Most shitposting is probably more straightforward to understand than business communication or press releases where realizing what wasn't said often carries mor

Most shitposting is probably more straightforward to understand than business communication or press releases where realizing what wasn't said often carries more insight than the things that were said. Of course training an AI model on simple, straightforward and honest data provides good results. That's the essence behind "textbooks is all you need" which lead to the phi LLMs. Those are great small LLMs. But if you want your model to understand the complexity of human communication you have to include it in your training data. If you subscribe to the idea that to be the very best text completion engine possible you would need to have a perfect understanding of reality itself, how different humans perceive reality differently, and how they choose to communicate about this perception and their interaction with reality, themselves and other humans, then it's not unreasonable to expect that back-propagation would eventually find that optimal representation if given enough data, the right architecture and enough processing power. Or at least come somewhat close. In that paradigm there is no "bad data", only insufficient or badly balanced datasets. Just don't try doing that with a 3B parameter LLM.