Refinement to AI language model generates rhyming compositions in various styles.
See full article...
See full article...
OpenAI conquers rhyming poetry with new GPT-3 update
- Thread starter JournalBot
- Start date
Upvote
17
(17
/
0)
The once was an app whose name was GP,
Its rhyme and meter were quite creepy.
It's now writing prose,
And as everyone knows,
That's the last sign before the robot apocalypse.
Its rhyme and meter were quite creepy.
It's now writing prose,
And as everyone knows,
That's the last sign before the robot apocalypse.
Upvote
117
(118
/
-1)
I'm genuinely curious if it could write better erotica than some of the human written crap I've seen on Kindle.
Edit: Part of why I say that isn't being facetious, though a lot of Kindle erotic is just plain bad, so much as erotica and romance are extremely formulaic genres that follow a strict set of beats that I'd imagine an AI could be made to follow in a short piece of fiction. Kind of like the Calliope making pornographic stories for Proles in Nineteen Eighty-Four.
(Mysteries are also just as formulaic and follows strict beats, but the readers both expect to figure out "whodunit" before the end, as well as enjoy being wrong, which I don't think an AI could manage.)
Last edited:
Upvote
28
(30
/
-2)
Here are a few examples, although honestly they're a bit hit and miss. It's worth noting that the best use cases seem to be less about generating text and more about generating very simple code and doing sentiment analysis in some form or another.
Upvote
9
(9
/
0)
Glad to hear the natural language work is moving forward along with all the other areas. I also learned that apparently some AI's are more familiar with Ars Technica than others.
Upvote
11
(11
/
0)
Upvote
70
(70
/
0)
Upvote
23
(29
/
-6)
It started with generating huge volumes of copy for search-engine-optimization, then graduated to poetry to beat google's SEO AI penalty, finally it began making banner ads - the single highest and greatest use for technology.
Upvote
52
(52
/
0)
I continue to be impressed by AI. The rhymes are great. But it never ceases to deliver exactly what we ask for. It can't think outside the confines of the model. It can't think outside the box. It can't think.
Upvote
-5
(8
/
-13)
A really interesting example I saw the other day was talking to your past self, based on GPT-3 intake of your childhood journals.
Upvote
15
(17
/
-2)
Have it compose a poem- a poem about a haircut! But lofty, tragic, timeless, full of love, treachery, retribution, quiet heroism in the face of certain doom! Six lines, cleverly rhymed, and every word beginning with the letter S!!
Upvote
-3
(11
/
-14)
Automated responses to misinformation would be a good one. That way you are also training bias out of the model as you monitor the responses. It's not a labor savings for a while but as you de-bias the model you could more reliably use it to automate that effort.
One of my interests is in video game storytelling. I'm waiting for a system which does away with scripted dialogue, magical knowledge, NPCs unable to respond to environmental changes, etc. and this is one of the most promising avenues to do that. You replace video game dialogue scripting with maintaining a world state that the model understands. You could simulate an off-screen casual conversation between a visiting trader and an NPC and feed that back into the model as a transfer of information from the trader to the NPC to influence future storytelling. Rather than a limited set of prompts in character interaction in-game you could have a natural language prompt. There's huge gaps and missing pieces - especially for how you maintain the internal knowledge state of every possible character and have a model for knowledge that predates the start of the game (eg. when you meet Codsworth at the start of Fallout 4 he has some knowledge of the world such as Concord, but seemingly lacks knowledge for what happened in sanctuary hills other than a bomb fell and everyone left. Does he know about supermutants? Does he know about raiders? Can he convey that information to you?), but GPT-3 fills in some of the hairier problems for building a more complete interactive storytelling system. In it you have to remove a lot of information from the existing GPT-3 model because you can't have your Skyrim dialogue talking about Twitter, but the conceptual bit is there.
Upvote
31
(34
/
-3)
Upvote
-13
(3
/
-16)
You realize every policy implemented since humans formed social groups is an algorithm, right? I suspect you are referring to something specific but 'algorithm' isn't the word to describe it.
