AI dictionary, from cybertriumph to wringie, and beyond!
December 2, 2019 7:30 AM   Subscribe

AI Dictionary is a Twitter bot that tweets a . I wanted to see how much OpenAI's language model actually knew, so I tried to get it to define words... but I accidentally had it set to 'random', not 'best'. [via mefi projects] Semi-related: OpenAI’s GPT-2: the model, the hype, and the controversy (Towards Data Science), and Experimenting with OpenAI’s Improved Language Model (short post on Medium), which notes "The public at large will need to become more skeptical of the content they consume online."

Previously: Dangerous But Not Unbearably So, in which Optical scientist Janelle Shane has helped demonstrate the inevitability of our AI dominated dystopian future by using the (scarily plausible) GPT-2 model from Open AI to show that AI tends to prefer names that sound like ships from Iain M Banks novels.
posted by filthy light thief (28 comments total) 25 users marked this as a favorite
 
Human error: I was going to note that this Twitter bot tweets once a day, more or less, but I got distracted. This was not a problem with the machine(s) (yet).
posted by filthy light thief at 7:31 AM on December 2 [2 favorites]


Wrong setting: MetaFilter, n. An online community of intellectuals, wizards, and accountants, who care more about the denotation of specific cultural concepts than subjectivity in general.

AMAZING.
posted by Foci for Analysis at 7:37 AM on December 2 [4 favorites]


I assume they output a few dozen a day and pick the best one to tweet, but even so, these are REALLY REALLY on point.
posted by rikschell at 7:41 AM on December 2 [2 favorites]


This is indeed the English to which I subscribe and practice daily...
posted by jim in austin at 8:02 AM on December 2


I assume they output a few dozen a day and pick the best one to tweet
That's also my assumption, but I'd have guessed only a few runs rather than a few dozen.

From Talk To Transformer using the full-sized GPT-2 model, my very first run with the prompt "Metafilter, n. ":
Metafilter, n. forum. One of the Internet's largest source of news, entertainment and original content, 2/3 of which appears to be misogynist propaganda. Incidentally, the site does not provide a search function for articles containing the words "rape", "rape culture" or "rape apologia". If you click on an article and that thing you are looking for is a feminist website, that is a feminist website. If you cannot see that text, do not waste your time here.
Second run was irrelevant, but the third run started out:
Metafilter, n. The place where shit hits the fan in the Internet's geekdom.
And the fourth:
Metafilter, n. the discussion board with the top active writers and commenters in the social media universe.
So whether we want ad copy for Buzzfeed or for the Wall Street Journal, I guess problem solved?
posted by roystgnr at 8:06 AM on December 2 [2 favorites]


So I played around with Talk to Transformer a bit and got this:

Metafilter, n.: A site where you can write things you find wrong, ask questions you don't have an answer for, or get advice on things you would not ordinarily want to.


Which, you know, is pretty decent. However, as part of that autogeneration, it gave me a definition for Aerospace Enthusiast:

Aerospace Enthusiast, n.: A people who have been exposed to a set of terms and strategies for thinking about the space industry that they are more or less skilled at, who make it their job to tell you if you are doing something stupid in this or that area, or if you are just making shit up as you go. See also Space Cowboy.

Which seemed great.
posted by nubs at 8:15 AM on December 2 [4 favorites]


And just now, these two:

Metafilter, n.: A place for everybody to have something nice to say.

Mantra, n.: A vague statement which can serve as a catalyst for the power of Thought.


Ok, that's enough - I can lose all day to this kind of thing.
posted by nubs at 8:17 AM on December 2 [1 favorite]




It's only a matter of time before someone trains a GPT-2 model on metafilter comments and gets it to learn individual poster's pet topics and idiosyncratic expressions. I'm not kidding--I've been playing with fine-tuning GPT-2 on IRC logs, and as I posted on twitter:
There's an IRC channel I've been on for about 15 years with basically the same group of people. We have 8 years of logs which I'm using to train a GPT-2 model, and it has learned our in-jokes, the particular way this group riffs on words, and our pet topics of conversation.

It's an uncanny mirror world where mirror-us are conversing imperfectly but recognizably like us. Not just individuals, but 10-20 lines of back and forth. It even mimics the channel's markov chain-based bot's particular brand of nonsense (sorry, @xashbot).

It generates fake URLs that we paste to each other and then half a page is us talking about and reacting to the topic of the article. e.g. http://www.theonion.com/articles/time-travel-in-back-and-forth-on,31691 … followed by us talking about time travel (except we apparently took The Onion seriously).
posted by jjwiseman at 8:37 AM on December 2 [5 favorites]


jjwiseman: ... except we apparently took The Onion seriously

Machine Learning bot, Twitter troll, or Conservative relative?

But seriously, this is definitely heading to the uncanny valley deepfake text, except the machine learning is now climbing out of the valley, and heading towards being good enough to fool people very often.

It's fun to joke about the times the machines get it wrong, but it's scary when we find a press conference from a president, which wasn't actually delivered (previously).
posted by filthy light thief at 8:49 AM on December 2


https://thegradient.pub/an-epidemic-of-ai-misinformation/

The current state of AI vs. the hype.
posted by Nancy Lebovitz at 9:12 AM on December 2 [4 favorites]


Until youse came in here with other GPT-2 stuff I was wondering if these were actually from an AI or if that was just an pretense to tweet funny human-written things. (Not that I have a problem with that either)
posted by aubilenon at 9:34 AM on December 2


I had a fun time generating dictionary definitions using Write With Transformer a few months ago, but only on the GPT-2 small model. The output was pretty random and silly, and didn't have a ton of bearing on the input word. Just looking at the definitions for MetaFilter on (what I assume is) the large model, I'm pretty blown away.
posted by little onion at 9:43 AM on December 2


Thank you, Nancy.

