Catch up on stories from the past week (and beyond) at the Slashdot story archive

 


Forgot your password?
Close
typodupeerror
×
AI News Technology

OpenAI Has Published the Text-Generating AI it Said Was Too Dangerous To Share (theverge.com) 62

Posted by msmash from the interesting-changes dept.
The research lab OpenAI has released the full version of a text-generating AI system that experts warned could be used for malicious purposes. From a report: The institute originally announced the system, GPT-2, in February this year, but withheld the full version of the program out of fear it would be used to spread fake news, spam, and disinformation. Since then it's released smaller, less complex versions of GPT-2 and studied their reception. Others also replicated the work. In a blog post this week, OpenAI now says it's seen "no strong evidence of misuse" and has released the model in full.

GPT-2 is part of a new breed of text-generation systems that have impressed experts with their ability to generate coherent text from minimal prompts. The system was trained on eight million text documents scraped from the web and responds to text snippets supplied by users. Feed it a fake headline, for example, and it will write a news story; give it the first line of a poem and it'll supply a whole verse. It's tricky to convey exactly how good GPT-2's output is, but the model frequently produces eerily cogent writing that can often give the appearance of intelligence (though that's not to say what GPT-2 is doing involves anything we'd recognize as cognition).
This discussion has been archived. No new comments can be posted.

OpenAI Has Published the Text-Generating AI it Said Was Too Dangerous To Share More

OpenAI Has Published the Text-Generating AI it Said Was Too Dangerous To Share

Comments Filter:

  • Fake News (Score:4, Funny)

    by weilawei ( 897823 ) on Thursday November 07, 2019 @02:41PM (#59391550)

    They haven't actually released it; the story was just autogenerated with the GPT-2 tools.

    • They haven't actually released it; the story was just autogenerated with the GPT-2 tools.

      First we have last night's update causing a bug last February, and now we have the "too dangerous to release" AI chatting about its own release. This is Skynet, I tell you. Guys, I'm getting really scared.

    • They haven't actually released it; the story was just autogenerated with the GPT-2 tools.

      Yes, but it's Fake News You Can Trust [shopify.com]!

  • Pretentious (Score:3, Interesting)

    by 110010001000 ( 697113 ) on Thursday November 07, 2019 @02:43PM (#59391556) Homepage Journal

    Could "AI Researchers" get any more pretentious? Such geniuses. Yeah we get it, you wrote a program. Good job.

    • And, was it as scary as they claim, or was the whole thing just a marketing gimmick to generate interest?

  • Play With It Here (Score:4, Informative)

    by bartle ( 447377 ) on Thursday November 07, 2019 @02:52PM (#59391598) Homepage

    The article links to this location where a user an enter their own custom prompt and allow their "AI" to generate additional content. [talktotransformer.com]

    Enter a few custom prompts and you'll soon get a sense for its strength and weaknesses. It seems to do best when prompted about some current news - presumably it has a wealth of source material to draw on. It's far weaker when the prompt contains fictional proper names.

    • Pehaps, but has anyone really been far even as decided to use even go want to do look more like?
    • It's literally just playing Mad-Libs with Wiki data. I threw in "Ur-Quan are" and it vomited back a page from Memory Alpha about a Delta Quadrant species and replaced its name with Ur-Quan. It at least knew enough that Ur-Quan was sci-fi, and it didn't grab text from a Star Control wiki, but it certainly didn't generate anything.

    • Re:Play With It Here (Score:4, Funny)

      by aardvarkjoe ( 156801 ) on Thursday November 07, 2019 @05:17PM (#59392168)

      This one is worth sharing:

      Two men walk into a bar. One of them is a man named James. The other is James. The bartender is named James. They sit down. The bartender drinks, James laughs. They talk about nothing, about their day and their job.

      The bartender walks into James' bar. James says, "Hi. How are you? How are you?" The bartender replies, "I'm good, thanks." James goes away.

      The bartender goes into James' bar. James comes in. The bartender says, "How are you?" James says, "I'm fine." The bartender replies, "Yeah. I'm really enjoying myself. You, on the other hand, have been in a lot of trouble."

      James goes, "Who told you that?"

      The bartender says, "A friend of mine."

      James says, "I didn't say anything. You've got to come in and have a drink."

      The bartender comes in...

    • I am programming late tonight and will be writing about this as I go through it.

      This will most likely be my last blog post for a while. I'm not quite sure what to do with myself at the moment.

      In the mean time, I will do my best to answer as many of the questions I receive as I can. I'm hoping to do this weekly, maybe twice in the next week or two. I will try to answer as many questions as I can and will be doing so as much as I can. It's going to be a really busy week!

