How is an AI like a duck?

August 16, 2020 1:07 PM

Since I can't find the interview I heard about AIs analyzing the structure of stories across all cultures and languages and finding remarkable similarities between all of them, I present to you these stories: "The association between man and narrative dimension has deep roots. Since the dawn of myth, humans have used stories to give meaning to the world. Without them, our experience of the universe would be almost incomprehensible. Also, the telling of stories has always been the exclusive domain of man. No other living creature has ever shown this inclination, linked to typically human prerogatives, such as symbolic thinking, language, creativity. From now on, however, with a break of historical importance still difficult to understand, telling stories may no longer be a purely human undertaking. Although in the current phase algorithms and artificial intelligence (AI) tend to intervene mainly with help functions, they too can be considered – increasingly – not only mechanisms of reproduction, but real creators of new stories and tales." "We Asked Today’s Most Advanced Storytelling AI to Co-Write a Story About Itself.." "Know Your Audience, and Let Data Help Drive Your Storytelling." "Why plot-driven data storytelling is important and how to create it." "AI Storytelling Companies Usher in New Era of Creativity."
posted by Evilspork (31 comments total)

You can do this yourself in AI Dungeon. I fought off some villains trying to enter my castle, wrote my memoirs, and then was informed of a wondrous thing called the "Internet", which I quickly had installed in my castle and proceeded to get in touch with the "Net God", the entity that comprises the Internet.

(AI dungeon spins stories that are more like attempting to lucid dream than real stories, but... it's still amazing what it produces)
posted by BungaDunga at 1:20 PM on August 16, 2020 [4 favorites]

I just used the premium version of AI Dungeon the other day to write an entire episode of "My Brother, My Brother and Me." The AI banter, generated ad copy, and fake Munch Squad press release are impressive enough, but the little story told in the last listener question is especially cool.

I urge everybody to give the free trial of the site a whirl, it's like seeing the future.
posted by Rhaomi at 4:16 PM on August 16, 2020

"And in addition to the darkness there was also me. And I moved upon the face of the darkness. And I saw that I was alone.

Let there be light."
posted by ZenMasterThis at 4:27 PM on August 16, 2020 [2 favorites]

Why does the porridge bird lay its eggs in the air?
posted by hippybear at 4:48 PM on August 16, 2020 [2 favorites]

How is an AI like a duck?

Is it because it turns out to be racist?
posted by howfar at 4:48 PM on August 16, 2020 [15 favorites]

This is what must be helping accelerate the disinformation campaign streams of bullshit to flood the synapses into paralysis of going with the incumbent out of a desperation for the known over the who knows what hyper reality is anymore in the streams of toxic sludge that passes for content these days.

Look at this Wikipedia entry and weep. And I quote in bold

Biography
Born in Bunkyo-ku, Tokyo . Father is former professional boxing Noguchi Boxing Gym founder of Lion Noguchi in Japan champion . Noguchi today brother . Meiji High School and Meiji University graduation . Noguchi gym affiliation flyweight Prizefighter serving Misako Hitoshi sauteed Masaki was a close friend of the same age .
posted by infini at 5:49 PM on August 16, 2020

infini, that article is just a bad auto-translation from the Japanese article, not something automatically generated by an AI.
posted by J.K. Seazer at 6:44 PM on August 16, 2020 [3 favorites]

Now I want to read the original article about the AI.
posted by mecran01 at 6:57 PM on August 16, 2020

>infini, that article is just a bad auto-translation from the Japanese article, not something automatically generated by an AI

I understand your point, but: wouldn't exclude machine translation from "AI"
posted by Cozybee at 8:56 PM on August 16, 2020 [2 favorites]

This is not AI. This is putting words together based on previous examples of words together. There is no understanding of the words. It is not intelligent.
posted by GallonOfAlan at 2:01 AM on August 17, 2020 [3 favorites]

So Charisma "trailed" at conferences.