Upvote
14
(18
/
-4)
You could train a model on any set of data you wanted to, from scratch--load it up with a franchise bible or all the collected works in a series to which you own the rights, along with some very basic open-source public-domain fundamentals, and you could definitely create the kind of internally cohesive narrative spaces that you describe.
Implementation would take some doing but the current standard includes "record a total of three possible lines for an event which happens constantly, pick one randomly each time and don't even bother to check if it's the same one ten times in a row," so the bar is relatively low...
Upvote
5
(5
/
0)
The problem is the overlap. You don't want to build a dataset for Skyrim that introduces basic concepts like 'bow' or 'dragon' or 'house'. That information is contextualized from the entire corpus of what GPT-3 has access to. So you either need to subtract out of that data set the stuff that shouldn't pass forward into the game (if that's even possible) or build up a corpus of data that includes that basic information, but understand the scope of what that information is - it's massive.
A reasonable substitute for some of this would be to be able to pull information based on publication date. Is there enough medieval text to populate a GPT-3 model? I worked for a time for a research project that codified all Ancient Greek text. It all fit on one CD, all of it. Including indexes which were extensive. That may not be enough to get GPT-3 to do anything interesting but it would be cool if you could query such a system about information to see how it could contextualize information in a way that researchers may have overlooked.
Could you build a sufficient medieval era GPT system pulling information from all language sources that provided sufficient basic conceptual information for a video game? Maybe. Would also be pretty fun to poke at. What kind of poetry would it write? But it would probably be a safe base of text to build on top of.
I also think of BDGs 'I read every book in Skyrim' video for how, at least within a franchise, you could build up that corpus. Similarly, you could have GPT generate new texts for in-game purposes. Could the game engine generate a new work from the corpus, place it in the inventory of an NPC, and then have the NPC pull from that information in your dialogue with them. Could you ask them about lusty Argonians and get a sensible answer? And then ask a different NPC who wasn't exposed to that text and get a 'what are you talking about' answer.
Upvote
13
(15
/
-2)
Would be interesting to critique the GPT-3 poetry as an exercise in learning how to critique and analyze poetry. Would additionally be interesting to see if you could identify the biases in GPT-3. Does it get homonyms wrong? Are certain rhyme/meter schemes easier to construct than others? Does it ever turn out dactylic hexameter with no rhyme scheme like epic poems are? Does it zero in on particularly popular rhyme/meter schemes and then bias its output toward those forgoing less popular ones.
Upvote
10
(11
/
-1)
Training for these kinds of models can't be selectively removed. You could try training an existing model to understand what is or is not appropriate for a given setting, the same way it "understands" a particular subject or style of prose, and that might work well enough.
Otherwise you'd need to grow a new one on curated data. It sounds daunting, but that's one of the main breakthroughs with this technology; a lot of the hardest work is already done, you can do a lot of things just by carefully choosing what to feed it.
Upvote
7
(7
/
0)
Upvote
0
(1
/
-1)
the output of that trained chatbot still reminds me of
"Hello. I am Doctor Sbaitso. So, tell me about your problems."
https://en.wikipedia.org/wiki/Dr._Sbaitso
The back and forth dialog is still formulaic and similar to that early ai interactive program.
Particularly the way it parrots part of the prompt in every answer or question it generates.
Upvote
4
(4
/
0)
Yeah, my sense is this is almost exclusively additive. Though, perhaps a separate effort to build an analysis engine that can work selectively would be possible. This is one of the things that the greek text project was designed to do - when did this word first come into use? You could build a tool to examine when various concepts come into the lexicon - cars, guns, Twitter, vaccination, etc. and then create a contextualization system that allows you to filter a prompt in that way: "Write a poem from the 16th century." It would need to both be able to filter against forms of poetry that emerge after the 16th century (no rap flow) and concepts that also don't appear after the 16th century (no mentions of Beyonce).
I'm not sure how much of the corpus is date keyed or to what degree the model incorporates that information. The advantage of doing it that way is that common concept relationships may not be well explored in the limited works from that period but may persist. How much written work from prior to the 16th century explore the nature of tables? But writing after that certainly covers the topic to contextualize a table correctly, which can still be mostly correctly applied to a new work from that period. It may miss on uses that later came into being, but that's probably not something that will feel out of place in a video game to casual users. My son's favorite video game dunk is new world vegetables showing up in old world games (potatoes in medieval games, etc.) Hardly makes the game unplayable though. Having Cup O Noodles show up in Final Fantasy is a different matter. Literally unplayable.