      Also, I'm hoping to write a more extensive blog post in the near future. It's something I've been wanting to do for a while, but haven't gotten around to.

      Totally out of context for the text I entered. That generic response could be used for ANYTHING. I am murdering late tonight and will be writing about this as I go through it.... I am eating late tonight and will be writing about this as I go through it... I am shopping late tonight and will be writing about this as I go through it.

    • The article links to this location where a user an enter their own custom prompt and allow their "AI" to generate additional content. [talktotransformer.com]

      Enter a few custom prompts and you'll soon get a sense for its strength and weaknesses. It seems to do best when prompted about some current news - presumably it has a wealth of source material to draw on. It's far weaker when the prompt contains fictional proper names.

      The danger is elsewhere - pervasive forum bots. I have noted 2-3 of those on the guardian lately posting request-response for the purposes of Brexit promotion. If building them will become a commodity, the whole Internet discussion culture will go to hell and Social Media will shortly follow.

    • It's total shit, like all chatbots.

  • 1) Go to site: https://talktotransformer.com/ [talktotransformer.com]
    2) Enter "Trump" and try it out
    3) Next, try "Clinton"

    Do the results seem a bit, I don't know, "one-sided?"

    • No, they don't? They simply seem to be pieces of actual news reporting. Like what I got:

      Trump's "fire and fury" response to North Korea's ICBM test.

      The missile launch, the first since Trump took office, comes as Trump continues to call for China to do more to rein in North Korea, a country that recently conducted its most powerful nuclear test to date.

      In a pair of tweets, Trump said he would "do what is necessary" in response to North Korea's "hostile actions" and vowed that "something will happen" as a result.

      He also warned North Korea that he would "totally destroy" the country if it threatened the U.S.

      Earlier Tuesday, U.S. Secretary of State Rex Tillerson was on Capitol Hill to discuss North Korea with congressional leaders and said he expects the United States and China will "have to take further actions" against the country.

      North Korea test-fired its most powerful missile yet on Tuesday, which officials said was capable of reaching anywhere

      ======

      Clinton, the former Secretary of State, the New York Times reports.

      The Clinton campaign will no doubt claim that the emails are the work of Russian hackers who stole them from Clinton's email server and leaked them to WikiLeaks to make a political point about the former Secretary of State's honesty.

      If they're the Russian's, what can the Clinton campaign say to explain why they have released them, after Clinton's campaign has been under continuous investigation by the FBI? If the emails are the work of Russian hackers, then why have the Russians released them now, after the FBI has cleared Clinton of criminal wrongdoing?

      Clinton's campaign says they're the work of Russian hackers, while the Russian government has denied it.

      So, what's going on?

      Not quite sure what you're seeing there.

      • Interesting. When I tried "Trump," it came back with two paragraphs full of loaded language about racism, fear-mongering, anti-muslim rants, etc. For "Clinton" it produced (literally) a small, inoffensive biography.

        I guess it's schizophrenic.

      • Re: (Score:2)

        by Qwertie ( 797303 )
        It produces a different result each time, sometimes pro-X, sometimes anti-X, sometimes neutral, but always fake news. I got a sort of ... conspiracy theory that Clinton threw the election for Trump?

        Clinton had no such luck. Clinton's attempts to reach Obama and his top advisers were rebuffed. She even sent a letter directly to Obama, which was intercepted by the White House before it was delivered. As for the DNC emails, Clinton had two options: to leak them herself or share them with the press. The former

    • Re: (Score:1)

      by elohssa ( 317266 )

      Interesting experiment. I was not able to duplicate your results.

      For Trump, I got a news article that he was down slightly in a recent poll.
      For Clinton, I got an article about drone strikes.

      I would judge both as being slightly negative. Overall balanced.

    • Re: (Score:2, Troll)

      by The-Ixian ( 168184 )

      Trump loves to grab women by the pussy. (Photo by Win McNamee/Getty Images) Getty Images

      Trump has also made some pretty shocking remarks about women.

      "I think I was elected because of women," he said in 2005 when asked about the number of female Republican voters he got in New York. "I get more votes than women."

      And during an MSNBC interview in April, he said that he was "very, very happy" with how women were treating Hillary Clinton in the debate. "The only thing she had going for her was her looks," he said. "She is a woman, and I don't think she has the looks."

      As for his comments about Alicia Machado, he's not exactly apologizing.

