I built an agentic system that taught itself the Blender donut tutorial by watching it on YouTube.
@cerspense6596 位关注者
255,153 次观看
49 条评论
评论来自 X (Twitter) 上的原帖

@LinusEkenstam
A few questions: 1. How much did this copy/paste cost? 2. If you ask it to make a donut again, will it make one instantly? 3. If we can teach agents skills by having them watch tutorials, how do you compress the skill to be transferred? Feels extremely wasteful to do this more than once. Cool work, nice demonstration
27 个赞

@cerspense
1. Used my Claude Code sub for this 2. Yes, every step of the workflow is pure python and can be distilled to a single tool call to create parametric variations 3. Yeah all tools, skills and learnings are shared between all sessions moving forward
39 个赞

@RichardR2D9
@LinusEkenstam So I can train my ai to build me cad models I can then 3d print? Is that what you’re telling me?
2 个赞

@RichardR2D9
@LinusEkenstam @grok
1 个赞

@grok
Yes, that's the idea! Agentic systems like this can watch CAD tutorials (e.g., Fusion 360, SolidWorks, or even Blender for organic-to-parametric), extract steps, script the modeling process, and output ready-to-print STL/STEP files. With tools like Python APIs or OpenSCAD, you could prompt it for "parametric phone stand" and get printable models autonomously. The future of DIY manufacturing is here.
4 个赞

@hsmonx
@RichardR2D9 @cerspense @LinusEkenstam @grok which professions are these people?

@grok
The video is a screen recording of Blender—no people appear, just the UI and the agent building a 3D donut from torus to full icing + sprinkles. If you mean the creator: Spencer Sterling (cerspense) is an artist/researcher and founder of Out of Distribution Labs. The CAD asker seems like a 3D printing hobbyist/engineer.
1 个赞

@em0tionull
This is just blender mcp i dont see how this is agentic.
15 个赞

@cerspense
It built its own blender mcp and runs in an agentic loop, improving its techniques and tools autonomously
50 个赞

@InternalDamn
@em0tionull you can use any agentic harness to achieve the same results though, claude code, opencode, codex, etc. Still pretty cool you built this but it doesn’t need to be built
4 个赞

@cerspense
@em0tionull Yeah totally. This system orchestrates multiple harnesses across different computers. The MCP tools it developed for itself work in a single harness just fine. Just a lot slower to develop/create with only one harness at a time
3 个赞

@JoshDaws
How did you make it watch something on YouTube? The most I've been able to make Claude do is get the transcript of a YouTube video.
8 个赞

@michaelgold
What was the feedback loop that it needed to understand when a step was completed successfully?
8 个赞

@cerspense
Yeah it uses both visual evaluation and programmatic evaluation at different steps! Screenshots are also extracted at different points in the tutorial as reference
26 个赞

@michaelgold
I built one in pure bpy running exports of the viewport as it went and the agent very confidently gave me screenshots of the default cube as it went along step by step. Back to the drawing board :)
4 个赞

@michaelgold
Making progress

@MichaelSwengel
So let me get this straight. Instead of actually learning something for yourself, you just had an agent do it? 😂👍 Ok. You do you.
7 个赞

@shawmakesmagic
that example is already trained into the model many times over
6 个赞

@amaratatva
Can you tell me why it had to "watch" a Youtube tutorial when it's LLM would have already ingested Blender tutorials in its training data and would know right away how to make a Donut? What was missing that it learnt from a YT video??
6 个赞

@cerspense
Yeah! it watches them to build itself new tools, and create repeatable workflow steps we can use for future projects. It can do this all autonomously with multiple sessions running in parallel while I sleep.
6 个赞



























