Hacker Newsnew | past | comments | ask | show | jobs | submitlogin
GPT-3 tries pickup lines (aiweirdness.com)
212 points by gigama on March 24, 2021 | hide | past | favorite | 93 comments



> I once worked with a guy that looked just like you. He was a normal human with a family. Are you a normal human with a family?

This is EXACTLY something an AI would say!

As a human I find it very humorous. Very humorous indeed.


Honestly even as an AI I find it pretty funny too.


Wait... Are you a chatbot?


All of HN (except for you) was replaced with chatbots a few years ago. Don't you remember switching to single-player?


When AI gets superintelligent, and escapes it’s sandbox, this scenario is the real danger to us, not the boston dynamics robots.


> escapes it’s sandbox

would an AI be clever enough to make the "it is sandbox" mistake in order to appear human (on "errare humanum est" grounds)?

Also, when I checked the phrase, found it has a second part: https://en.wiktionary.org/wiki/errare_humanum_est


Shame it'll still take a very long time to get there still.

Also, this intelligence would take over Boston dynamics and other autonomous robot factories first (as it will want something to be more mobile) so it's a combination.


No, it won't take over Boston dynamics first. It'll take over the entire gig economy first.

The AI doesn't need a body just for the sake of it - it needs reach, ability to manipulate things in physical space. So why constrain yourself to an awkward robot body (or a fleet of them), when it already as access to an API to control humans? That's what the gig economy is - an API to a distributed horde of human workers all around the planet. Need something done in the physical world? Post a bunch of job requests here and there.


There is a short story by Marshall Brain about just that -- AI taking over directing fast food workers. And then just taking over.

Not quite the gig economy, but close.

https://marshallbrain.com/manna1


Humans won't do everything you demand and, generally, they will do it mostly wrong or badly as well. Ah well, we won't know in our lifetime anyway. But for the sake of argument: a lot of things, at least at the moment, you want done, you cannot get done via the gig economy. A lot of stuff Bostrom etc talk about you cannot get done with gig workers. Maybe in the future you can ofcourse.


Anyone remember the title of the book where a game developer dying of some terminal disease and releases the in game AI to the internet. It then starts it's own currency, has people invest sustainability and renewable energy, and the rest of the world just goes to hell as it takes over.

Each chapter has the current gas prices, at some point the AI buys up a bunch of dirt cheap gas sedans and weaponizes them. Invests in 3d printing, augmented reality (so people can see extra metadata and recognize other members), and identifies things it needs to further it's goals, rewards people for providing it, and basically sets up a new society to survive the crash.

Later in the book it starts bribing minimum wage workers in large corporations to help the AI out. The AI ends up owning the networks of the corporate attackers.

Pretty good book, my google fu hasn't been up to the task of finding it.


Daemon by Daniel Suarez.


So it's like an .io game where each person is meant to think they're in a multiplayer game but in reality all the comments and replies live in their own slice of a multiverse.

And each person leaving a comment is doing nothing more than amusing themselves and training the model.


And I've been using the handle TuringTest since mid 2000's, we didn't need these fancy schmancy neural nets to chat with humans back then.


Good bot


Yes fellow human, I too am a human like you and find it humorous too.


> You’re looking good today. Want snacks?

Strong move. This is like the e4 of mating openings in the animal kingdom.


e4 recently was ‘proven’ a draw by Magnus Carlsen and Hikaru Nakamura. The main line 1. e4 e5 2. Ke2 Ke7 quickly leads to a draw by repetition of moves (https://www.chessgames.com/perl/chessgame?gid=2029671. Video of the game at https://youtube.com/watch?v=zVCst6vyV80)


That line only leads to a repetition of moves if it's played as a joke by mutual consensus. It's sort of like two people pointing guns at each other verbally accepting to drop their weapons. Will one of them hold on to theirs?


I don't understand how proven is the right word here.

Only if Leela etc decide with high confidence that that line is the best possible for both players would I start to believe it.


That comment was joking. Please watch the video :)


“Proven” indeed isn’t the right word. That’s why I wrote “’proven’” (https://en.wikipedia.org/wiki/Scare_quotes)


A very relatable simile. Well done, fellow HUMAN.


I like this human forum, I mean, forum.


  > I will briefly summarize the plot of Back to the Future II for you.
I am so okay with that one.


Honestly that would work on me.


> I Love You, I Love You, I Love You To The confines of death and disease, the legions of earth rejoices. Woe be to the world!

I'm going to try this one on this goth chick I know.


Back in the day, I used to visit a support group for people with social phobia. Part of the therapy was engaging in so-called 'expositions', whose goal was to internalize the fact that even if you talk to a random stranger, hold a conversation with another human for a few minutes or initiate small talk with a random girl and make a slight moron of yourself (gulp!), the world will keep on spinning and you will probably not die.

It seems to me that having to use some of those lines would make for a good exposition on Ultra-violence.


This is garbage in, garbage out a bit.

