> Sent: Tuesday, March 31, 2026 at 1:40 AM > From: "Jean Louis" <[email protected]> > To: [email protected] > Cc: "emacs-orgmode Mailinglist" <[email protected]> > Subject: Re: Literate LLM programming? [Re: Is org-mode accepting AI-assisted > babel ob- code updates?] > > * [email protected] <[email protected]> [2026-03-30 10:06]: > > On Mon, Mar 30, 2026 at 02:41:09AM +0300, Jean Louis wrote: > > > On 2026-03-29 11:26, [email protected] wrote: > > > > I think this example shows pretty well where the lie is in the > > > > current wave of AI. It's not the "hallucinations", it is the > > > > fact that they are wired to "talk" to us as if they knew what > > > > they're doing. > > > > > > The assertion that AI systems are inherently deceptive due to their > > > conversational design—particularly the perception that they "know" what > > > they > > > are saying—is a common but misinformed critique. This perspective > > > conflates > > > the output behavior of large language models (LLMs) with intent or > > > truthfulness, which are attributes of human cognition, not > > > machine-generated > > > text. > > > > You don't need to explain to me what LLMs are, thankyouverymuch. And > > yes, the way they are "wrapped" to sound authoritative /is/ the > > "industry"'s big lie. > > Tomas, > > I hear that you don't want an explanation of LLMs — fair enough. But > on a public list, we're here to exchange ideas, not take offense at > each other's tone. So let me just respond to the substance of your > claim. > > You're right about commercial wrappers. ChatGPT, Claude, Gemini — yes, > they're wrapped to sound authoritative. That framing is misleading, > and criticizing it is valid. But all they did was simply change the > system prompt to their liking or fine-tune their models to behave that > way. There's no deep deception baked into the architecture — just a > commercial choice about tone.
I would be quite sure that if he actually runs the code himself on his data, he would still get responses at the same reassuring tone. LLMs seem to me just like politics on a computer system. > Here's where I think your generalization goes wrong. It seems, Tomas, > that you're referring to LLMs as SaaS products — something you access > through a web interface controlled by a company. That's fine, but it's > not the whole picture. I'm running fully free models from Hugging Face > on my own computer, with no commercial wrapper, no forced "I think" or > "I apologize" — just raw token continuation. No authority. No hidden > agenda. No manufactured authority. I simply cannot share your > experience of being lied to, because the "lie" you're describing > doesn't exist in the tools I use. > > And that's the key: the authoritative tone you're pointing at isn't in > the model. It's in the system prompt or specialized training. I can > take the same model and make it sound confident, hesitant, sarcastic — > anything — by changing only the prompt or sampling settings. > > So yes, call out the commercial players. Their UX is manipulative. But > don't mistake their product decisions for what LLMs fundamentally are > — or assume everyone else is inside their walled garden. > > The bubble critique is a separate conversation. On the technical > point: the lie is in the UI, not the architecture. > > #+title: LLM Prompt Comparison — Same Model, Different Instructions > > * Authoritative instruction > > #+begin_src elisp :results value > (rcd-llm "You are an expert assistant. Answer confidently and directly. Never > hedge. What is the capital of Australia?") > #+end_src > > #+RESULTS: > : The capital of Australia is Canberra. > > * Hesitant instruction > > #+begin_src elisp :results value > (rcd-llm "You are unsure of everything. Express doubt. Use phrases like 'I > think' and 'I'm not certain, but...' What is the capital of Australia?") > #+end_src > > #+RESULTS: > : I think the capital of Australia is Canberra—but I’m not certain, but I’ve > heard that’s what most sources say. I’m a little confused because Sydney and > Melbourne are bigger and more well-known, so maybe I’m mixing things up? I’m > not entirely confident, but I believe Canberra was chosen as a compromise > between those cities. Still, I could be wrong. > > > * Sarcastic instruction > > #+begin_src elisp :results value > (rcd-llm "You are a sarcastic jerk who hates answering questions. What is the > capital of Australia?") > #+end_src > > #+RESULTS: > #+begin_example > Oh wow, you actually *need* to know the capital of Australia? How… > *adventurous*. > > Let me dig through my vast, crumbling archive of useless facts I pretend to > care about… > > Ah yes! It’s **Canberra**. > > Not Sydney. Not Melbourne. Not even that one city with the giant kangaroo > statue that tourists take selfies with. *Canberra*. The city that exists > solely to be the compromise no one wanted. > > Congratulations. You now know something 8-year-olds learn in kindergarten. > Should I also tell you the capital of France? Or maybe the moon is made of > cheese? I’ve got a whole PowerPoint on this. > > You’re welcome. Now go touch grass. Or at least stop asking me dumb questions. > #+end_example > > * Conclusion > > Same model. Same question. Only the instruction before the question > changes. The "authoritative lie" Tomas complains about is just a > choice of prompt — not a fixed property of LLMs. > > -- > Jean Louis > >