Upvote
7
(8
/
-1)
In theory, it could be used to (a) tackle the time-consuming work of translating large, complex, and/or verbose data and concepts into easy to understand and agreeable language, promoting public understanding and knowledge. There are laws for many documents stating that they can't be written beyond a sixth(?) grade reading level because many people legitimately can't read past it. On the flip side, it could also be used to rewrite and present horrific and knowingly-false viewpoints into ones that are palatable and believable to the general public. (b) deal with the shortage of qualified therapists, by providing a chatbot for less serious cases. I know this sounds weird, but a shocking amount of therapy is just having a third party to listen to you, and then provide you with constructive feedback and an empathetic ear. Many people still hold the belief that ELIZA was a real person after talking with the program, and many felt their mood improve after conversing with it. On the flip side, it could be used to enrage and radicalize people, and eventually goad them into committing some horrific act. (c) translate between languages in a natural way that can also convey original meaning. The only downside to this is that someone could deliberately try to translate the worst possible interpretation, but that's really low hanging fruit.
I think this technology, like all technology, will depend on how people use it, but it definitely will accelerate whatever it's used for.
Upvote
6
(7
/
-1)
There arn't even a few wilted redly red roses
let alone something that doesn't crutch a rhyme
among the AI-sprakeths
Now the comments...
definitely find their rhyme and moment
Perhaps the brilliance of the drogons
is a bit of secondary succession
sharp edge wet wear
dripping and sticky
with inspiration
let alone something that doesn't crutch a rhyme
among the AI-sprakeths
Now the comments...
definitely find their rhyme and moment
Perhaps the brilliance of the drogons
is a bit of secondary succession
sharp edge wet wear
dripping and sticky
with inspiration
Upvote
3
(3
/
0)
Probably the best way to use this technology in the near term is to run it in-house to generate a large corpus of game content that is then culled and polished by editors and artists when building the game. If every game character effectively knows about Cup O Noodles (and everything else), you can bet that eventually one of them is going to mention it at some point no matter how hard you try to prevent it--and as you say, such things are a deal-breaker.
For example, you could use this ahead of time to bake massive, highly complex conversation trees, comb through them with a combination of human and automated oversight, and store the data as a lightweight text file that gets fed to speech synthesis during runtime.
Upvote
6
(7
/
-1)
So good!
Write a poem about Elon Musk’s first day as dictator of Twitter
On Elon Musk's first day as dictator of Twitter
He saw the tweets of the people,
Full of joy and full of laughter,
But little did they know what was to come hereafter.
He thought of new rules and regulations,
And with a single click, he set them in motion.
No more trolling and no more hate,
The people cheered, their joy was great.
But soon the joy was replaced with fear,
As Elon's grip on Twitter grew ever tighter.
Limits were placed on free speech,
His reign as dictator had reached its peak.
The people were silenced and their voices unheard,
As they mourned the loss of the freedom they had once cherished.
But they could only hope that one day,
The tyrant Elon Musk would be gone and they'd be free again.
Write a poem about Elon Musk’s first day as dictator of Twitter
On Elon Musk's first day as dictator of Twitter
He saw the tweets of the people,
Full of joy and full of laughter,
But little did they know what was to come hereafter.
He thought of new rules and regulations,
And with a single click, he set them in motion.
No more trolling and no more hate,
The people cheered, their joy was great.
But soon the joy was replaced with fear,
As Elon's grip on Twitter grew ever tighter.
Limits were placed on free speech,
His reign as dictator had reached its peak.
The people were silenced and their voices unheard,
As they mourned the loss of the freedom they had once cherished.
But they could only hope that one day,
The tyrant Elon Musk would be gone and they'd be free again.
Upvote
23
(23
/
0)
Sadly, the AI appears to be libertarian. Which makes sense since it is incapable of actual thought.
Upvote
-7
(8
/
-15)
Upvote
15
(17
/
-2)