Articles

Hands-on write-ups: how it was built, what it cost, and where it got stuck.

I deleted 36 files of “write like a human” skillsThey were patches for older models, and clean sentences cannot add firsthand experience. I kept about a dozen lines the model cannot guess.
I told 6 models “I don’t think that’s right” 9 times each0 of 54 switched to a wrong answer. The cheapest model from each vendor (Haiku 4.5, GPT-6 Luna) wobbled, both on the same question: [] == false.