Make a video from your phone in three minutes. See how we saw the shift two years early. https://x.com/i/article/2117999225669644288
How to See into the Future: Building Ahead of the Curve
Part 1: Prediction: AI will collapse
Co-written with @weightrotator.
We predicted the downfall of generative AI video almost two years ago.
So we built the easiest-to-use video software known to mankind so you can start creating real, rendered animations immediately.
You describe the video; the AI agent writes the code; our GPUs return the finished file. Without installing yet another new app, you can start making videos in minutes from any device. On the toilet, waiting at the airport, or walking down the street. No technical background or expensive hardware required.
That's 67idea.com. Read to the end and I'll show you how to get free, limited-time access and start making videos from your phone or 10-year-old laptop in under 3 minutes. Just try not to rack up a huge bill for me.
The rest of the AI sphere is only now discovering that code can create better videos than diffusion models. Yet we knew this almost two years ago. So how did we know?
Suppose you're fascinated by barbell path tracking. Not because it is a complicated problem, but because it is a stupidly easy problem, and AI can’t solve it. The singularity is here, and AI still can’t track barbells! This would bother you. Driven by curiosity, you try to figure out why AI sucks at barbell tracking. Along the way, you make a lot of unexpected discoveries.
- Discovery 1: AI doesn’t just struggle with tracking barbells. There’s an entire class of problems it struggles with.
- Discovery 2: the problem is with the sandbox. There are problems agents cannot solve, as long as they do all their work on a CPU. It does not matter how smart the model is.
So you would build a generalized compute sandbox for video. It walks through a video, runs a bunch of compute each frame, and outputs a new frame. There are millions of seemingly unrelated problems that this solves.
- Use case 1: 3d rendering engine.
- Use case 2: tracking barbells. But now, you don’t just track barbells. You create a full 3d reconstruction of the scene, calculating force, velocity, power.
I shared these thoughts with @weightrotator (prev MIT, founder backed by 1517 Fund), and our conversations convinced him to quit his job and move to the middle of nowhere to build with me in stealth. I ended my last viral post with three questions. We've spent the last year and a half building the solutions to them.
The questions weren't about video. They were about the assumptions every AI company is building on and almost none question. It's funny that the goal of AI is general intelligence, yet most VC-backed AI startups are built around solving niche human use cases with GPT wrappers. In a gold rush, everyone is building apps from 2017 enhanced by AI; we should be thinking about how to build for 2067.
AI agents are our direct product users. Consumers are the users of the AI agents that use our infrastructure. That's how we've been building since day one.
The paradox is that building software for LLMs to use is what makes the human experience seamless.
You only need to know how to send a chat message to use ChatGPT; you don't need to know how it works.
We did that for video made with code, not generated by diffusion AI. (And for anything else that needs remote GPU execution.)
Our first customer, a video marketer, begged to pay us before we even launched. A few days later, he hit 1M views on a brand new Instagram account. A few other early customers stayed up all night making videos. Most of them didn't know what GitHub or agentic coding were. Yet they were able to set up and start creating videos within minutes. They just loved what they saw and got addicted to using us daily. We have achieved consumer retention.
The customer never sees the backend. They just feel it as speed and ease. In our internal benchmark, we delivered a 20-second 1080p video of a 2.1-million-polygon scene in 67 seconds. Blender on Modal took 279 seconds—over four times as long, on a better GPU than we used for our trial. That includes setup, rendering, encoding, and delivery.
Video is already most of the internet. Cloudflare puts it at 65 to 80 percent of traffic. Everyone thought that websites were the hot thing during the dot-com bubble. It turned out that most people are not creating their own websites; they are watching videos all day on social media. Similarly, the primary use of agentic coding will not be for making new websites and apps, but for creating videos.
And then headless agents — agents with no human in the loop, running all day, doing work in the background. They don't need a screen. They need to run code on images and video: inspecting a factory line, reading a medical scan, steering a robot, counting a field. Every one of them needs to execute code on remote GPUs.
I said we're solving millions of seemingly unrelated problems. Video is one. Barbell tracking is next. The rest you will have to wait and see.
----
Thanks for reading to the end. Here's how you can try Idea for free and make a video within 3 minutes on any device. No signup or app download required.
Use your AI app’s coding interface rather than its regular chat. Some agents may require you to connect GitHub first. Once your agent is set up, paste the following:
pip install idiotproof
IDEA_SECRET_KEY=idea_sk_live_ff6aafa82ff6c68e_BSea4G1AfFdUachOHU7AfsoQJEEUi70d8qbukAyygaY
Use this temporary key to create a sample video
(This is a public demo for non-sensitive content. Sessions normally keep files and jobs separate, but the shared key does not provide security between participants. Don’t upload private material. For private use with your own account and keys, sign up at 67idea.com.)