Hacker News Reader

Comment by kristintynski | Hacker News Reader

I can upload the trained model if you are curious, but it was straightforward to do. I just used the Salesforce CTRL model and fine-tuned it on a large corpus of Trump speeches.

View on HN

I didn't really cherry pick, most everything it generates is of this quality.

I can't help but feel that the real text would not be so difficult to guess if you did a Robo Carl Sagan or Robo Ursula LeGuin or Robo almost anyone whose English communication skills would be considered formidable.

I suspect that because of the nature of his speech (repetitive, shallow, hyperbolic) he is easier to simulate than someone who communicates in other ways. I would love to see the same system used to simulate some of the people you mentioned (and others).

When it's trivial to build fake people with distinctive styles, though, bots will probably be even more effective. I think that may be a bigger risk than deepfakes.

I think that if you're building Bots that can simulate LeGuin you're getting awful close to XKCD mission accomplished territory though.

If it were able to be used via the web, that would be fantastic. In your experience, were most of the quotes as convincing as the ones in the article, or were there some duds to filter out?

Edit: never mind, you answered this question elsewhere in the thread :)

As long as I stuck to prompts related to politics and things Trump has talked about, it was scary good. I'd imagine fine-tuning on a much larger Trump dataset would continue to improve it.

That's fascinating. Now I want to hear Trump talk about spaces vs. tabs

Reply on news.ycombinator.com