Between programmers and marketing, a whole lot of this stuff is WAY over-hyped. Programmers, especially moderately experienced ones, consistently overestimate the capability and reliability of their software. And marketing, of course, sells software that doesn’t even exist.

That said, if these programs do become sufficiently capable of fooling lots of folks, I wonder if the solution is to simply flood the internet with their products, until either the world descends into complete chaos, or people improve their critical thinking skills. I mean, this software, like encryption software and standards, is/will be open-source, open-standards and easily available, so it’s foolish to think you can control it.

On the up-side, Twitter and Facebook would probably just implode.

Metafilter, of course, would be left standing.

I am kidding, of course...

I think...
posted by shorstenbach at 9:47 AM on December 2


So I guess the cybertriumph will have to wait a few more years yet?
posted by rikschell at 9:54 AM on December 2 [1 favorite]


shorstenbach, the hype problem is caused by journalists as well as programmers and marketers.
posted by Nancy Lebovitz at 10:03 AM on December 2


unkwatch, n. the time each month during which it is customary to review your spelling, spelling etiquette, onomastics, and grammar and punctuation.

This seems like it should be a thing. Spelling etiquette and onomastics are of particular importance in the Internet age. Maybe just before or after the new moon?
posted by Not A Thing at 10:11 AM on December 2 [1 favorite]


Math teacher, n. a person skilled at making use of arithmetic and algebra. This small class is almost destroyed by the large class of pseudo-mathematicians, who enjoy using word roots and calculating the results of complex equations.
        — Ambrose Bierce rides again, this time on a cybernetic steed.
posted by Schadenfreude at 10:15 AM on December 2 [4 favorites]


"NOT HOT DOG"
posted by gimonca at 10:23 AM on December 2


fish, n. a person or thing which is not a fish, or which has a different shape, size, or shape than the fish.

Paging Douglas Hofstadter....
posted by Foosnark at 11:16 AM on December 2 [2 favorites]


To be clear, when I say that the GPT-2 model I trained "has learned our in-jokes, the particular way this group riffs on words, and our pet topics of conversation.", at least 50% of its output is boring or nonsense. But a fraction of it is clearly demonstrating novel creations relevant to the personalities of the people in the IRC channel, as expressed in text. It's not close to a 100% accurate simulation, but it is able to do things that other AI systems have not. The capabilities of these systems don't need to be exaggerated to be worthy of excitement.
posted by jjwiseman at 11:26 AM on December 2


Suppose it didn't have to be coherent over large stretches. A fake interjection that plausibly belongs to Joe in any correspondence, chat, or comment board will pass unnoticed, because we often express ourselves badly. With that, the evidentiary power of all text logs is moot. That is a huge change: like photoshop but for text. It's the small changes that are caustic, not the magic future in which the fuckers become sentient.
posted by stonepharisee at 12:54 PM on December 2 [1 favorite]


fish, n. a person or thing which is not a fish, or which has a different shape, size, or shape than the fish.

This seems like a paradox, but I think the definite article in the second clause sheds some light: in using the term "fish (1)," one refers to a particular fish, while "fish (2)" refers to all things (including other fish) that are not that particular fish.
posted by Not A Thing at 1:31 PM on December 2


KIRK: Have you ever been on Metafilter?
SPOCK:
No.
KIRK: Well, then I've got a lot to learn.
SPOCK: Starfleet is not about manners, it is about responsibility.
KIRK: A human is taught manners at a very young age. And then etiquette. And then decorum.
SPOCK: Duty.
KIRK: Being a gentleman is not about being mean. It's about being thoughtful.
SPOCK: When are you going to learn how to dress
posted by RobotVoodooPower at 2:27 PM on December 2


PICARD: Data, bring up all records on Metafilter.
DATA:
Over 25,000 posts have been posted thus far.
LAFORGE: Sorry, Commander, the bottom line is the level of comment is too high for the system to handle.

PICARD: Data, what is CAT-SCAN.COM and why do people wedge cats in scanners?
DATA:
In my belief, it was conceived to create a way to fulfill the inevitable conclusion of a generation of patients who'd suffered for more than twenty-five years, becoming life-forms incapable of sleep and unable to conduct complicated natural activities.
PICARD: If that's the case, there's no doubt in my mind that CAT-SCAN.COM is an appropriate technology. It may even be therapeutic.

- yes, this is actual GPT-2 output!
posted by RobotVoodooPower at 2:38 PM on December 2 [6 favorites]


I can't play with this at work, but I wish I knew its definition of "ezowowowowowowowoowo-woowowow"?
posted by klausman at 3:20 PM on December 2


ceiling, n. An object that keeps people from trying to look upward.
posted by floomp at 4:21 PM on December 2 [2 favorites]


Oh gosh, this is gloriously Ben Marcus: redefine words, create dystopia.

As a former dictionary typesetter, I compliment you on your layout. A real paper dictʸ might be a little more compressed to save paper, but this is really pleasing
posted by scruss at 2:09 PM on December 10 [1 favorite]


« Older a morbid fascination with their strange appearance   |   Live in Lesotho Newer »


You are not currently logged in. Log in or create a new account to post comments.