worked on; it does not announce a new product or research finding.
The writer joined study groups in agentic systems, practical deep learning and mathematics, and took part in a short effort to examine an open-source board-game AI. The groups combined reading and discussion with practical exercises. In Agentic Adventures, for example, participants built a remote sandbox for an agent and explored a separate sandboxed setup for local agents using Ollama and Docker.
Other activities included using minGPT to train a small text predictor on Romeo and Juliet, and applying a basic LoRA technique to make a small model speak more conversationally. The writer also finished dodo, a mini programming language begun as part of an application to Recurse Center. They say they wrote its code themselves while using language models to help prepare specifications, tests and example programs.
In a series of paired sessions, the writer and another participant implemented a DEFLATE decompressor in Rust, working from the format specification. The post describes DEFLATE as combining LZ77-style backreferences with Huffman coding. The project followed a talk about ZIP file behavior and the security risks that can arise when software handles archive contents inconsistently. The account describes the implementation as a learning exercise; it gives no performance results or independent evaluation.
What the Projects Showed
The account illustrates how a retreat organized around peer-led study can give programmers time to move between theory and hands-on work. The writer describes moving from introductory machine-learning material to model experiments, and from a talk about archive formats to implementing part of a compression standard. These are examples from one participant’s experience, rather than evidence about outcomes for the program as a whole.
The projects also show different ways participants used AI tools. Some sessions explored agents and language models directly, while the writer says they kept model-generated code out of their own implementation of dodo. In that project, they used language models for specifications, test cases and examples instead. That distinction is relevant to readers trying to understand how one programmer incorporated AI assistance while retaining responsibility for the code.
The DEFLATE work connects a relatively small programming exercise to a broader software-security concern raised in the talk: if tools that inspect and extract an archive interpret it differently, they may not agree about its contents. The post does not report finding a new vulnerability; it describes the issue as motivation for learning how the format works.
From Study Groups to Code
The writer says a friend, Cory, recommended spending the summer at Recurse Center. The post describes the center as a programming retreat in Brooklyn where participants can organize regular study groups. One example was a Friday gathering to work through older Advent of Code challenges. The writer joined Agentic Adventures, Practical Deep Learning and Math Monday, among other activities.
Practical Deep Learning used the first half of a book by Jeremy Howard and Sylvain Gugger. According to the writer, the group progressed from classical machine-learning approaches and prebuilt models toward building neural networks. Math Monday, started with Sophia after a coffee chat, paired learning with making things: participants worked on Project Euler problems, mathematical puzzles, fractals, Voronoi diagrams, Boolean identities in the Rocq proof assistant and Hilbert curves for a pen plotter.
A separate short sequence examined the Keldon AI for the board game Race for the Galaxy. The group first learned the rules and played against the program, then inspected its code, which the writer describes as a two-layer neural network with curated features. The post says the group used Claude to help interpret network weights and discuss the strategies the AI appeared to favor. The writer’s takeaway was that economic strategies seemed stronger than military ones in the base game, an observation they said aligned with player sentiment.
““I didn’t let them write any of the code for me.””
— The post’s author
What the Post Does Not Establish
The account is a personal retrospective, and it does not provide independent checks of the projects’ results. It gives no benchmark or comparison for the language-model experiments, no full evaluation of the board-game AI, and no performance or compatibility results for the decompressor. It also does not specify the exact dates or duration of each project, or report outcomes for other Recurse Center participants.
Projects After the Retreat
The post does not announce a next project, publication or follow-up milestone. It says readers can try dodo through an online REPL, but provides no further development schedule. Any later work on the language, decompressor or other projects remains unreported in this account.
Key Questions
What is the news development?
A programmer published a first-person recap of projects and study groups from a summer at Recurse Center in Brooklyn. The post appeared on September 11, 2026.
What did the writer build?
The writer finished a mini programming language called dodo and worked with a partner on a Rust DEFLATE decompressor. The post also describes group experiments with small language models and a remote sandbox for an agent.
Did the writer use AI to write the dodo code?
The writer says they did not let language models write the code. They used them to help draft a specification and create test cases and example programs.
Did the DEFLATE project reveal a new security flaw?
The post does not report discovering a new flaw. It says a talk about ZIP-file inconsistencies raised a security concern and prompted the writer and a partner to implement a decompressor as an exercise.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
