Hands-on tests of how AI tools really behave when you push them. Real builds, real data, full source. Each experiment is reproducible — pick one and look at the receipts.
Click into any experiment for the full interactive report.
One app idea, planned five ways, built 25 times by Claude Code and 25 times by Codex, each graded by the other company's AI. Every plan produced a running app; what the plan left out mattered more than how deep it went.
AI does blast out code, and that part is real. Twenty four numbers on what it costs: 17x the code for 1.3x the software, review time up 441 percent, and zero studies measuring what it does to the people writing it.
Karpathy's CLAUDE.md vs the internet's most-liked CLAUDE.md / AGENTS.md files — plus a few of my own. The results aren't what you'd expect.
Prompt-specificity curve (3+ specificity levels), tool-use reliability across model versions, agent-vs-orchestrator latency. New experiments published as runs complete.