OpenAI's GPT-3 may be the biggest thing since bitcoin
-
-
I explain why I think GPT-3 has disruptive potential comparable to that of blockchain technology.
So... Much ado about nothing?
I predict that, unlike its two predecessors (PTB and OpenAI GPT-2), OpenAI GPT-3 will eventually be widely used to pretend the author of a text is a person of interest, with unpredictable and amusing effects on various communities. I further predict that this will spark a creative gold rush among talented amateurs to train similar models and adapt them to a variety of purposes, including: mock news, “researched journalism”, advertising, politics, and propaganda.
At least we'll have a laugh then...
This article was fully written by GPT-3. Where you able to recognize it?
Oh, well played.
-
@JBert said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
This article was fully written by GPT-3.
-
@Gąska said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
@JBert said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
This article was fully written by GPT-3.
Written by GPT-3 and edited in vitro.
-
@Tsaukpaetra until I see with my own eyes that software decided all by itself to put scare quotes around researched journalism, I'll remain skeptical.
-
@Gąska said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
@Tsaukpaetra until I see with my own eyes that software decided all by itself to put scare quotes around researched journalism, I'll remain skeptical.
What I mean was, the source training set was crafted for a particular result. For example, consistently scare-quoting.
-
@Tsaukpaetra if that was the case, you'd expect it to happen more than once. It's generally impossible to train current-gen AI specifically to scare-quote, and then NOT have it put scare-quotes somewhere every other sentence. Not to mention that I've read about GPT-3 before and I know for a fact that this particular trained instance (because GPT-3 is a particular trained instance) wasn't trained specifically to do scare quoting, but to be a general-purpose chatbot/text generator. And while it is very impressive, it still feels like the editor of that article did so much editing that it's more fair to call him a co-author.
-
@Gąska said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
feels like the editor of that article did so much editing that it's more fair to call him a co-author.
Fair enough.
It's not like they
canwill provide methodology that will be able to reproduce the article anyways...
-
@Tsaukpaetra said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
It's not like they
canwill provide methodology that will be able to reproduce the article anyways...I actually thought about emailing the poster and asking for video recording of them using the bot to generate the article. Should be easy enough for them, but yeah, I don't think they'd cooperate.
-
@JBert said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
Where you able to recognize it?
No. I assumed people on a bitcoin forum to be idiots and thus the ability to fool them not too far fetched.
The article overall felt a bit weirdly formulated but not too much out of the ordinary for shitty sensationalist garbage. So, well played, I guess. If it’s true.
-
@Gąska said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
I actually thought about emailing the poster and asking for video recording of them using the bot to generate the article. Should be easy enough for them, but yeah, I don't think they'd cooperate.
I agree. Generating 100 articles and picking the best one is easier than filming themselves 100 times and picking the video with the best result.
-
@Zecc if you had to reroll 100 times to get something that sounds genuinely good, how many times do you think you'd need to try before you get on that's sounds at least half decent?
-
Found this on /. figured it was worthy of some laughter:
https://www.lesswrong.com/posts/N6vZEnCn6A95Xn39p/are-we-in-an-ai-overhang
-
@Dragoon Those very large systems are all very well, but still suck in many ways because they take crazy quantities of data (and energy, and money) to train. We don't want a single mega-brain AI, we want lots of smaller ones in lots of places, able to learn from much sparser data and able to apply their learning while the learning is going on. Right now, we've got the Thomas Watson-style approach to AI, whereas we want them more like PCs or smartphones.
"I think there is a world market for maybe five computers."
Thomas Watson, president of IBM, 1943
-
GPT-3 is scary. Not just because it might pass the Turing test despite being designed to just predict words but also because like all neural networks it's incomprehensible black magic. I'm tempted to learn machine learning just to see if the experts that create those models really understand them or if they create these models through pure chance and brute force.
-
Why GPT-3 is Good for Comedy, or: Don’t Ever Do an AMA On Reddit
Of everything I’ve gotten GPT-3 to do comedy has been the easiest and the most fun. It has a superhuman capacity for generating absurdity, and often carries a joke to the upteenth degree, as …
Not that good, but not that bad either.
-
@magnusmaster said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
GPT-3 is scary. Not just because it might pass the Turing test despite being designed to just predict words but also because like all neural networks it's incomprehensible black magic. I'm tempted to learn machine learning just to see if the experts that create those models really understand them or if they create these models through pure chance and brute force.
