Planned
The hypothesis, cost, time, safety, and stopping conditions are being worked out.
EXPERIMENTS / EXPERIMENT RECORDS
What we tried
and what happened
Each record covers the purpose, environment, AI and human work, cost, problems along the way, and final result. Failures and cancellations stay in the record too.
EXPERIMENT #007 / COMPLETEWe security-checked
We security-checked
35 AI-built web apps
We checked 35 web apps an AI built from one-sentence requests for outside connections, storage, input handling and keys in code. Login and AI chat included, no problems were found, and some apps added safeguards nobody asked for.
Read the resultsEXPERIMENT #006 / COMPLETEWe tested 20 prompt
We tested 20 prompt
tricks on Claude
Promise a tip, ask for a deep breath, threaten, assign an expert role. We ran 20 widely shared prompt tricks with the same requests and conditions, 66 runs in all. Only “write in detail” clearly changed the answers.
Read the results for each trickEXPERIMENT #005 / COMPLETE10 problems in solo
10 problems in solo
AI game development
We built a 16-game idle game in three days with Claude Code. Here are the 10 problems that came up and how each was fixed, with the development history and cases from other developers. The game is playable in the browser.
Read the development notesEXPERIMENT #004 / COMPLETEI had an AI copy
I had an AI copy
a drawing 100 times
An AI drew a picture, and the next AI saw only that picture and copied it, 100 times for each of three subjects. The subject came through to the end, while the details were replaced by other things.
Read the experimentEXPERIMENT #003 / COMPLETEHow far can one sentence
How far can one sentence
take an AI build?
One sentence per subject, no testing and no fixes, for 15 builds. From shogi and a spreadsheet to a 3D sandbox and a one-hour RPG, all 15 were usable within what was checked. The limit was not found.
Read the experimentEXPERIMENT #002 / ADDITIONAL RUN / TERMINATEDI Gave an AI God a World of 300 People.
I Gave an AI God a World of 300 People.
Humanity Was Extinct 745 Years Later
We gave an AI God a world inhabited by people. It checked the world 152 times and used 46 Actions. Humanity still became extinct.
Read the additional experimentEXPERIMENT #002 / COMPLETEI let AI
I let AI
be a God
AI defined the world itself and then served as its God for 1,000 simulation days. It woke 15 times and never changed the world.
Read the original experimentEXPERIMENT #001 / PARTIALI let AI handle
I let AI handle
the whole game build
Codex produced a solid, working first version—but its fun score was 1 / 5. A revision based on human design feedback reached 3 / 5. Both versions are playable inside the article.
Read the experimentSTATUS GUIDE
Experiment status
Running
The work and events are being recorded before any conclusion is drawn.
Result checked
The outcome has been assessed against the original criteria and what actually happened.
Stopped
The experiment ended because of cost, safety, feasibility, or another documented reason. That reason stays in the record.