Pascal's Playground
In progress: being built now. The first three builds are approved.
A ladder of harder and harder overnight builds, starting with interactive 3D web pages, that Pascal attempts using only local models. A separate checker judges each night's try, and the rung where Pascal starts failing shows what local models can really do.
- Why
- My other build pipeline uses frontier models for review, so it never showed what local models can do on their own. I want a real answer, not a benchmark score.
- How long
- Requirements written in October 2026
How I built it
- I had Claude Code interview me and write the requirements: who proposes, who approves, who checks.
- Pascal researches what other people have built on similar setups and proposes builds with a stated difficulty.
- I approve, edit or reject each one in the morning report. I approved the first three rungs.
- Each night Pascal makes one attempt, and a checker that Pascal cannot touch decides whether it works.
- To copy it you need
- A Mac with enough memory for local models, Ollama, an agent that can run on a schedule, and Claude, or the AI of your choice, to plan. Gotcha: the thing that builds must never be the thing that grades.
Try it yourself
Copy this prompt and paste it into ChatGPT or Claude to try it yourself.
I want to find out what the AI models on my own computer can really build. Korab Eland is setting up Pascal's Playground (on whatarewebuilding.ai): an agent proposes a ladder of builds, starting with interactive 3D web pages, makes one attempt a night using only local models, and a separate checker decides if each attempt works. The rung where it starts failing is the answer. Help me plan a small version. First ask what computer I have and suggest a local model that fits. Then help me write three rungs, easy to hard, with a plain pass or fail test for each. Korab's lesson: the builder must never grade its own work, and 3D pages need scripts, so don't block them. Ask me for anything you need.
Tools I used for this project
