I think Twitter is not a great place to dig out training data in general. Most of its data is not well structured and/or tagged. Its signal to noise ratio is relatively low. Its texts are generally very short and very dependent on context. Twitter has been largely failing as a short-form video platform. There's some trace of algorithmic generation of interest/topic-based feed on Twitter, but you know, its quality was never great. I guess it's just a hard problem given Twitter's environment.

Its strength is freshness and volume, but I guess these can be achieved without Twitter if you have a strong web crawling infrastructure? Also, the current generation of LLM is not really capable of exploiting minute-level freshness... at least for now.