Everyone seems to agree on one rule: if you want an AI-generated video, you need an AI video model. Sora, Veo, Kling, pick your favorite, type a prompt, wait for the render, pay the credits. That’s the whole playbook, right?
Turns out, no. I just came across a post from this LinkedIn creator who produced a full animated explainer video without a single AI image model or AI video model touching it. Not one frame came from a diffusion model. The entire thing was 100% AI-coded, and it honestly reframed how I think about the phrase “AI video.”
The myth: AI video has to come from a video model
Video models are amazing at one thing: dreaming up pixels. You describe a scene, and they hallucinate something that looks like it. That’s great for mood, drama, and cinematic shots. It’s terrible when you need a specific layer of silicon to sit exactly where it belongs in a diagram.
The author’s approach flips the logic. Instead of asking a model to imagine what a process looks like, you ask an AI agent to research the process and then build it, piece by piece, in a real 3D tool. The output isn’t a guess. It’s a rendered scene made from code, which means every object, camera move, and timing decision is deterministic and editable.
What the original poster actually did
The workflow is surprisingly short, and that’s the part I love. No sprawling pipeline, no ten tools stitched together.
- Asked GPT-6 Astra on Codex to research how an AI chip is made
- Had Astra create an animation of that process using headless Blender (Blender running from the command line, no interface open)
- Generated a song with Gemini
- Asked Astra to edit and finalize the video with that soundtrack
The result? The author never touched Blender, but received a 3D model. Never opened a video editor, but received an edited video that met every requirement. The agent handled the research, the modeling, the animation, the editing, and the final assembly.
Why it matters: this is the moment where “I can’t animate” or “I don’t know Blender” stops being a blocker. The skill that counts is knowing what the video should say, not knowing where the buttons are.
Where the code-based route wins
The creator is clear that this opens up new possibilities, especially when precision matters more than visual effects or drama. Let me spell out why I think that’s true.
- Accuracy: A wafer has a set number of steps. Code will render those steps in order, every time. A video model might merge two, skip one, or invent a fourth.
- Cost: The author says the bill was a tiny fraction of what the same video would cost through AI video models. Rendering from code is cheap compute, not per-second generation credits.
- Iteration: Want the camera to hold two seconds longer on the etching stage? Change a line. With a video model, you re-roll and hope.
- Consistency: The same chip looks the same in scene one and scene nine. No drifting shapes, no morphing labels.
Put those together and you get exactly the kind of content the creator predicts will move this way: most explainer videos and procedural animations. Think product walkthroughs, manufacturing processes, medical mechanisms, architecture flythroughs, software architecture diagrams that actually move.
The real formula: domain knowledge + AI execution
This is my favorite point in the whole post, because the author is refreshingly honest about it. They admit they have no domain knowledge of semiconductors, so their chip video is, in their words, the low bar. Someone who understands how chips are actually fabricated, plus a bit of AI skill, could produce a much better video than the example.
Read that again if you’re an expert in anything. The advantage isn’t with the person who knows Blender best. It’s with the person who knows the subject best and can direct an agent well. Chemistry teacher? Mechanical engineer? Logistics planner? Your knowledge is the scarce ingredient. The rendering is now the easy part.
The honest downside
This isn’t a replacement for video models, and the author says so directly. Photorealistic, cinematic shots are harder to achieve through code. Not impossible, but a top AI video model is usually the better and faster option when you want lighting, faces, atmosphere, and emotional punch. Blender through an agent can get there, but you’d be fighting for every frame.
So the smart split looks like this: drama and realism go to the video model, precision and procedure go to code. Pick the tool that matches the job instead of defaulting to whichever one is trending.
How to try this yourself
If you want to test the approach without a big time investment, this is the path I’d take based on the creator’s setup.
- Pick a process you already understand deeply. Your knowledge is what makes the video good.
- Ask a coding agent to research and write out the steps first, then review that outline yourself before any rendering happens.
- Have the agent build the scene with headless Blender and render a rough, low-resolution pass so you can check the flow.
- Generate music separately with an audio model, then hand the agent both files and ask it to cut, sync, and export.
- Iterate on the script, not the pixels. That’s where the cost savings live.
The author’s closing advice is simple: for education and design videos, consider the code-based approach too. It may open up completely new territory for you. After seeing this, I believe it!
Head over to the full LinkedIn post to watch the chip animation itself and see exactly what a coded video looks like when no video model is involved.