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.
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.