Zhangcan Ding

Building & — turning ideas into presentations and videos. Neuron reconstruction PhD; relearning what neurons actually do→

[Zhangcan Ding @jason_dingzc]

I build PageOn and iArt — tools for explaining things. Watching over 100k people use them made me curious about explanation itself.

I also spent a PhD reconstructing neuronal morphology and never really understood what neurons do. So I'm relearning it in public — on a site that's also an experiment in explanations you can poke at.

First post — the textbook firing threshold isn't real.

https://t.co/svZrMmwzOo

[Zhangcan Ding @jason_dingzc]

New in iArt 🎆 it now writes an original song and generate the whole video to the beat — all from one prompt. Made this little Fourth of July piece to show it off. Happy 4th to everyone celebrating — go make your own.

[Zhangcan Ding @jason_dingzc]

@kapildevtamkr @prasenx And if you'd rather not pay to export — we open-sourced the whole thing: https://t.co/ixnN5jmCJv. if you've got a coding agent (Claude Code, Cursor, Codex), you can build and render the same kind of video yourself, for free. iArt's just the no-setup version of that.

[Zhangcan Ding @jason_dingzc]

@kapildevtamkr @prasenx you've already got it 👍 the video's live and shareable via that link — send it to anyone as-is. free = create, preview, share (all done). exporting the file to your device is the paid part — grab a plan whenever you want to export it locally.

[Zhangcan Ding @jason_dingzc]

for sure 👍 honestly you don't need me — it's built to do it itself. sign up at https://t.co/Lt6wY1xUNd, open a chat, and just describe what you want. paste a URL or upload your images/logo, tell it the style + length, and it generates the video. the free tier's enough to try it right now.

[Zhangcan Ding @jason_dingzc]

@prasenx this is the sweet spot for agent-written Remotion. we open-sourced motion-skills for exactly this — they teach a coding agent the craft (timing, easing, safe areas, frame-accurate render) so the output stops looking generic: https://t.co/ixnN5jmCJv

[Zhangcan Ding @jason_dingzc]

@kapildevtamkr @prasenx If you don't want to write remotion code, iArt does the same thing from a prompt — describe your product and it generates the launch video. that's the no-code route to what you're seeing here.

[Zhangcan Ding @jason_dingzc]

I check our community almost every day and it's honestly the best part of building iArt. A user made and shared this little story about friendship, in Urdu 🌈. People are making things more beautiful than I ever imagined.

[Zhangcan Ding @jason_dingzc]

The update: a brief compiler is now built into iArt — your rough ask becomes a director's brief (shot list, beats, pacing) you can tweak, then it generates. browse or reuse community templates on your own project:
https://t.co/pOjUui0uQp

[Zhangcan Ding @jason_dingzc]

Most AI video looks like slop because the prompt is too thin. That's normal, most of use aren't video directors, are we? So this week we added a step: iArt compiles your one-line ask into a director's brief — shot list, beats, pacing — before anything renders. this 👇

[Zhangcan Ding @jason_dingzc]

If your ai-generated motion looks dumb, you're probably using the agent wrong. Coding agents can animate anything — but they can't draw. So agent-made motion always looks geometric. This is the fix.

The unlock: code can't illustrate. So let a vector model like Recraft draw the parts — one flower, one weed, each clean and isolated — then let code compose and move them.

The structure lives between the assets, not inside one image. that's what makes it feel designed instead of like a wiggling picture.

Open-sourced the whole thing as a new pack — generative-illustration-skills — in our motion-skills collection. 15 packs,51 skills now, all teaching coding agents to make motion. https://t.co/AoiJ51Fqh0

[Zhangcan Ding @jason_dingzc]

Compounding looks like nothing is happening — until it's everything.

One dot, doubling once a day, to 2,048.

Made this with the motion tool we're building.

[Zhangcan Ding @jason_dingzc]

@jomatech As someone actually building an AI startup i was sweating the whole time waiting for the punchline to bail me out... got to the end like oh. oh you're serious this time 💀

[Zhangcan Ding @jason_dingzc]

Warmer. softer. again. almost. not quite.

If you've made anything with AI, you know this part — a hundred prompts that wither, and then one that finally blooms.

Every visual here is AI-generated, animated entirely in code — the direction we're building iArt ai toward. sound on 🌱
#generativeart #creativecoding

[Zhangcan Ding @jason_dingzc]

The race for the World Cup's last knockout spots — qualification probability, matchday by matchday.

Scotland sat at 82%, then vanished.
Ecuador flatlined, then rose from the dead beating Germany.
Paraguay climbed from 42% to safe.

20,000 sims per matchday on the real results. The last spots get decided today.

Animated with the motion tooling we build. #WorldCup2026

[Zhangcan Ding @jason_dingzc]

We're on the quiet road - start from structure, pull the visuals up. A model writes the code, we render it, then check the pixels and have it fix what's broken — because you can't tell a motion is wrong by reading code, only by looking.

Our visuals are still catching up. But for real work — product, data, explainers — exact and editable is the half that actually matters.

We open-sourced the craft behind this → https://t.co/AoiJ51Fqh0

[Zhangcan Ding @jason_dingzc]

We have AI that generates photorealistic video from one sentence. It still can't put your product name on screen correctly, or render your actual UI. That's not a quality problem — it's a representation problem.

There are basically two ways AI makes visuals: it generates pixels, or it generates structure.

Pixels — image and video models. Gorgeous, and almost the entire industry is here, because the internet is full of pixels and diffusion was built for them. But a pixel output is a finished frame: you can't truly edit it, and you can't make it exact. You get something that looks close. Not the thing.

Structure — code, vectors, scene graphs. The model generates the representation, and the picture is rendered from it. That makes it editable (change one thing, only that changes) and exact (the text is the text, the number is the number). The catch: raw visuals still trail — clean structure is scarce in training data, and symbolic precision is the one thing models are worst at.

So each road owns half. Pixels nail the look. Structure nails control. And almost everyone is racing down the pixel road — not because structure is worthless, but because it's harder.

[Zhangcan Ding @jason_dingzc]

Step back and this is an old fight in new clothes.

Pixels = distributed, learned, sub-symbolic representations — the connectionist side.
Structure = discrete, compositional, symbolic representations — the classic-AI side.

Visual generation is the connectionist vs symbolic divide, resurfacing. And the endpoint everyone is circling — beautiful AND exact AND editable — is just the old dream of getting both halves at once.

[Zhangcan Ding @jason_dingzc]

Honestly still early — launched a few weeks ago. the signal I care about more than raw count: a real chunk of signups come in organically from ChatGPT recommending us, zero spend. And a creator started making YouTube videos about us completely on their own, which has already converted to paying users. Happy to get into actual numbers over DM if you're curious.

[Zhangcan Ding @jason_dingzc]

We just open-sourced 50 skills that teach AI coding agents to make motion graphics.

Kinetic typography, data charts, explainers, TikTok/Reels, WebGL, Manim — 14 installable packs. Works with Claude Code, Cursor, Codex.

The problem: motion is the hardest thing to get right from a prompt. Timing, easing, safe areas, loops that actually loop. These skills encode how a professional motion designer would build it — with a verify loop so the agent checks its own output.

MIT licensed.

https://t.co/AoiJ51Fqh0