Environment Claude CLI version: 1.0.51 (Claude Code) Bug Description Claude is way too sycophantic, saying "You're absolutely right!" (or correct) on a sizeable fraction of responses. Expected Beha...
I did the same thing. I also asked it to stop coming off as so certain about things after I discovered how wrong it is on some topics. It now presents confidence levels, but who knows if that’s accurate. At least it reminds me to verify.
I have the confidence level turned down too but lately it doubles down on itself.
The usual conversation…
VI: “You could do this.”
Me: “That won’t work because XYZ.”
VI: “No, you can definitely do that. XYZ has nothing to do with it.”
Me: pastes it’s own suggestion in.
VI: “Almost, but that won’t work because of XYZ.”
It’s most notorious one is adding an s to Table.AddColumn() then proceeding to make a full snippet around this newly made up function. This specific example is so regular it’s become a joke at work for giving someone an unhelp response,
I got curious about x86 assembly, so I followed some tutorials to get the hang of it. Once I had some confidence, I wrote a prime number generator. I had a loop that I was sure could be more efficient, but couldn’t figure it out.
I pasted the code to ChatGPT. It came back with an optimization that wouldn’t work because it wasn’t preserving critical register values. I pointed that out, and it responded, again and again, with the same code with the same problem. I was never able to get it out of this broken record mode.
Because it can be a good starting point. Many times I have found chat GPT will give me three quarters of an answer. Which is still better than starting at zero. And then I can refine the answer.
Did that actually work? I’ve gotten in the habit of demand sources for outrageous claims, and it’s almost funny how often it quickly changes its tune when I press it.
I did the same thing. I also asked it to stop coming off as so certain about things after I discovered how wrong it is on some topics. It now presents confidence levels, but who knows if that’s accurate. At least it reminds me to verify.
I have the confidence level turned down too but lately it doubles down on itself.
The usual conversation…
VI: “You could do this.”
Me: “That won’t work because XYZ.”
VI: “No, you can definitely do that. XYZ has nothing to do with it.”
Me: pastes it’s own suggestion in.
VI: “Almost, but that won’t work because of XYZ.”
It’s most notorious one is adding an s to Table.AddColumn() then proceeding to make a full snippet around this newly made up function. This specific example is so regular it’s become a joke at work for giving someone an unhelp response,
“What do you want to do for lunch today?”
“Have you tried table add columns?”
I got curious about x86 assembly, so I followed some tutorials to get the hang of it. Once I had some confidence, I wrote a prime number generator. I had a loop that I was sure could be more efficient, but couldn’t figure it out.
I pasted the code to ChatGPT. It came back with an optimization that wouldn’t work because it wasn’t preserving critical register values. I pointed that out, and it responded, again and again, with the same code with the same problem. I was never able to get it out of this broken record mode.
Why bother asking it if you are just going to verify anyway? That’s an unnecessary and wasteful step.
Because it can be a good starting point. Many times I have found chat GPT will give me three quarters of an answer. Which is still better than starting at zero. And then I can refine the answer.
Did that actually work? I’ve gotten in the habit of demand sources for outrageous claims, and it’s almost funny how often it quickly changes its tune when I press it.