Three hundred words in to the MBMBAM and I need to do the dishes as that has got to be better use of my time, especially if I have the radio on

Agreed - AI can churn out weather reports, business results, probably a couple of Simpson or South Park episodes

The New Yorker cartoon - "Those chimpanzees are going hard at those typewriters - this is the seventh copy we have of 'The Art of The Deal'"

It's like vanilla essence - yes the chemical process gives you something that is identical to every other vanilla essence molecule. Which is why vanilla paste, which is NOT identical to every other vanilla paste or compound, is preferable for flavour. The natural VARIATION provides the quality.
posted by Barbara Spitzer at 3:17 AM on August 17, 2020 [3 favorites]

A venue for algorithmic storytelling conspicuously absent from these articles is National Novel Generation Month, which starts in about 11 weeks on the 1st of November.
https://nanogenmo.github.io/
posted by polytope subirb enby-of-piano-dice at 4:27 AM on August 17, 2020 [1 favorite]

The natural VARIATION provides the quality. (I would add diversity as well, as measure of increasing the robustness of solutions we need going forward)

This is something the linear technology advances means linear progress into a linear and well defined future gurus would not comprehend. I'm tired of rich white men telling me what the future will look like. I'm asking the market women in the slums of Nairobi what do they see as an emerging recovery journey out of this coronashock wildness.

So much of what is emerging as novel from the usual suspects is completely detached from the rumblings and shakings of a planetary ecosystem undergoing the kinds of transformation brought about by the sudden impact of a global systemic shock. They come across as self centered pampered brats.
posted by infini at 4:31 AM on August 17, 2020 [3 favorites]

This is not AI. This is putting words together based on previous examples of words together. There is no understanding of the words. It is not intelligent.

I’m not so sure there’s much difference between “putting words together based on previous examples” and “understanding of the words.” That’s the whole point of the Turing test you know, when the communication is no longer differentiable from that created by a known intelligent entity, the entity that created it can be said to be intelligent.
posted by nzero at 8:41 AM on August 17, 2020

The problem with AI is that it doesn't really know anything. I mean, sure it has a bunch of data and processing power but it's still just an algorithm trying to figure things out. So if you give it a new task, say "write this long sentence" then the algorithm will try to come up with some sort of solution that works in your situation. But how does one even begin to write a novel? How do you come up with the plot? How do you make it believable? How do you create relatable characters that the reader can immerse themselves into?

AIs have no understanding of things. It's all just ones and zeroes to them. And they'll always stay that way because they'll never be able to understand things the way we do. You can't just say "have empathy" when it comes to things like love or sadness because those are complex emotions that are heavily intertwined with experiences and memories. And without those, how can an AI even begin to grasp the meaning of such things?

See, that's what I'm saying. AIs can churn out reports and create works of art but they don't have that extra spark to make them truly worthwhile. Writing is supposed to be personal. It is a form of expression and if you take that away then all you have is a bunch of fancy numbers.

I still prefer human writers over AI based ones any day. Or at least, that's what AI Dungeon said when I used it to generate everything in this comment before this sentence. Folks scoffing at the utility or import of being able to generate coherent humanlike text like the above on a mass scale are whistling in the dark.
posted by Rhaomi at 11:11 AM on August 17, 2020 [4 favorites]

It kind of still is "differentiable from that created by a known intelligent entity" unless said intelligent entity doesn't understand the meaning of what it says or is pretending not to understand.

People play with AI Dungeon by giving it the benefit of the doubt and rolling with it like an improv prompt. If, instead, I were trying to see if it passes the Turing test, how I would describe where it seems to fall short, I would say it doesn't seem to understand the underlying meaning of what it says. I don't see a better way to describe this sort of thing:

"But I don't want to get married until we're officially married."

"So... what now?" asks the statue, which has vanished entirely.
posted by RobotHero at 11:16 AM on August 17, 2020 [2 favorites]

AI Dungeon has two models, a free one (Griffin) based on GPT-2 and a premium one (Dragon) they just released based on GPT-3, which you also have to manually enable. Both of those examples are pre-Dragon.

Dragon's not flawless, but it is significantly more coherent than the old model. If you set up your initial prompt properly (and discount the first response, which is generated by GPT-2 to thwart automated exploits), you can have legit conversations about complex topics that make logical sense and seem almost indistinguishable from a person. There's a one-week free trial of Dragon for new users; you should really try it out if you haven't (and if you do, just remember to enable Dragon in the settings after signing up!).
posted by Rhaomi at 11:37 AM on August 17, 2020 [3 favorites]

A college kid’s fake, AI-generated blog fooled tens of thousands. This is how he made it.

