[adeeds]

I tried using the new 1558 model on the headlines of news articles and the results were surprisingly decent. Maybe 1/5 of the articles were plausible, and 1/5 were comically awful (Things got pretty zany when I plugged in an LA Times article titled "Yes, Elsa and Anna wear pants in 'Frozen 2.' And yes, it’s a big deal."). Of the remaining 60% of the articles, a lot could be made plausible with some light human editing. I also didn't fine-tune the model or tweak many parameters, which I'm assuming would lead to even better results.

I did notice that some types of articles tended to be easier to fake. A sports article asking a coach about his team's performance would end up spouting the same handful of platitudes, and some random sports statistics that sound realistic unless you look into them. Some articles that were more on the wonk side of things (lots of statistics and not a lot of context) also seemed to produce decent results, although I'm sure they're ridiculous if you're in expert in the subject matter of unemployment figures or auditing a city for overtime charges.