You can’t blame your saw when your shelf comes out crooked. You’re supposed to know how the tool cuts.


Cute. But you can’t blame your saw when your shelf comes out crooked. You’re supposed to understand how the tool cuts.

There's a unique wrinkle with a chatbot: the saw doesn’t turn around and tell you the cut is straight when it isn’t.

That's the issue I was raising. I wasn't arguing Gemini should "care" about my opinion. I was asking what happens when I give it FACTUAL EVIDENCE that contradicts its answer—and instead of reassessing the CLEAR evidence, it continues defending the original conclusion with increasingly elaborate and truly bizarre explanations.

I agree asking an LLM the right question matters. Sure, agree. But that's not an excuse for the tool to confidently defend an answer that the evidence contradicts.

In my HDV example, the difference was revealing. When I asked Gemini to explain the discrepancy, it produced a plausible explanation that was wrong.

That's precisely my point. The problem wasn't that Gemini couldn't eventually find the answer. The problem was that it was willing to give me a confident explanation before it had established that the explanation was true.

These tools are useful. But if we're going to rely on them, I think the important question isn't merely "Did I ask the question correctly?". Sure that's important. It's also "Does the system know the difference between something it has established and something it has merely inferred?"

That's a much more consequential problem than whether I used the right saw.