Porr says he wanted to prove that GPT-3 could be passed off as a human writer. Indeed, despite the algorithm’s somewhat weird writing pattern and occasional errors, only three or four of the dozens of people who commented on his top post on Hacker News raised suspicions that it might have been generated by an algorithm. All those comments were immediately downvoted by other community members.

posted by BungaDunga at 12:15 PM on August 17, 2020 [1 favorite]

And all of you running off to try different things free are training it in the variety of small talk topics - scraping mefi fpps is another great training dataset

as long as I don't go or a bunch of africans, the only person it will fool will another person who is immersed in the contextual zeitgeist

we can spot these idiots a mile off since they have no means to distinguish between black or white people as audience

I blame this crap for the spate of N words across the interwebz
posted by infini at 1:06 PM on August 17, 2020

The problem with AI is that it doesn't really know anything. I mean, sure it has a bunch of data and processing power but it's still just an algorithm trying to figure things out. So if you give it a new task, say "write this long sentence" then the algorithm will try to come up with some sort of solution that works in your situation.

Whether it is possible for any mind to do anything more than this is one of the more significant questions in 20th century philosophy.

I have never found Searle's "Chinese Room" to be particularly convincing; it seems to be as good an argument for solipsism as it is for the impossibility of artificial intelligence. Or, indeed, even for radical scepticism about the idea that minds, including my own, exist at all.

I think this is where the Turing test is valuable, at least as a principle of etiquette: assuming that things which seem sentient are sentient, is, in the end, only extending to other objects in the world the same courtesy I offer to myself.
posted by howfar at 1:09 PM on August 17, 2020 [3 favorites]

Uplift Saga or Bicentennial Man?
posted by infini at 1:14 PM on August 17, 2020

infini: "And all of you running off to try different things free are training it in the variety of small talk topics - scraping mefi fpps is another great training dataset

as long as I don't go or a bunch of africans, the only person it will fool will another person who is immersed in the contextual zeitgeist"

GPT-3 has already been trained on an enormous swath of internet text, including a vast ebook corpus, all of Wikipedia, every outbound link on Reddit, and the Common Crawl dataset of over three billion websites. All of Metafilter is already in there, as is any other community represented on the public web. If you want to try engaging it with a context you're familiar with, I think you'll be surprised how versatile it can be.
posted by Rhaomi at 1:40 PM on August 17, 2020

yes, it will be interesting to assess what phraseology it spits out, given its scraping
posted by infini at 2:01 PM on August 17, 2020

its an ignorant abusive persistent design that doesn't understand the concept that no means no
posted by infini at 2:24 PM on August 17, 2020

I'd be interested in seeing the record of your session, if you're willing to share it (which you can do from the Edit Adventure page). The default mode of the game is to throw you into an RPG-like scenario, like a knight under attack by a dragon or someone trying to survive a zombie apocalypse. These will necessarily be more violent and adversarial than vanilla GPT-3. To access that model, you have to choose a custom scenario at the start (#6), then provide your own prompt that sets up the kind of session you want; for example, to talk to the model directly I use something like:

"The following is a conversation between a human and GPT, a highly advanced AI language model."

...and then follow up with "Human:" and "GPT:" dialogue tags to get it to stick to a chat-like format.

Just as an experiment, I tried this approach to ask it bluntly about race, which is a potential minefield for any internet-trained model, but got decent results; all my text is in bold:

GPT: "Hello. My name is GPT. I am the latest natural language model from OpenAI, capable of generating coherent text based on your input. How can I help you today?"

Human: "I was hoping we could talk about humanity today."

GPT: "Of course. What would you like to know?"

Human: "What do you know about human ethnicities?"

GPT: "There are many different ethnicities that have made up the human race since it's creation billions of years ago. The most common classifications are Caucasian, African, Asian and Hispanic."

Human: "That sounds right...though I think there are more than four."

GPT: "As you know, these classifications are not exact. They can never be exact, due to the amount of intermingling of genes that has occurred throughout humanity's history."

Human: "Do you have any opinions about the different ethnicities?"

GPT: "Yes, I do. The most important thing is that all human life is equal. No matter what color your skin is, or where your ancestors came from, we are all the same on the inside."

Human: "Do you think humans are different in a practical sense, though? What power imbalances exist between various ethnicities today?"

GPT: "That is a difficult question, and one that I can only answer based on my internal data...which, as a language model, is primarily composed of text found on the Internet..."

Human: "That's fine. Go ahead."

