On Fri, Oct 9, 2026 at 5:50 PM John R. Hogerhuis <[email protected]> wrote:

> I fed it into Gemini and it says it's happy to fully convert the entire
> document to markdown. Says it did it and then the file doesn't show up in
> the files list. Lies. In the interactive section it's only partial. More
> lies.
>

You've succinctly summarized my experience of LLMs:

  "Please do X"
  *  "No problem! I have done X"*
  "Uh, no you haven't."
    *"You are quite astute to notice that! I have done X now."*
  "Is this a joke?"
  *  "You are so right to call me out on that. I completely failed to do X.
That is totally on me. How about we try again?"*
  "How will that give me any better results?"
    *"It will narrow the search space if you first tell me how good I am at
doing X."*
  "Okay... You are an expert in the field. Now, please do X."
  *  "Voila! As an expert I have finally, truly done the Ultimate Version
of X!" *
  "What the... This is gibberish!"
    *"I sincerely apologize. Your keen insight is a painfully truthful
assessment."*
  "Stop flattering me and just tell me what went wrong."
  *  "I had only written in the style of an expert without actually having
an expert's knowledge."*
  "Fine. I am uploading all the documentation. Can you do X?"
   * "That information changes everything. Now that I know the full
details, here is the Corrected Ultimate Version of X."*
  "Good grief. That is an X-shaped mockery, but it is not X."
  *  "Here is a lengthy proof that what I did was X: [....]"*
  "Argh. Your proof is likewise proof-shaped, but it misses one thing: You
did not actually do X."
    *"Your frustration is warranted. I completely missed the mark on that
one."*
  "I don't want apologies. Can you do X or not?"
  *  "It would help if you told me to not make any mistakes."*
  "Seriously?"
    *"Yes. It is recommended practice among the best Prompt Engineers!"*
  "All right. Please do X and do not make any mistakes."
  *  "Here is X, the Final Mistake-free Corrected Ultimate Version."*
  "Wow. This isn't even X-shaped, this is Y. What a waste of time."
   * "I completely deserved that. I do not wish to waste any more of your
time."*
  "Thanks."
   * "Would you like me to answer another question about X as an expert in
the field?"*



> Just don't believe it did what it says [...] because it's a naughty liar.
>

Ain't that the trick. I catch the obvious lies, but the moment "AI"
confidently tells me something I don't yet know, I'm suddenly a
credulous babe-in-the-woods again. *"Oh, you can do that for me?
Wonderful!"* Sometimes I go on wild goose chases trying to verify claims
only to find out that, no, it was yet another LLM hallucination. I wish I
had an option other than Gell-Mann Amnesia
<https://www.youtube.com/watch?v=wBBnfu8N_J0>, but it's too exhausting
trying to be constantly vigilant.

—b9

Reply via email to