      Clinton loves to grab women by the pussy. You should watch her go after Carly Fiorina on Saturday because she went after Carly with a vengeance. Carly was a good person. You shouldn't vote for her. But she was an interesting candidate. She went to Washington as the first female secretary of state. I'm sure it was exciting for her, and it would have been for a lot of women, to see a female president. But you know what? That's a bad role model for girls. You know? You look at who our first female president was. Hillary Clinton.

      "She's been to Russia five times as secretary of state. I mean, come on. Come on."

      But what's not clear here, as Clinton makes clear in the video above, is that Trump is the one trying to grab women

      • Funny that I get moderated as a troll.

        All I wrote was "Trump loves to grab women by the" and "Clinton loves to grab women by the"

        The algorithm did the rest in both cases.

        It appears to roast both equally.

    • Re: (Score:2)

      by HiThere ( 15173 )

      If it's based on internet news stories, you should expect bias. How one-sided it would be and biased in which direction would depend on the particular sites it was using as sources a the moment.

      OTOH, I'm not sure I'd call that AI.

  • Try ' See how a modern neural network completes your', part of the first line from their own web site and see how bad it handles that.
    Mine told me to go to web site that was down and when I looked it up it had been closed because it had been used to spread malware

    • Re: (Score:2)

      by Pulzar ( 81031 )

      Try pasting the first paragraph of a current political news story. It generates all kinds of semi-believable fake news... In that context, it's both impressive and a bit worrying. The Kentucky democrat in my "story" was suspected of felonies, hiding money, receiving undocumented donations, and his wife was also hiding money on the side... each with a full background story at a click of a button. It's really easy to generate 30 of those, pick the best 3 or 4 and start spamming them out.

      It does a lot worse wi

  • Max Headroom (Score:3)

    by lobiusmoop ( 305328 ) on Thursday November 07, 2019 @03:07PM (#59391654) Homepage

    This needs to be hooked up to a digital virtual head generator and voice synth and it can generate hours of mindless daytime filler TV.

  • I tried it with the first sentence from the Verge article... the generated text is... pretty good. Although the fact that the AI came up with the idea that "if people in the future want to take actions they "should expect to be killed"." is somewhat... worrysome...

    The research lab OpenAI has released the full version of a text-generating AI system that experts warned could be used for malicious purposes.

    While there's no evidence of its being used to hack any system, OpenAI and its founder Greg Brockman are

  • First post slashdot was an obvious attempt to "get back to the future." The question is not why did they attempt to create an echo chamber. The question is why did they fail?

    I have no doubt that there is some sort of conspiracy in the comments. I do believe that the Slashdot staff knew something was wrong. But they made a poor choice of words and their management failed to communicate the seriousness of their situation.

    One final comment about the slashdot comment from one of my readers:

    "It's amazing how a p

  • There worried that is will spread fake news and misinformation? We had entire networks calling themselves news outlets for that. Are they worried that news anchors will be replaced? Is that why these news anchors are so supportive of UBI?

  • I just stopped consuming news, except for entertainment purposes.

    What's real is what I experience personally. Period. End of the line.
    The people and nature around me.
    What do I care what some government supposedly decided? What I care about is the cop right in front of me who actually influences my world, and what he believes is real. What I care about is the project that I finished and can now play with, sell, etc.
    Those are real.

    News is just shit. To manipulate you. Sure, you can assume its existence, as th

  • I got this! Pedophilia fiction. (Score:3, Interesting)

    by Nicopa ( 87617 ) <nico@lichtmaier.gmail@com> on Thursday November 07, 2019 @04:06PM (#59391916)

    I typed jokingly "Se likes sex. A lot" to see what this AI tool will generate. This is what I got:

    She likes sex. A lot. Like a lot a lot."

    The girl with the most sexual prowess is the one you see first: a little boy, perhaps, who looks at you with wide-eyed wonder when you come near him. His face lights up as you make eye contact. Your hands begin to wander down his body, and you feel yourself grow more and more excited by the little boy's every move. He is like an irresistible treasure you'd like to explore every way you can, but you've learned that it's better to let the little boys be the explorers, the ones who know what they're looking for and will gladly show you everything they have. They're a little safer than you, anyway. The older you get, the more comfortable you become with the idea of children. Your daughter is about to

    This is the danger of feeding your AI with random Internet texts. I can imagine this AI can also be used to generate other kind of ugly texts (racist, etc).

    • Re: (Score:2)

      by PPH ( 736903 )

      This is the danger of feeding your AI with random Internet texts.

      We really miss Tay over on /pol/.

