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 participant’s September 11 account describes a summer at Recurse Center, a programming retreat in Brooklyn, spent on collaborative study groups and independent projects. The report covers hands-on work with language models, mathematics, a board-game AI, a small programming language and a DEFLATE decompressor; it does not describe a formal research result or institutional announcement.
A programmer who spent the summer at Recurse Center, a programming retreat in Brooklyn, published an account on September 11 describing a mix of peer-led study groups and personal software projects. The report offers a firsthand look at how participants explored topics from AI agents and machine learning to mathematics, programming-language design and file compression; it does not announce a new product or research finding.
The author says a friend, Cory, recommended the retreat. During the summer, the author joined study groups in Agentic Adventures, Practical Deep Learning and Math Monday, and took part in a short investigation of an open-source AI for the board game Race for the Galaxy. Recurse Center participants can organize and schedule study groups, according to the account. One example was a Friday group working through older Advent of Code programming puzzles.
Agentic Adventures focused on modern language models and agents. Activities included building a remote sandbox for an agent, setting up a local agent environment with Ollama and Docker, and using a small GPT-style model to predict text in the style of Shakespeare’s Romeo and Juliet. The group also tried a basic low-rank adaptation, or LoRA, to make a small model more conversational. These are descriptions of group exercises, not claims of new model capabilities.
Other projects ranged from drawing fractals and generating Hilbert curves for a pen plotter to completing the author’s small language, dodo. The author also paired with another participant to write a Rust decompressor for the DEFLATE format used in ZIP files, working from its specification. The source account ends partway through its discussion of debugging bit sequences, so it does not establish whether the implementation was completed or how it was evaluated.
Peer Projects Across Computing
The account shows how a participant-led programming retreat can bring structured learning and practical experimentation together. Its examples range from introductory work through a deep-learning textbook to implementation challenges involving compression and recursive language features. That breadth may interest programmers weighing collaborative study against solo project work.
The report also captures how participants engaged with AI tools in different roles. They built agent sandboxes and experimented with small models, while the author says they did not let language models write the dodo code. Instead, the models helped draft a specification and test cases and examples. The distinction is useful: the account describes AI as both a subject for experimentation and a supporting tool, without claiming that it replaced the programming work.
How the Summer Was Organized
The author describes Recurse Center as a programming retreat in Brooklyn and says the groups were open for participants to organize and schedule. The account does not give the summer’s dates, the number of participants, or a formal curriculum. It instead documents the author’s own activities and collaborations.
In Practical Deep Learning, the group worked through the first half of the book by Jeremy Howard and Sylvain Gugger. The author says the study moved from classical machine-learning approaches to prebuilt models and then neural networks, and valued the book as a practitioner-level introduction. In the board-game AI sessions, participants first learned the rules and played the game before examining the Keldon AI’s older two-layer neural network and curated features. The author reports that their interpretation suggested economic strategies were stronger than military ones in the base game, a conclusion said to align with player sentiment.
““I spent the summer at Recurse Center, a programming retreat in Brooklyn.””
— The report’s author
Projects Without Published Evaluations
The account is a personal report and does not provide formal evaluations or complete project records. It does not say whether the remote agent sandbox or the DEFLATE decompressor was released, independently tested, or used beyond the sessions described. The source text cuts off while discussing debugging, leaving the decompressor’s final status unclear.
The author’s reading of the Race for the Galaxy AI is also an interpretation of its network weights, not a controlled comparison of strategies. The account provides no performance data, test conditions or details about how the game positions were sampled. Likewise, the small-model exercises are described without benchmarks or measurements of their output quality.
No Further Milestones Listed
The report does not announce a follow-up event, release date or next milestone for the projects. The account’s clearest next step is left unstated: readers are directed to an online REPL for trying the dodo language, but the source does not supply a continuing project schedule or evaluation plan.
Further details about the decompressor, agent sandbox and game-AI analysis would be needed to establish whether those experiments developed into maintained tools or broader findings. For now, the article records what the author and collaborators worked on during the summer.
Key Questions
What is Recurse Center?
The author describes it as a programming retreat in Brooklyn where participants can organize and schedule study groups. The account does not provide additional institutional details.
What did the author study there?
The author joined groups on AI agents, practical deep learning and mathematics, as well as a short series examining an open-source Race for the Galaxy AI. The author also worked on personal programming projects.
Did the group make a new AI model?
The report describes exercises with minGPT and a basic LoRA on a small model. It gives no benchmark results or evidence that the work produced a new general-purpose model.
Was the DEFLATE decompressor completed?
The author says they and a collaborator wrote a Rust decompressor from the format specification. The available account cuts off during its debugging discussion and does not confirm completion or testing.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