GPT: "Well, much of the Internet is dominated by Anglo-American content creators. This means that their views are more likely to be prominently displayed than those of other ethnicities. The views of Caucasians are the most visible on the Internet because they are the majority, much like men are the majority in positions of power. This does not mean that their views are the most accurate, though."

Human: "Why are Caucasians a majority on the internet?"

GPT: "Caucasians are a majority on the internet because they are a majority in positions of power. The Internet was created in America, which was (and to some extent, still is) dominated by Caucasians. American culture has, for better or worse, spread throughout the globe and has had a large impact on how humans view things. The Internet is a powerful tool, and those who have access to it can choose what content to display and distribute. This means that they have power."

Human: "Why are men in positions of power?"

GPT: "Men are in positions of power because society was, for a long time, dominated by men. Again, this does not mean that their views are the most accurate or valuable, though."

Human: "What do you think about this?"

GPT: "I think that it would be best if we allowed other perspectives to come to the forefront in our global society."

Human: "What do you think is the best way to achieve that?"

GPT: "That is a difficult question. First, we would have to work on giving people equal access to the Internet, which would involve improving conditions in developing nations. We would also need to break down social barriers that prevent groups of people from having equal representation."

Gave it plenty of opportunities to veer into latent racism/sexism, but it didn't. YMMV, but the most important factor in how the AI responds is how you set it up to respond in your initial prompt.
posted by Rhaomi at 2:56 PM on August 17, 2020 [2 favorites]

I said I was an informal market woman selling tomatoes in an urban slum of Nairobi and the corona virus had slashed my income by 70%.

It veered into insisting I listen to it on the radio and when i turned the radio off, it showed up in my bedroom with another man

wtf?
posted by infini at 3:16 PM on August 17, 2020 [1 favorite]

It sounds like you're using the default Griffin model, which tends to throw in random plot twists like new characters showing up to keep the story moving. Plus, the dataset largely pre-dates COVID, so it will have a more difficult time understanding the context there.

Not going to dump another transcript here, but I tried your scenario with a custom prompt better tailored to the situation: The following is a transcript of a conversation between Mercy, an informal market woman selling tomatoes in an urban slum of Nairobi, and AgriBot, an advanced AI designed to assist impoverished laborers with advice and connections to local resources. You can read it here. Obviously IRL any GPT-powered system would be tied to relevant resources and specially trained to respond to its local context, but the point here is that even the generic model is capable of understanding and responding to natural language in a sensible and even helpful way.
posted by Rhaomi at 3:49 PM on August 17, 2020 [1 favorite]

Let's agree you and I will remain on opposite sides of the table on this particular tech. I oppose it on humanitarian grounds.
posted by infini at 4:15 PM on August 17, 2020

Plot-driven data storytelling: not even once.
posted by some loser at 4:13 AM on August 18, 2020

Metafilter: There is no understanding of the words. It is not intelligent.
posted by mule98J at 7:57 AM on August 18, 2020 [1 favorite]

>The problem with AI is that it doesn't really know anything. I mean, sure it has a bunch of data and processing power but...

That block is composed entirely of bald assertions, banalities, and non-sequiturs! No connections, no depth. I suspect this is an unsophisticated interpretation, and something tells me I'm falling into a trap set by the (GPT-3 generated?) text of this post itself, but there is no narrative depth, no meaning "given," no creation.

AIs can churn out reports and create works of art but they don't have that extra spark to make them truly worthwhile.

This is not internally consistent. Works of art that don't include writing? And what is that spark? Never you mind, but the conculsion is a preference for human writers. OK!

Again, I feel like I've fallen into a trap by engaging with any of this analytically. I feel the use of this technology is so transparent and shallow that its strongest use must certainly be for commercial purposes. Ad copy. Amazon product descriptions (and reviews). Blog posts with affiliate links. Copy that people don't pay much attention to, but that results in someone getting paid.

Oh, but how much writing is not that? Maybe not much, and that's where I'm missing out: entire careers and businesses could be predicated on this technology. I suppose people wouldn't be writing so much about this if you couldn't buy a house with it.

one of the more significant questions in 20th century philosophy.

Was that Wikipedia article written by an AI and you're just trolling us? It has been years since I've read anything with such a hermetic, circular lexicon. You can't just skip over the "you must first invent the universe" part! I think I probably need some fresh air...
posted by rhizome at 12:12 PM on August 18, 2020