I trained this by fine-tuning the Salesforce CTRL transformer model, then surveyed 1,000 to see if people could tell the difference between RoboTrump and real Trump. They can't.
This doesn't really surprise me. Trump's speech patterns are juvenile: lots of repeated words and phrases, simple vocabulary, almost no subtext or complex ideas. A perfect training set for a simple probablistic text generator. This would be more interesting trained on the data set of an better orator, maybe Obama, Reagan, or Kennedy?
That's terrifying. I got 70%, but really only because I'm an aficionado of these bonkers quotes and recognised some phrases.<p>I do think that Trump is kind of an easy target, here, as his speech and writing is fairly incoherent anyway, but it's a scary demo all the same.
I don’t get it. If the point is that the text is indistinguishable from real Trump speeches, this is because it says precisely the kinds of things that the actual Donald Trump says. What would be dangerous would be persuading people that he said things that are substantially different to what he really thinks or says.
Got 9/10 but it was a challenge - between the way real Trump mangles English and fake Trump strains to flagrantly fail the Turing test (h/t @JanelleCShane) it was a struggle - nice job, train it more, and hook it up to it's imaging equivalent.....then let them all sort that mess out . :-D