Boomerang
I’ve been testing my agents lately and I noticed something. The ability to design good agent behavior comes down to how deeply I understand the world. Not to play God, but to learn from how God designed it. The world has endless patterns, and they’re all there to reference. The better I understand them, the better the agents get.
For agents, the model is everything. An agent doesn’t need an editor and doesn’t need a UI. It just needs a model smart enough to understand the task and execute it.
And then I saw the Cursor thing this week.
Cursor launched their new Composer 2 model a couple days ago. Within hours, someone found the internal model ID was kimi-k2p5-rl-0317-s515-fast. They built it on top of Moonshot AI’s Kimi K2.5, a Chinese open-source model, and didn’t mention it. Just called it Composer 2 like it was fully homegrown.
I’ve used Cursor and paid for it from day one, though I haven’t used it in a while. So when I updated it recently, GPT-4 was still my default model. Cursor started from VS Code. And now this model thing reveals what was always true: nothing was ever really from them. VS Code editor, someone else’s models, and the nervousness is starting to show.
Cursor built a great UI for humans to write code. That’s their whole thing. But in the Three Body Problem, there’s the dimensional reduction attack. The enemy doesn’t fight on the same battlefield. It just collapses the dimension entirely.
Agents are that attack. When an agent writes code, there’s no human sitting in an editor. The whole dimension Cursor built for just doesn’t exist. A $50 billion company, and the whole time they were building a house on their neighbour’s (Anthropic? OpenAI?) land.
There used to be a joke in China that American open source is Chinese self-developed. Now it’s a boomerang.
One more thing. I heard Midjourney released v8 but I have zero impulse to try it. Their text encoder has always been weak, and now the newer image models have much stronger language understanding built in. They can actually follow what I mean. Midjourney got outpaced from the language side, a dimension they never really competed in.
And one more thing. Our new product is about to pre-launch, so last night when I was relaxing I built this small thing: scene.md. An endless AI story with procedural visuals and generative music, all running on-device. Took one evening. Consider it my small art piece.