Its a hive of misinformation, disinformation and toxicity. Its succinct I guess, but nothing is eloquent or descriptive because of the character limit. And its full of repetitive "filler" information.
Who wants that in a foundational LLM dataset?
Maybe its OK for finding labeled images... But that still seems kidna iffy.
Now, this has been severely degraded by the changes that Musk has made. The spam in direct messages is off the charts now, whereas in the past I would get maybe a spam per year. And when one of my areas of interest has a post that gets popular, I have to scroll past all the insipid clout-chasing replies from blue check marks which get floated to the top of replies in an attempt to reward some of the worst people on the internet. Also the long form tweets that need to be expanded are a big deflation of user experience, as reading and replying to those are suboptimal compared to a tweet thread.
But this is also the general internet: 99% spam plus 1% quality. And the quality of the 1% of good Twitter is some of the very best of timer material out there.
And since LLMs have been trained on this same mix... they seem to be mostly good at filtering. But they do lie an awful lot.