cross-posted from: https://lemmings.world/post/21993947
Since I suggested that I’m willing to hook my computer to an LLM model and to a mastodon account, I’ve gotten vocal anti AI sentiments. Im wondering if fediverse has made a plug in to find bots larping as people, as of now I haven’t made the bot and I won’t disclose when I do make the bot.
Reddit is different than fediverse. They work on different principles and I argue fediverse is very libertarian.
Is there anyway you can rule out survivorship bias? Plus I’m already doing preliminary stuff and I looking into making response shorter so that there’s less information to go on and trying different models
What kind of models are you planning to use? Some of the LLMs you run yourself? Or the usual ChatGPT/Grok/Claude?
So far I’ve experimented with ollama3.2 (I don’t have enough ram for 3.3). Deepseek r1 7b( discovered that it’s verbose and asks a lot of questions) and I’ll try phi4 later. I could use the chat-gpt models since I have tokens. Ironically I’m thinking about making a genetic algorithm of prompt templates and a confidence check. It’s oddly meta
I often recommend Mistral-Nemo-Instruct. I think that one strikes a good balance. But be careful with it, it’s not censored. So given the right prompt, it might yell at people, talk about reproductive organs etc. All in all it’s a job that takes some effort. You need a good model, come up with a good prompt. Maybe also give it a persona. And the entire framework to feed in the content, make decisions what to respond to. And if you want to do it right, and additional framework for safety and monitoring. I think that’s the usual things for an AI bot.
If you manage to do it, maybe write some blog post about the details and what you did. I think other people might be interested.