A dream that I had last night involving AI presented an interesting concept

1 points by intrasight 2y ago ↗ HN

I'll preface this post by explaining that I have some crazy, weird, complicated - sometimes dystopian - dreams. Most have soundtracks that I record when I wake.

Last night's checked all those bullets. Of course I only remember snippets when I wake up.

I was driving through a large tunnel. I was driving one vehicle and a guy and a young woman were driving in a different vehicle. He was sort of directing our operation. I don't remember precisely what the operation was, but wasn't good - like we're trying to escape from something - what exactly I wasn't sure even in the dream. But my sense was that it was some kind of AI agent run amok.

We stopped and walked for a bit to check things out. I scounted ahead and came back. The guy then explained the AI research program that he was involved in at his university. This was the part that I found interesting enough to post to Hacker News.

They had a federated, competitive AI system. Instead of users interacting with a single app, they were interacting with several LLMs within a dashboard. Each LLM was sort of competing to give the best answer, and the user would choose one of them and copy and paste the results into whatever it was they were doing. The AI's were scored based upon how many times their answer was copied and pasted.

Anyway, I'm curious if such federated, competitive AI systems are a real thing.

1 comment

[ 3.9 ms ] story [ 8.5 ms ] thread

ekimehtor 2y ago ↗

A few thoughts, First off neat dream! I only had a dream about watching a fish tank . Now onto your concept, I recall a video put out by open AI where or two separate AIs played a game against each other but there was no scoring done by multiple humans and the AI were not based on LLMs. After pondering your concept for a while I did come up with the idea of a Pictionary like Game where to Dalle –ISH AIs our task with drawing complex images based on a verbal description and are judged by multiple peoples guessing.

TLDR; you could have two AI using different LLMs compete in a whose line is it anyways type competition, where they are both given a topic and required to tell a joke which multiple humans can vote on how funny it is.

PS, You might also consider using Limited language models or LI LMs as a sort of weight class System. This would involve restricting the number of weights or vectors used in training in order to offer an equal playing field were those with limited hardware funding are not dominated by giant corporate models, a.k.a. Heavyweights!