I highlight .fr08, a legendary 64kb demoscene production by Farbrausch. It is a milestone in procedural generation, packing high-fidelity 3D graphics and audio into a tiny executable file.
I checked out Princeton's 3D search engine, which lets you find objects based on shape. It is an interesting experiment, though I am skeptical about the actual use cases for the models it retrieves.
I share Innovation team demos: gesture-controlled 3D molecules, videos with evidence-backed critiques, audio search beyond transcripts, cheap machine-fault detection, and governed AI agents.
I looked into speech editing tools and GPT-4o Mini's impressive performance. I also listed dozens of text-based formats LLMs can generate to create physical artifacts like 3D models or circuits, and explored self-hosted PaaS solutions like Piku.
I explored OpenAI's custom tool types for DSL generation, new Node.js 2025 features like single-executable bundling, and Markdown directive syntax. I also shared my Claude Code usage costs and thoughts on vibe coding.
I found an argument that visual AI is more powerful when it generates code: render, inspect, and revise editable SVGs, interfaces, animations, or 3D scenes instead of static pixels.