  • AI brainstorming (Score:3)

    by Chromal ( 56550 ) on Thursday November 07, 2019 @05:04PM (#59392112)
    (When the velociraptors achieved faster-than-light travel, their empire began a period of rapid expansion to the stars.) This time, Velociraptors were able to travel to the distant star Azalea, which was nearly identical to the star known as the Antares, except for one thing: its location in the universe, which is not the same as our own. The Velociraptors settled here and took the name Velociraptor Prime. Over the course of a thousand years, the Velociraptor population grew. The Velociraptors conquered planets across the galaxy, and many of the worlds they conquered became their homes. But in a moment of arrogance, a few Velociraptors lost their lives. For the rest of their lives, the Velociraptor species kept a small piece of Azalea to honor their fallen brethren. (see also: https://talktotransformer.com/ [talktotransformer.com] )

    • Re: (Score:2)

      by Chromal ( 56550 )
      Researchers recently discovered that the Trump administration actually consume the souls of small children for sustenance. "Snuffing out the weak is great, two or three children a day is perfect!" according to one top official. "We don't have to do anything except feed them. They just eat." The children of America. Their souls are a food source for the administration. According to a new book, "Inside the White House," published Tuesday by Simon & Schuster, the White House is so obsessed with consuming

    • Re: (Score:2)

      by Kaenneth ( 82978 )

      But the future refused to change.

  • About ten times out of eleven, attempting to contact reCAPTCHA results in an error for me. And you need to contact it twice in a row to pass it, with the result that I basically cannot use anything protected by reCAPTCHA. I use Pale Moon, but in Chrome (which I keep around for emergencies) I don't even get a reCAPTCHA to begin with.

    • Try spoofing your user-agent. Of course, Google might be implementing some kind JS that works well in its V8 JS engine but not well in others.

    • Hey noobs, not being able to use the website that demos this thing because I can't reCAPTCHA is not offtopic.

      Why do they give noobs modpoints?

  • The ultimate answer to life, the universe and everything is

    This site may also have information on
    Our purpose

    The Universe is made up of
    Everything in it

    You are here because of your
    Wondering why?

    It's time to explore the
    Ultimate Question

    Why is there something rather than nothing?
    Everything in the Universe exists for a reason.

    The purpose of the universe is to bring us into balance.

    In other words, our purpose is to be better people, more like YOU.

    Your purpose in life is to be more like God.

    The purpose of God is to bring about a better life for everyone.

    What is the ultimate purpose of existence?
    There is no ultimate purpose.

    There are only
    Possible Purpose

    I'm a seeker of knowledge and the ultimate purpose of my existence is to be more like God, who is
    The greatest of all beings

    • Sponge Bob was an author...

      who wrote a book about the SpongeBob SquarePants universe, called SpongeBob's Spatula: The Great Flood, which was published in 2005. In it, he wrote that he wrote the book as an apology to those who had bought his original comic strip SpongeBob SquarePants and wished to find the author so that they could "come to terms with their misdeeds." However, he claims that the book is in reality "an official apology from the author to his readers."

      A sequel to the book was released in 2014.

  • This is essentially what journalists do. Take a very few facts and pad them out into a longer story, which has very little content beyond the original.

    But maybe they need not worry just yet. I put in "It will be hot tomorrow." thinking about the weather (it is hot here today). This is what I got. Maybe I'd like to read the paper that this too generates...

    It will be hot tomorrow, and I won't have a problem getting it to you if you want."

    "Oh, sure," I said, feeling a little giddy and happy to know I was

  • It actually will. People have trouble coming up with something given a vague prompt.

    So enter the topic, and see what it generates. If the first does not work, then try generating another. Give it a couple of goes, picking out interesting paragraphs from each one. Then glue them together.

  • New books by dead or inactive authors are what come to mind for me.

    LK

  • It's scary and dangerous for journalists because it's something that can significantly reduce demand for their jobs.

  • This reminds me of the board game, Balderdash, where you are given a topic (like a movie title, a person's name, or a name of an object, phenomenon, etc. - these are all the kinds that most people don't have a clue about), and all the players write the answer that they can make (if know the actual answer, they can write that, too, and get some points, but that happens very rarely). I wonder how good a player the AI would be - perhaps it could be included in the play!

  • I've become jaded about the extended family of chatbots and so-called "A.I." text generators. Not one of them has ever exhibited the slightest ability to generate context-aware content. I tested "Transformer" online just now and got pretty much the sort of gibberish I'd expected.

    [--------]

    Original text: "Reluctantly, the chipmunk ventured out from its den. A scent of roasted almonds wafted over the entrance."

    Generated text: "It paused a moment, before taking to the air. It paused, and then took flight. It w

Slashdot Top Deals

You are always doing something marginal when the boss drops by your desk.

Close