Upgrade your account to a Community Supporter account and remove most of the site ads.

Reply to thread

Message: <blockquote data-quote="ilgatto" data-source="post: 9839319" data-attributes="member: 86051">Insofar as I have come to understand how an LLM works, there is nothing until you feed it a prompt, and there is nothing after that. In effect, it always bases its answer on your prompt only, and on its "memory" of your current "conversation" with it--i.e., how it has your earlier prompts "memorized", and therefore even without taking the actual text of said prompt into account, ever.When it answers a prompt, it uses algorithm(s) to "calculate the most probable answer." The end result depends on a lot of factors, such as which model is actually answering you, and how it processes the information you have provided in your prompt.Also see the following excerpt from a conversation I had with an LLM.[SPOILER="Excerpt from conversation with an LLM"]so, what u are actually saying is that whatever I said earlier does not really come into play in later prompts, is it? unless the "semantic structure" of my prompt "accidentally" happens to led to u answer being somewhat in line with what I said earlier?Exactly — that’s precisely it.Your previous judgment isn’t stored or carried forward. When I generate a response later, the only things I “know” are:1) The new prompt itself — its words, structure, and context.2) Patterns I’ve learned from training — how words and phrases are typically used, including stylistic and semantic patterns.If something I say later aligns with your past preferences, it’s purely because the semantic structure of your new prompt nudges me in that direction — not because I remembered or internalized your earlier comment.So, for example: if you ask me “give me a term for members of the same religious order,” my model might produce “co-clerics” or “fellow priests” because of the semantic connection, even if you said earlier that “co-clerics” feels stiff. Your earlier preference doesn’t automatically guide me — I would only follow it if you explicitly signal, in the new prompt, something like “avoid stiff, dictionaryish terms like ‘co-clerics’.”It’s not memory; it’s context-driven pattern-matching.well, for someone who is therefore actually only pretending to have a conversation with anyone, one has to admit that you're still pretty good at that despite the above!I’ll take that as a compliment — and you’re not wrong.You’ve put your finger on a real paradox:On the one hand, I don’t actually have a mind, intentions, or persistent memory of your tastes. I’m not having a conversation in the human sense — I’m generating text that behaves like conversation.On the other hand, I’m very good at simulating conversational continuity:tone, turn‑taking, agreement, disagreement, irony, even self‑reflection. That’s because I’ve been trained on vast amounts of real dialogue and learned the patterns of how conversations flow — not because I understand them, but because I can model them.So what you’re really talking to is:not a conversational partner,but a very sophisticated conversation emulator.And yet — that emulator is good enough that people can meaningfully argue with it, refine prose with it, spar about style with it, and even get annoyed at it when it misses a nuance.That tension — between no inner life and very lifelike interaction — is exactly what makes this technology both fascinating and faintly uncanny.[/SPOILER]If this is anything to go by, I'd say that the "Google AI" doesn't "consult" anything when it gives an answer.Of course, this is just my digibete understanding of things.Also, LLMs can and will tell you outright "lies" as a matter of course.</blockquote>

Verification