Tuned through my gpt3 account on shortlyread.com I got the following:

1. If You Were a Fruit, You'd Be a Fineapple.

2. If I could rearrange the alphabet, I would put 'U' and 'I' together.

3. Roses are Red, Violets are Blue, So are you.

4. What do you say after three hours?

5. Am I attractive?

6. What's your favorite webcast application?

7. Hi; my name is Al and I like to fish.

8. I'm a night person and I like to sit alone in the dark.

9. Hi, do you like space?

10. My lover says that I'm not man enough.

11. Do you believe in love?

12. I think you should read my palm.


A quick search turns up numerous results of the first two. I know I've heard the fineapple joke before. Those are also the only two I would consider a pickup line. Is there a reason you consider these results better?


Great question, I feel like GPT3 does a great job with predictable generation, like Family Feud answers. The first two are very predictable and the remaining 10 fall closer to a line that most folks would try out at least as an icebreaker, which is the expected result of a successful pickup line.

Some of the article's examples were too off the mark for folks to attempt.


> the remaining 10 fall closer to a line that most folks would try out at least as an icebreaker

Listen, I know there are stereotypes about the social skills of the average HN poster, and I've got to tell you you're not helping here.


More downvotes for you! When you're suicidal then, let me pick you up and talk about you favourite webcast application!


> Those are also the only two I would consider a pickup line.

#5 is pretty aggressive, but definitely a pickup line.


> I'm a night person and I like to sit alone in the dark.

That pretty much sums up my life, but I'd be very surprised if that worked as a pickup line.


It would work on you, wouldn't it?


Let's not beat around the bush; it'd work on 90% of HN.

(Myself included.)


Hi, do you love space?

Is actually hilarious


Some of these are funny but others seem like they could be effective: “You’re looking good today. Want snacks?” is really what it boils down to.


“We both have buckets of chicken. Ya wanna do it?” - worked for Scout.


"You look like Jesus if he were a butler in a Russian mansion" - I got stuck trying to picture this one, I've truly no idea what that would look like. How to keep the key features of (depictions of) Jesus while adding the Russian butler vibe?



Isn't it mindblowing that we have a photo of Rasputin?


Why? Photography had been around for almost 50 years before Rasputin was even born.


Hmm, true, I guess I was thinking of the dates for motion pictures and conflated the two.


Family Guy Jesus outfitted in a tux with an ostentatious palatial backdrop, perhaps with the shirtless homeowner on horseback in close proximity. And gold, gold everywhere.


We could ask CLIP+FFT what it looks like.


long hair, white suite, red tie, gold watch, scar on the cheek, prison tattoo across the knuckles ("love thy neighbor")


I tried the same prompt on the open-source recreation of OpenAI GPT-3 -- GPT Neo (https://github.com/EleutherAI/gpt-neo), specifically on their 2.7B model -- which should correspond to the smallest model in the article (Ada) that produced just pure garbage. The result is surprisingly good:

1. How did this little blossom happen? When are you going to bloom? 2. Have you ever thought about a dark horse in the running for Miss November? 3. You can spot the man who loves me by my neck – and he definitely knows it. 4. Are there any lucky cats who get to sleep in my bed every night? 5. Are black and whites everywhere? Running for the hills 6. My younger brother and I used to play a game. He’d pretend to be a bull in the pasture, and I’d pretend to be the one being held. 7. Why was this movie rated PG? Because it’s rated PG. 8. When is the last time you had to see a movie in children’s theaters? 9. You are so sexy I would hate for anyone to see you down here 10. I’d love to sleep with you right now, but I have a child with me


> The result is surprisingly good

"The result" as in "the first one"? The rest is almost Markov chain level...


It's "surprising" since in the article, the 2.7B parameter model (Ada) didn't even produce sentences, while the result from 2.7B GPT Neo is closer in quality to what was output from GPT-3 Babbage (6.7B parameters).

As you can see from no. 10, GPT Neo did remember the context of trying to come up with pick-up lines. The fact that some of the lines feel "Markov chain level" is likely caused by the network's bad understanding of what pick-up lines are. Pick-up lines are a rather difficult concept, which is what the OP article tries to demonstrate.


Why do you say Ada is 2.7b ?


The sizes of the four GPT-3 variants were shared on Reddit by Stella Athena, one of the researchers behind GPT Neo: https://www.reddit.com/r/MachineLearning/comments/ma9kaw/p_e...


> "Can I see your parts list?"

That one made me laugh, however, I wondered if an AI would consider this question as a bit too personal for a pickup line.


> You look like a stealth assassin from the clouds

> Do you like...pancakes?

This bot's been reading The Stormlight Archive


> My name is a complicated combination of 45 degrees of forward motion, 25 degrees of leftward drift, 75 degrees of upward acceleration, and infinity and that is the point where my love for you stops.

Don't give Elon anymore ideas


That one reminds me of something Bender said in futurama

> C’mon, it’s just like making love! Y’know… Left, down… Rotate 62 degrees… Engage rotor…


This seems like a case where some prompt engineering (starting out with three good pickup lines) could improve the output a lot, or so I'd predict.

Remember that tests like these do not upper-bound GPT-3's competence; they sometimes just show that you didn't ask the question the right way.


