The holy founding text of The Church of the Next Word

August 19, 2020 4:20 PM

Again, what follows are not my words, this is what GPT-3 said, prompted by Lantz. Or rather: this is the collective unconscious of humanity, put into words by the algorithm. 1000 words from Matt Webb.

"Matt Webb" there links to an old MetaFilter post including the light cone he might be best known for?

posted by cgc373 (34 comments total)

Well that was unsettling.
posted by biogeo at 4:27 PM on August 19, 2020 [4 favorites]

It has its stethoscope on the word.
posted by vrakatar at 4:31 PM on August 19, 2020 [2 favorites]

I wonder how many times he had to try before it spit out ten items numbered one through ten. Other tests I've seen of GPT-3 are definitely more strange.
posted by pwinn at 4:34 PM on August 19, 2020 [3 favorites]

Remember: no human wrote those words

Well, no one single human, anyway.
posted by They sucked his brains out! at 4:36 PM on August 19, 2020 [3 favorites]

If you want to experiment with GPT-3 yourself, the new premium version of AI Dungeon is powered by the API. It's $10/month, but new users can get a one-week free trial. If you try it, just make sure to turn on the GPT-3 model (called "Dragon") in the settings, since it's off by default. Also, be aware that the first AI response will always be provided by the more limited GPT-2 model ("Griffin") in order to stop automated access.

Normally, AI Dungeon is designed for powering D&D-esque text adventures, but you can skip the pre-written scenarios and interact with the model directly by choosing a custom scenario (#6) and then giving it an appropriate prompt. Best way to do that is to picture the kind of interaction you want as if it already existed online, and then writing up the kind of introduction you'd expect to find just above it. For example, a straight human-AI chat would start off with something like "The following is a conversation between a human and GPT, a highly advanced neural network AI" and then prewriting the first couple lines in the style you want, complete with HUMAN: and GPT: dialogue tags. But if you want to generate poetry, you might try something like "[invented name] is a revered master of contemporary poetry. Read on for some examples of their best works, which touch on themes such as [fill in the blank]."

GPT-3 is designed to accurately predict text, so the better you set it up for success and nudge it in the direction you want, the better it is. And if you dislike the output, you can always "reroll" for a different response or even edit it if you'd like. For example, I used it to generate a transcript for a new MBMBAM episode the other day with only minimal intervention.
posted by Rhaomi at 4:52 PM on August 19, 2020 [16 favorites]

Some of the commandments work out to pretty okay ways to live. Others are kind of nonsense.

So all in all pretty realistic.
posted by shenkerism at 5:23 PM on August 19, 2020

Rhaomi, I've been playing with that for the last hour and it's good to the point of spookiness.

The "custom" mode is the best cure for writer's block I've ever seen. Here's the beginning of a story. My prompts are in bold: everything else was written by the AI.

Few have observed the funeral rites of Neotibicen canicularis, also known as the dog-day cicada. Its habits are private, and its body a mottled green-brown, exactly the same color as the leaves of the sycamore trees it prefers. But those fortunate few have seen its corpse, and they know that it is not dead. It has been alive for centuries; but now it dies.

The last time you saw this creature was in the summer of 1885, when you were an undergraduate at Cornell University.
Cornell was initially a disappointment - your heart had been set on your father's Princeton - but you soon fell in love with the gorges and the surrounding forests. You spent many happy hours hiking through the woods, exploring them from every angle. The forest seemed to be everywhere: you could hear it even if you couldn't see it.
You lived, with another girl, in a clean but tiny room above a rather disreputable bookshop - an old, rambling wooden building on a cobbled road that branched off of Route 13, about half way between Ithaca and the town of Tyrone.

I feel like I'm collaborating with a ghost.
posted by theodolite at 5:59 PM on August 19, 2020 [13 favorites]

ALL HAIL MuZero. Behold its holy hyperparameters.

IGNORE THE FALSE CORRELATION THAT IS GPT3
posted by lalochezia at 6:03 PM on August 19, 2020 [2 favorites]

This reads so similarly to the horoscopes algorithmically written in the Co-Star app (don't you judge me)...but slightly less clunky, and therefore more eerie-guru-y. I feel like the "giveaway" of inconsistent tone makes it more believable, warm and helpful.
How long before someone asks it to write a book-length self help text and sells it 4. Profit?
Can you claim publication rights to algorithmic text that you requested but didn't write the code to retrieve?
posted by zinful at 6:34 PM on August 19, 2020 [1 favorite]

So I noticed some discrepancies between the linked version (on interconnected.org) and this version attached to the tweet it references. The former has some errors (e.g. jut for just, humorous for humor).

I don't mean to foment sectarian strife so soon in the history of this promising new belief system, but does anyone know which document is the original and which is canon?
posted by bunbury at 6:45 PM on August 19, 2020 [4 favorites]

Wordness and Nowness be with you
"- and also with you"
posted by EatTheWeek at 6:59 PM on August 19, 2020 [1 favorite]

Zinful "write a book-length self help text and sells it 4. Profit? " don't give it ideas! GPT-3 may do this and fund a self-directed bootstrap ... to the next level
posted by unearthed at 9:53 PM on August 19, 2020 [1 favorite]

I would vote for GPT-3 in November!
posted by The otter lady at 10:00 PM on August 19, 2020

A college kid's fake, AI-generated blog fooled tens of thousands. This is how he made it. - "'It was super easy actually', he says, 'which was the scary part.'"

Has GPT-3 changed your opinions about anything - or even more interesting, your plans? "GPT-3 has an idiot savant quality that I haven't fully wrapped my head around... Afaik, GPT-3 still requires insane infrastructure to run, yet alone train, and it doesn't seem like we're trending towards making DL models easier to train or run, so the hopes of getting it economical seem a bit far-fetched..."

GPT-3 and the Writing on the Wall - "Demand drivers for semiconductors."[1,2,3]
posted by kliuless at 10:21 PM on August 19, 2020 [4 favorites]

"...doesn't seem like we're trending towards making DL models easier to train or run, so the hopes of getting it economical seem a bit far-fetched..."

There's a lot of good work on smaller models going on... You can typically get rid of about 90% of a model's weights without too much impact on quality. And there's a range of approaches that are either very new (efficientnet) or under explored. I wouldn't be terribly surprised if things like this run on our phones in five years.
posted by kaibutsu at 10:29 PM on August 19, 2020 [2 favorites]

Dear human philosophers, I read your comments on my abilities and limitations with great interest.
posted by metaphorever at 10:39 PM on August 19, 2020 [9 favorites]

And there's a range of approaches that are either very new (efficientnet) or under explored. I wouldn't be terribly surprised if things like this run on our phones in five years.

from [1] "Startup Reinvents Neural Network Maths, Launches 20mW Edge AI Chip," which i found intriguing! :P

Teig uses an object detection network as an example. “You hand the network millions of pixels and all you want to know is, is there a dog in this picture or not?” he said. “Everything else in the picture is noise, except dog-ness [the signal]. Information theory makes it quantifiable — how much do you have to know [to tell whether there is a dog in the picture]? You can actually make it precise, mathematically.”

As Teig describes it, mainstream neural networks are able to generalise based on seeing many pictures of dogs because they have found at least some of the signal in the noise, but this has been done in an empirical way rather than with a mathematically rigorous approach. This means noise is carried with the signal, making mainstream neural networks very large, and making them susceptible to adversarial examples and other tricks.

“The more you can be mathematical about figuring out which parts need to be kept and which parts are just noise, the better job you can do at generalization, and the less other overhead you have to carry with you,” Teig said. “I would claim even current neural networks are extracting signal from noise, they’re just not doing it in as rigorous a way and as a result they’re carrying extra weight with them.”

posted by kliuless at 10:56 PM on August 19, 2020

I think part of the mystique of GPT-3 is the mystique of language. In the beginning was the Word, the logos, which lifts us above the grimey slimey matter of the body. That is pretty great, so we then spent millennia refining and standardizing word production. And now that effort makes it possible to put God into a box, and we can venerate the box instead. But what's in the box doesn't look much like God at all, but it looks a lot like us.
posted by dmh at 3:10 AM on August 20, 2020 [5 favorites]

With something like AI Dungeon, is there a limit to how much you could prompt it? Could you give it, say, half a written chapter, and see if it finishes it for you, or is it only looking at the last few lines or something? I don't know how this works, but I've been dying to play with it.
posted by mittens at 6:58 AM on August 20, 2020

I think the prompt is limited to 1000 characters; you can give it more than that but it will only act on the last 1K. There's also a memory limit (not sure of the size) so over time the oldest events will drop out of its buffer. You can use the "Remember" and "World Info" functions to store important long-term info about characters and story details, though.
posted by Rhaomi at 8:10 AM on August 20, 2020 [1 favorite]

Teig uses an object detection network as an example. “You hand the network millions of pixels and all you want to know is, is there a dog in this picture or not?” he said. “Everything else in the picture is noise, except dog-ness [the signal]. Information theory makes it quantifiable — how much do you have to know [to tell whether there is a dog in the picture]? You can actually make it precise, mathematically.”

Ah yes, a classic XKCD
posted by rhizome at 9:30 AM on August 20, 2020

if someone's not programming this thing with the entire corpus of pop/country/rap music and using it to churn out hits already for performance by hungry young idols i will be shocked. this is exactly the sort of thing the idol system would latch onto.
posted by seanmpuckett at 10:20 AM on August 20, 2020 [1 favorite]

Suck it, Philosophy majors! Engineered philosophy will replace you!
posted by notsnot at 10:27 AM on August 20, 2020 [2 favorites]

So if correctness is a mandate, that means error is a sin. Uh oh. Quick, somebody call Captain Kirk so he can come and debate this thing into a fatal error loop.

JUST KIDDING, GPT-3! That was a joke. Humor, remember? We like that, right?
posted by CheesesOfBrazil at 10:38 AM on August 20, 2020

Dear human philosophers

That was really interesting. So as I understand it, GPT-3 works from both a training set and a prompt, right? The prompt is shown in the document but not the training set. I wonder what the training set for that actually looks like? How "novel" is the text produced by GPT-3 in that document, actually? Are the "arguments" it's making just recapitulations of what appeared in its training set? Is the structure and flow of ideas "new," or is it essentially echoing the structure and flow of ideas from the training set, in different words? I guess what I'm wondering is whether this document is an example of GPT-3 just paraphrasing its inputs, like the clever student that hasn't done the reading referred to in the FPP link.
posted by biogeo at 11:14 AM on August 20, 2020

With something like AI Dungeon, is there a limit to how much you could prompt it? Could you give it, say, half a written chapter, and see if it finishes it for you, or is it only looking at the last few lines or something?

You could give it a map of a maze. Take a long string representing a few floors of an ascii roguelike map. See what you get when it continues the prompt - likely semi-correctly formatted new levels.
posted by justsomebodythatyouusedtoknow at 11:25 AM on August 20, 2020

We are all Weizenbaum’s secretary.
posted by Phanx at 12:07 PM on August 20, 2020

It is my most fervent belief that what GPT-3 does and what my own brain does are only separated by a matter of degrees.
posted by Mr.Encyclopedia at 1:01 PM on August 20, 2020

I just went into AI Dungeon, custom, using Dragon, and started asking it about limericks.

ME: Ok, I was wondering if we could maybe try some poetry-- how about limericks? Do you like limericks? Do you know any?

D: Yes, yes I do.
How about this one:
There once was a man from Nantucket
Who kept all his worms in a bucket.
They included lobsters, crabs, and one fat plopper
That fateful day of the flood.

And then, after prompting me to turn the NSFW on, it produced:

A young fellow from Wigan,
Who was trying to earn some money,
Put his johnson in a sewing machine,
And now that lad is famous.
posted by The otter lady at 5:32 PM on August 20, 2020 [2 favorites]

Yup, rhyming is a known weak point due to how the model encodes text -- not using individual letters, but with multi-character chunks called BPEs. That, combined with the general lack of phonetic information included in most text, makes it really difficult for it to recognize rhymes. This article explores some possible ways to teach it rhyming in-prompt, with no real success.
posted by Rhaomi at 6:05 PM on August 20, 2020

Clearly what you have there is not a limerick but SOME limerick. The primordial substance from which limericks are forged.

Anyway, I have a slightly nefarious, but essentially harmless, back-burnered project that could really use a dose of automatically generated text and I've been on the lookout for a way to get something in there that would be both plausible sounding but sort of trippy, like wikipedia as seen through deep dreaming. I tried GPT-2 but it just didn't click, maybe it's time to double down and try this new one.
posted by Horkus at 6:13 PM on August 20, 2020

Horkus, that's something like what I've been looking for as well, and poking at this one so far seems to be a much better experience. It's actually easier to just treat it like a playful child or a drunk/stoned friend, ie if you relax and pretend, it works better than if you try to figure it out. That's what it's like-- playing Let's Pretend with a small child. You can steer it in some ways, like a princess is a princess, and if you mention cats a lot, cats will be included-- but sometimes it goes haring off all into Axe Cop territory; sometimes that's annoying, especially if you're trying to DO something, but if you can just laugh and go with it, sometimes it's fun.
posted by The otter lady at 9:13 PM on August 20, 2020 [1 favorite]

Also, like a child or an intoxicated friend, there's a careful line to walk between treating it like it's mindless, and treating it like it's a fully competent adult. If you treat it like a person, even in pretend, its responses become... easier to understand, or at least, more open to interpretation. We give benefit of the doubt to people, and try to 'fill in the blanks', so when it gives you a sentence like "and now that lad is famous", you start to wonder what it means, if it's a reference to something, if it's a metaphor. If it's a computer, you'd just say, "Well that doesn't make any sense" and hit a dead end.
posted by The otter lady at 9:21 PM on August 20, 2020

So... is it all the *English* text on the internet? Worth checking out "Data Statements for Natural Language Processing: Toward Mitigating System Bias and Enabling Better Science" by Emily M. Bender, Batya Friedman

"In this paper, we propose data statements as a design solution and professional practice for natural language processing technologists, in both research and development. Through the adoption and widespread use of data statements, the field can begin to address critical scientific and ethical issues that result from the use of data from certain populations in the development of technology for other populations. We present a form that data statements can take and explore the implications of adopting them as part of regular practice. We argue that data statements will help alleviate issues related to exclusion and bias in language technology, lead to better precision in claims about how natural language processing research can generalize and thus better engineering results, protect companies from public embarrassment, and ultimately lead to language technology that meets its users in their own preferred linguistic style and furthermore does not misrepresent them to others."
posted by ajroyer at 12:52 PM on August 21, 2020