What I did at Recurse Center

The ability to build and share LLM-powered web applications independently is now accessible, removing previous barriers to experimentation.

Patrick spent the summer at Recurse Center, a programming retreat, engaging in various study groups and personal projects. He participated in Agentic Adventures, exploring modern LLMs and agents, building sandboxes for AI agents, and training small models with LoRA. In Practical Deep Learning, he worked through a book on machine learning fundamentals and neural networks. Math Monday involved problem-solving sessions on topics like fractals and Voronoi diagrams, with a focus on generative art and mathematical puzzles. He also delved into reverse-engineering a board game AI, understanding its strategy through neural network analysis. Patrick completed his mini-language, dodo, and implemented the DEFLATE compression algorithm in Rust with a partner. A significant portion of his time was dedicated to "vibecoding" with LLMs, creating projects like a rhythm game, a fantasy map generator, and experimenting with agents in games like "Just One" and a diplomacy game. He also developed tools for sharing these browser-based projects, such as a component for using personal access keys or WebLLM for GPU access. Towards the end of his summer, he explored cross-linguistic word similarities through tree search and began developing a new constructed language.

The summer at Recurse Center highlighted a broad exploration of programming concepts, from classical machine learning to cutting-edge LLM applications. The participant's engagement with study groups and personal projects demonstrates a drive to understand and apply new technologies, with a particular focus on agentic behavior and practical deep learning implementations. The successful development and sharing of LLM-driven web applications, including components for user-provided API keys and WebLLM integration, indicate a maturing ecosystem for democratized AI development.

This summer's activities showcase a trend toward leveraging LLMs for both creative and analytical tasks, as seen in the fantasy map generator and the analysis of board game AI. The challenges encountered, such as the difficulty in coordinating agent behavior in games and the complexities of cross-linguistic phonetic analysis, reveal current limitations and active areas of research. The successful reimplementation of DEFLATE in Rust also points to a growing interest in understanding fundamental algorithms and building foundational software tools.