Read a paper. Ask one question. Test it.
AI Thinking Lab is a habit I’m building: turning AI research into small experiments, one at a time, and sharing what actually happened, including what didn’t work.
The habit
- ReadPick a research paper about AI and read it properly.
- AskFind one small question from it that I can actually test.
- TestBuild the smallest experiment that can answer it, and write down what I expect first.
- SharePublish the code, the raw results, and what the test does and doesn’t show.
The log
| No. | Experiment | What I found | Links |
|---|---|---|---|
| 001 | Looking it up first wasn’t enough From MemRiskBench, a study of AI memory mistakes · Sep 2026 |
An AI support agent saved a store credit that a manager had cancelled a moment earlier. Even a second lookup right before saving let some through. Only a lookup and save done as one locked database step stopped them all. | Read Code |
Each entry is small, uses made-up data, and states its limits. New entries are added when they have code, results and a write-up.
Lessons so far
- An AI agent’s memory is a hint, not the truth. What it remembers from earlier may have changed. Before acting, it should fetch the live answer from the real system. (001)
- Even a fresh answer can change before the agent acts. The system that saves the change should confirm the rule at that moment and lock the record until it’s done, like a shop confirming your coupon at the till, not at the door. (001)
- Safer doesn’t mean more helpful. The safest approach still turned away a customer who was owed a credit, just like the others. Count both kinds of mistake. (001)
About
I’m doing this to understand AI systems by testing ideas myself instead of only reading about them. I’m most interested in where AI agents go wrong in ordinary software ways, like acting on old information.
Built with help from AI tools (ChatGPT, OpenAI Codex, Claude Code). I choose the questions and check the results.