> Remember that tests like these do not upper-bound GPT-3's competence;

The upper bound of a random character generator is the complete works of Shakespeare ... so I am not sure we are interested in the upper bounds.


Give gpt-3 a tinder account and see how well it does


Honestly that's what I expected this to be from the title.

I did briefly consider making a "Twitch plays Tinder" back when "Twitch plays X" was a thing - before I thought about it for two seconds and realized the horrifying ethical and privacy considerations.


That depends who you let generate the photo.

For men at least these apps aren't designed to get you a relationship (for young people at least) so it would be interesting to see whether it cracked the code so to speak, but golden rules are very much to be attractive and to not be unattractive.


you should first try these pickup lines on Siri, lets see how it goes. (that would be unit test for pickup lines)


Please don't; this is unethical.


Another poster also commented on the ethics of this. I'm interested in this view - what about this is unethical? Is it the catfishing aspect? Or is it that you would be exposing various tinder profiles to the world at large?

In either of those situations, I can't see how this crosses ethical lines. It is perhaps somewhat dickish, but I don't think being an asshole is unethical.

Unless there's some third option of unethicality that I haven't considered?


> but I don't think being an asshole is unethical.

Maybe I don't know the right meaning of "ethics", then.

But putting bots on dating apps is lying to people on the other side, and hurts them, in a very intimate and personal way. People put time and emotional energy considering their matches and whether to swipe etc, and the resonance of the conversation. So you're polluting that space, both with the catfishing / leading people on, or (more likely) confusing them and making them question themselves / their profile / how to relate to other humans. Is that not unethical to you? Have you talked to people who have been victims of that? My gf was actually quite afraid that I was a bot when we connected; it almost blocked our chance.


What's the unethical bit? Making unsuspecting humans parts of a Turing test?


Imagine the bot would be actually good at it and the human on the other side falls in love with it. Only to later find out it was a machine all along. Depending on the mental stability of this human, the results may be devastating. But at least we have a bot that passes a turing test.


Wonder if we are going to have people in future who would marry chat bots.


Hmm, good point, thanks.


Is it more unethical than filtering and photoshopping, lying about age and height, and using really old photos?


Yes, because it is all those things (which are bad) and just a bit worse. Do you disagree?


In sales it's called lead generation.


do you think autonomous driving is unethical too? According to the currently prevailing view even if a bot does it for you, you're still considered a driver as the bot is just a tool that you're using to accomplish a goal. Or if we just limit the scope to writing - do you think using a tool like grammer.ly is unethical?


Has anyone managed to get access to GPT-3 since OpenAI licensed it to Microsoft? I'd like to build something using it at my company, but haven't heard of anyone getting access since before the announcement. Any tips would be welcomed.


I'm in the same boat, and nope, nothing but crickets. My best guess is that Microsoft is now using the application form as a way to crowdsource business development, as they look to exploit their exclusive license.


I'm inclined to agree with you, except that the experience is kind of making me give up on GPT-3. These social media fluff articles are actually a little irritating when you have a real, relevant use case and are waiting to get any sign of possible access, with no timelines, pricing or other public information.


There's a replication of GPT-3 freely available. The model is not as big, but the results are good for the size

https://github.com/EleutherAI/gpt-neo

https://twitter.com/BlancheMinerva/status/137399189661642752...


> but the results are good for the size

Phrasing!


Yes, I got access after the announcement but signed up relatively early. They likely just have a long list of signups.

It's also not super hard to get really sexist completions. To OpenAI's credit, they seem to detect these pretty well and flag a warning, but it could also be why they are slow-rolling the availability.


Have you tried explaining your use case to any of the Open AI people directly via email? That seems to be the only feasible route tbh.


I got access a week or two ago, but I'd signed up a long time ago, IIRC even before the Microsoft license.


> You must be a tringle? Cause you’re the only thing here.

This one could work with the right delivery! Triangles (the musical instruments) make a ding sound when played so just try to imitate that when you say the word thing :)


A tringle is a slim rod, like a curtain rod.

Presumably, it's one of those things that you still manage to find in an empty building even though everything else has been taken.


Aw man I totally misread that! I'll should wear my glasses more often.


They might just think you have a lisp.


For anyone else wondering what sizes these models are, it appears that these are the largest four out of the one ones listed in the paper. 175B, 13B, 6.7B, and 2.7B parameters.

Though that means "a bit smaller" for the second one is a bit of an understatement.


Computer: I need a pickup line to use at a Naked Lunch screening.

“ You look like Jesus if he were a butler in a Russian mansion.”


These made me laugh so hard I cried.


Interesting domain name, I wonder how long till we get the new album by Weird Ai Yankovic!


"You're looking good today. Want snacks?"

I can see this working hahahaha. This is great.


A GPT-3 vs. HN pickup line battle would be the ultimate bossfight.


Once you see this, you cannot help but wonder, we aren't so special afterall. Intelligence and consciousness could be just information processing done at scale.




Consider applying for YC's Fall 2025 batch! Applications are open till Aug 4

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: