When I first started leaning on a model for real work, I’d take the answer and run. It sounded certain. It was formatted like something true. And it was so fast that stopping to question it felt almost rude.
Then it burned me a few times — confidently, fluently wrong, in a way that would have been obvious if I’d only asked it to slow down and reason out loud. So I changed how I use it. Now I rarely just take the conclusion. I ask it to show the work.
“Walk me through why.” “What are you assuming here?” “What would make this answer wrong?” Watching the reasoning unfold is where I catch the flaw, because a hidden assumption looks like confidence, but a stated one looks like something I can check.
It changed the relationship. The model stopped being an oracle I either trusted or didn’t, and became more like a sharp colleague who has to defend a claim before I’ll act on it. Sometimes the reasoning is airtight and I move faster. Sometimes it falls apart in its own explanation, and it just saved me from a mistake.
An answer I can’t inspect is a guess in good handwriting. I make it show the work now, every time it matters.
– Serguey Asael Shinder
Leave a Reply