Show HN: RoboTrump – I made a Trump text generator that is fully convincing | Hacker News Reader

Show HN: RoboTrump – I made a Trump text generator that is fully convincing

lawsuit.org

33 pointskristintynski27 comments

This is basically useless without a place where you can use it to generate new random bodies of text on the fly. It's easy to cherry pick 10 good samples out of hundreds with the intention of attempting to fool people - maybe not so much if they're generated on the fly.

It's also kind of useless if you don't compare it to other subjects: maybe Trump is easier to imitate than, say, Obama.

He almost certainly is - he doesn't have nearly the same gift for public speaking that Obama did. As pointed out elsewhere in this discussion, his speech patterns are pretty simple and repetitive, and thus ideal for RNN training.

It feels like maybe your bias for Obama is clouding your judgement for computers to discover and reproduce patterns in general. Like, what evidence do you have that Obama's speech patterns are easier or harder for ML to discover and reproduce patterns? And if his patterns are harder, what evidence do you have that the increase in complexity is meaningful?

Here's some analysis that shows Trump comes lowest in complexity among the last 15 presidents that were analyzed.

https://www.newsweek.com/trump-fire-and-fury-smart-genius-ob...

>The words were run through a variety of lexicological analyses, besides the Flesch-Kincaid, and the results were the same. In every one, Trump came in dead last. Trump also uses the fewest "unique words" (2,605) of any president—Obama was the best at 4,869—and uses words with the fewest average syllables, with 1.33 per word, compared to positively multi-syllabic president Hoover at 1.57.

>"By every metric and methodology tested, Donald Trump's vocabulary and grammatical structure is significantly more simple, and less diverse, than any President since Herbert Hoover, when measuring "off-script" words, that is, words far less likely to have been written in advance for the speaker," Factba.se CEO Bill Frischling wrote. "The gap between Trump and the next closest president ... is larger than any other gap using Flesch-Kincaid. Statistically speaking, there is a significant gap."

Simple speech is often less deceptive. A surprising amount of political speach is deliberately dualistic. By any measure, President Trump actually does what he has said he will attempt to do. Bigly.

Trump's speeches/sentences can be twisted either way you want and nobody can't tell if it's just gibberish or real statements so the computer has fewer chances to produce "bad" samples. Just like a broken record!

Again, this all makes sense anecdotally, but how can you measure how significant this to produce believable text?

Maybe a more diplomatic description could be to liken it to modern poetry.

It's more about conveying a feeling than following strict rules of grammar or composition.

As shown in sibling comments, it's pretty easily picked up on both statistically and just by reading their speeches that his are simpler. I'm not really a ML expert, so I can't necessarily comment to that making it easier to make a Trumpbot, but my gut reaction is that it would if for no other reason that this sort of text works better for generators like Markov chains.

For what it's worth, I don't like Obama - but it would be churlish to deny he's a top-tier orator.

I have actually done what they have done in the paper. The primary difficulty that I had is that when attempting to fool people they would basically believe nearly anything that the system would say.

Its relatively easy to reproduce see http://trumptwitterarchive.com/archive and then you just follow the directions on https://github.com/minimaxir/gpt-2-simple

how about with this > https://TrumpTweets.io ;-)

Next steps: Link it to those NN video and voice generators, stream endless trump on twitch.

that is the next logical step in this fake news dystopia isnt it. Seems like someone could relatively easily set up a pipeline to generate text, turn text into Trump's voice, then apply to Deepfake video, churn out millions of videos... profit. kill democracy. etc.

Discussion from yesterday:

https://news.ycombinator.com/item?id=21315699

Is this like when the Japanese created Mechagodzilla to defeat Godzilla? Are the Democrats' best hope for 2020 Mechatrump?!

I scored 8/10, and the two I got wrong were both successes of 'RoboTrump' convincing me. In the other cases, it fell down by not having the characteristic short sentences and self-interruptions of real Trump -- so real, people don't believe how real -- speech.

This is remarkable, but is there a way for us to access this Trump text generator without lawsuit.org's curation? I'd love to pull back the curtain and see how many "tries" it takes to extract a convincing quote.

Edit: honest question, why are people downvoting this comment?

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.

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

The thing about Donald Trump is this: Every time he utters a word or Tweets something, the media goes nuts and publicizes it. It seems completely crazy.

But remember: Donald Trump was never a politician before running for President as his first run. He was totally outgunned by several well-funded Republican rivals. He was totally under-funded compared to Hillary Clinton, who also had huge advantages in media and campaign machinery.

Yet Trump won. With his goofy, outrageous, cant-ignore-them tweets and statements. He won, 'bigly'. (Haven't heard that one in a while, huh?)

Some day, he'll be remembered as one of the most shrewd self-promoters in history. Deny it at your peril.

Reply on news.ycombinator.com