Maybe we underestimate people a bit. The assholes tend to be more impacting to us, but most people aren’t like that, and we tend not to notice the several neutral or good interactions the same way.
Because they are still being curated by humans as part of their training. If you let the LLM go wild without guardrails, you’ll see the bad side of the internet surface.
They have RLHF (reinforcement learning from human feedback) so any negative, biased, or rude responses would have been filtered out in training. That’s the idea anyway, obviously no system is perfect.
Then why are they all still smarmy assholes?
That’s what was said. LLMs have been reinforced to respond exactly how they do. In other words, that “smarmy asshole” attitude, you describe was a deliberate choice. Why? Maybe that’s what the creators wanted, or maybe that’s what focus groups liked most.
They’re talking neutral by default, but they absolutely talk trash if you prompt them to.
They do.
which llm are you using?
4chanGPT maybe?
I want this so much.
It was a thing, but Huggingface removed it due to the surrounding drama
Oh wow
Yeah, I can imagine it’d be horrible and dark, a satire of the dark-side of humanity to the point of hilarity.
hmm, it has a torrent available on web archive https://archive.org/details/gpt4chan_model_float16, but it seems to be in very outdated .bin format. If you have experience with llama.cpp tooling you might be able to convert it into something usable… just be careful, this old format isn’t reinforced against malware
with ChatGPT you can tell it to behave in certain ways. With Claude it’ll just start mimicking you.
Any would talk however you prompt it to talk.
