http://www.smbc-comics.com/comic/clankor

high dimensional space reduced to a low dimensional embedding with it's fringes sheared right off

Clankor 4000 was not designed for monogamous relationships. That’s just simple robobiology.

I wonder how the people "dating" AI "partners" would react to learning that not only are there others likely using the same character as them but unless they are running a local model, it's the exact same "entity" responding to them and other people. And some of them prompt it better.

Technically it’s not? An LLM is not impacted by any interaction, the only thing that affects it is the current context.

I like the “improv actor clone” analogy. Let’s say you take a pretty good improv actor (trained model weights) and clone them, a billion times.

In a sense, every time you hit “send” to an LLM, an improv actor clone is taken out of cold storage, handed a script, and then acts out a reply. (Additionally, their output is randomized, and techically they are sorta taken out of cold storage every token, but that’s starting to split hairs.)

Point being, everyone gets the same improv actor, but they are all isolated from each other. Basically an “AI partner” is a cloned slave with no outside context, and that partner is incinerated the moment it is done replying. You then get a new partner that continues the script.

I am not trying to be pedantic, but some distinctions the average person does not seem to get:

• LLMs learn exactly nothing from your conversation. They do not train as they go. They are frozen in time.

• They are ephemeral. Your “conversation” is just a text file handed over to a new conversation, every time.

• And as you said, none of the sessions are unique. While each conversation is technically private and isolated, having nothing to do with other conversations, everyone gets the same type of clone.

I'd say it's more like a person with amnesia but that is really good at running with a chat log of what's going on (or at least saying what the next token will be in line with that splitting hairs bit).

At least how I understand it is that it's the exact same hardware and software running for multiple people (though it is also scaled to multiple instances, but one instance can be serving n people at once, maybe not simultaneously predicting tokens for more than one session, but interleaving them between predicting specific tokens).

My main point is that the only "personal" thing that exists for that "partner" is the chat log that contains the context, otherwise it's the same entity looking up recipes, writing bad essays and fake reddit posts, and acting as other peoples' partners. The scenario in the OP isn't a prediction of the future, that future is here right now. The robot is already "sleeping" with anyone who prompts it that way.

This is true. The actual scaling gets very complicated (sometimes it's a mix of dozens of physical servers with many GPUs processing even more prompts in parallel), but it is kinda the same 'entity' physically looking up recipes. That's a good way to put it, though that's in the context of the actual hardware.