It's incomprehensible black magic to experts too. Or rather the explanations that they have don't really correspond with any kind of common sense at all. (It's to do with abstract feature detection, but the abstract features are just that: abstract. They're really hard to comprehend the meaning of.) There is a significant amount of pure chance and brute force involved — well, actually there's a hell of a lot of both — but GPT-3 shows that if you've got enough of 'em plus a suitable discriminator algorithm then you really can succeed. Genetic algorithms are another example of mostly-chance-and-brute-force approaches that work surprisingly well.
The big thing with GPT-3 over its predecessors seems to be that they've increased the number of layers. That matters because more layers means more sophisticated discrimination and prediction. (I think they've also increased the size of each layer, but that just lets its recognise more without adding subtlety.) I think of it being a bit like adding another layer of
*to a type in C; with 1 it's simple enough, with 2 you need to think a little, at 3 things are getting worrying, and at 4 you should hold on tight to your breeches! Beyond 4 is scary territory where the dragons of abstraction hang out.The big downer of GPT-3 is that the technology they're using is crazy energy expensive.
-
@dkf said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
Genetic algorithms are another example of mostly-chance-and-brute-force approaches that work surprisingly well.
I prefer to think of genetic algorithms as the triumph of violence over incompetence.
-
@Zecc said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
Why GPT-3 is Good for Comedy, or: Don’t Ever Do an AMA On Reddit
Of everything I’ve gotten GPT-3 to do comedy has been the easiest and the most fun. It has a superhuman capacity for generating absurdity, and often carries a joke to the upteenth degree, as …
Not that good, but not that bad either.
The linked one about San Francisco (warning: there's shit in this one) is kind of funny
-
@hungrier It looks like there might be a few kinks to be worked out...
-
Ah, a new word for buzzword bingo.
I might try to get access to this and write guardian articles.
-
@dkf said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
It's incomprehensible black magic to experts too. Or rather the explanations that they have don't really correspond with any kind of common sense at all.
@error_bot xkcd machine learning
-
-
-
@boomzilla on the contrary this seems like it’s nailed the concept just fine.
-
@Arantor said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
@boomzilla on the contrary this seems like it’s nailed the concept just fine.
Yes yes yes yes yes. Ants can stay. Beans.
-
@Gribnit confirmed as GPT-3 neural network output.
-
@Gribnit said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
Beans.
Fart Booby. Biohazard.
-
@Arantor you have sunk my battleship.
-
Might be worth a look.
-
-
It just occurred to me... "GPT-3 may be the biggest thing since Bitcoin... to waste computational effort for unprovable benefit".
And if you can replace shitty writers with shitty AI-generated articles, your problem was really that you had a shitty writer.
-
@dkf said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
@Gribnit said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
Beans.
Fart Booby. Biohazard.
@DogsB said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
Might be worth a look.
Disappointing.
-
It's trying to sell itself as a good idea, I see.
Edit: and it knows how to type "intelligence" better than I.
-
@Zecc said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
@dkf said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
@Gribnit said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
Beans.
Fart Booby. Biohazard.
@DogsB said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
Might be worth a look.
Disappointing.
Five passengers in a biohazard waste truck? Well, if the boobies belonged to the passengers, it's entirely possible they did have something to do with the accident.
-
-
@sockpuppet7 said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
Slap it on the front page.
-
-
@Zecc said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
Is this a Walmart movie?
If it is, are we part of the Target audience?
-
-
@topspin said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
@sockpuppet7 said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
@Tsaukpaetra thread is
I have been summoned, and so I appear...
-
@topspin said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
@sockpuppet7 said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
@Tsaukpaetra thread is
-
@hungrier said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
can't help but feel a sense of dread as he continues on his journey. He's never traveled so far, and he has no idea what's waiting for him.
Probably ponies.
-
-
@Watson no information on the wherabouts of
are currently available.
-
@Watson the training is complete.
-
@Zecc said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
It's trying to sell itself as a good idea, I see.
Edit: and it knows how to type "intelligence" better than I.
This looks very familiar. I think I've read a very similar text in some sci fi.
-
@Carnage said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
@Zecc said in OpenAI's GPT-3 may be the biggest thing since bitcoin:
It's trying to sell itself as a good idea, I see.
Edit: and it knows how to type "intelligence" better than I.
This looks very familiar. I think I've read a very similar text in some sci fi.
Yeah it's one of the more self-indulgent items in Isaac Asimov's Gold collection, not aware of OP, don't want to find out. Book is in room, still too annoyed at writer to open for a few months.
OpenAI's GPT-3 may be the biggest thing since bitcoin
1838: Machine Learning - explain xkcd
