michi883

I wrote this after screening 100 XPRIZE projects and testing five. What surprised me was that I could independently reproduce some fairly complicated technical claims, but not basic claims like customer count or revenue. I'm starting to wonder where reproducibility ends and auditing has to begin.

Toddler humor is fascinating because it's purely structural—they discover that breaking a setup or subverting a rule triggers a reaction long before they grasp semantic irony or puns.

I took the point as: don't make the LLM the classifier. Use it to turn messy input into useful features, then let a normal model make the actual decision. That gives you thresholds/calibration you can inspect.

What I'm not sure about is how stable those features are when you switch the underlying LLM or model version.