Simon Willison

Highlight

Simon’s standout post details a hands-on generative AI experiment chaining ChatGPT Images 2.5 with Codex (GPT-6 Astra) to build a 3D Blender model and launch an in-browser .blend URL Viewer tool. It showcases how multimodal AI pipelines and custom local coding agent skills are making 3D asset creation and browser utilities frictionless.

Posts

.blend URL Viewer Simon details a multi-step generative workflow where he used ChatGPT Images 2.5 to generate an image of a Fabergé egg themed after the TV show Pluribus, then passed that image directly into Codex running GPT-6 Astra (high) with a custom blender local skill. Over 17 minutes and 51 seconds, the agent autonomously constructed several .blend 3D model files based on the generated image. To easily share and inspect the output, Simon added a web viewer to his personal tools collection that renders .blend 3D files directly in the browser via a URL parameter. This post highlights the practical power of combining image generators, CLI coding agents with local skills, and small browser tools to accelerate 3D design workflows.

Quoting Terence Tao Simon highlights a commentary from mathematician Terence Tao warning that AI-powered research efforts are rapidly mining open, fruitful mathematical problems as soon as rumors of work surface. Tao points out that this aggressive automated effort risks depleting accessible research problems and could incentivize mathematicians to stop sharing promising research directions publicly. The key takeaway underlines the profound impact AI automation is having on the culture of open science and academic research incentives.

Project Pulse

Today’s posts capture Simon’s dual focus on practical, vibe-coded toolmaking with multimodal coding agents alongside tracking AI’s disruptive ethical and cultural impacts on open scientific research.

🛠️ Would you like to explore setting up custom CLI coding agent skills or local workflow tools inspired by Simon’s Blender agent experiment?