Wierd AI

AI - The Meme Lord

Episode Summary

Can AI create memes? Be funny? In this episode, we explore the dank world of AI generated content.

Episode Notes

Can AI create memes? Be funny? In this episode, we explore the dank world of AI generated content.

 

GPT-3 Actually useful for explaining jokes.

https://medium.com/ml-everything/using-gpt-3-to-explain-jokes-2001a5aefb68

Actually Jokes

https://www.gwern.net/GPT-3#navy-seal-copypasta

Papers

https://github.com/schesa/ai-memes

https://ai.facebook.com/blog/hateful-memes-challenge-and-data-set/

https://arxiv.org/pdf/1806.04510.pdf

Questions?

Adrian Spataru Linkedin: https://www.linkedin.com/in/spataru/
Bohdan Andrusyak Linkedin: https://www.linkedin.com/in/bandrusyak/

Episode Transcription

--- | Adrian | 00:00:00 ----

Welcome to the Wierd AI podcast. We’re your hosts, Adrian Spataru.


---- | Bohdan | 00:00:10 ----

And Bohdan Andrusyak.


---- | Adrian | 00:00:12 ----

And in this episode, we're going to talk about


---- | Bohdan | 00:00:14 ----

AI: The Meme Lord.


---- | Adrian | 00:00:22 ----

What's the difference between a computer and a dog?


 

---- | Bohdan | 00:00:26 ----

I don't know.

---- | Adrian | 00:00:28 ----

The difference is that when you tell someone a dog is stupid, they don't believe you. Now this joke was generated by GPT-3. So, AI can generate memes. It can generate jokes. Sometimes.

---- | Bohdan | 00:00:48 ----

Sometimes even funny jokes.

---- | Adrian | 00:00:60 ----

And on this episode, we're going to talk about how AI can make humorous online content, whoever uploading names, maybe making good joke, maybe even explaining jokes and more today.


 

---- | Bohdan | 00:01:05 ----

I think it's a good timing for, like, last several episodes, we had quite serious topic, and we need to get to our roots to our memes over mamas and our spicy memes.


 

---- | Adrian | 00:01:21 ----

Why is Meme Lord important? Why are we talking about this topic? First of all, this is a job position. So, I sent you a few days ago, Bohdan, a photo of a company who's hiring a Meme Lord or Meme Lady for their social media campaign, and especially when it comes to Bitcoin everybody on Twitter is posting memes like “Dogecoin to the moon” and stuff like that and making it like a meme and not really like, “Oh, you should buy Bitcoin because over the financial benefits it can provide”, you know, rocket and memes and stuff like this. So, there's a lot of financial incentive to be a good Meme Lord.

 


 

---- | Bohdan | 00:02:05 ----

Yes, exactly. Now my parents wouldn't be ashamed of me of being a Meme Lord. Now I can actually make money making memes. Yeah, but just put some numbers.

It's a multi billion industry now. So, especially like this recent boom of Dogecoin, it's 10 billions market camp or something like this at this moment, and it's all continue growing. And also GME, GameStop short squeeze push was completely powered by Wall Street Bads, and its spicy memes.

So, we're talking about millions and billions of dollars going into the market just because somebody made like, really good meme.


 

---- | Adrian | 00:02:55 ----

So, the first question is, can AI generate a meme? And the answer is yes and no. So, first of all, AI doesn't really. There is no way AI out there, which can generate a new meme format.

What the current AIs do is they take a very known meme format and then maybe put a caption on it. So, the first paper, which I found which talks about memes and AI is “Dank learning: generating memes using deep neural networks” and was published in 2018.

So, it's basically a glorified image captioning algorithm. As input you get you put an image of a very popular, meme format, which you've seen, I don't know.


 

---- | Bohdan | 00:03:47 ----

I think it's like a meme format from classic era of memes like “success kid”. It's like, those memes are so old that I already like, forget what they were...


 

---- | Adrian | 00:04:02 ----

Yeah, back in the days, you know, names were lasting like months and now they're like every week it's a new meme.


 

---- | Bohdan | 00:04:09 ----

Exactly like a state of the art artificial intelligence. It becomes irrelevant in less than a week.


 

---- | Adrian | 00:04:15 ----

Exactly. So anyway, you take a popular meme format, you would then pass that image to a convolutional network and then you wouldn't pass us that embedding tow a LCM, which would then generate a text for you.

So, give it an image, it would give you a caption. And this was the first alliteration of a meme. Nowadays they're a bit more advanced. The state of the art in meme generating is called, an algorithm called MemeBot and as we progress with technologies, we don't use LCMs now.

Now we're using transformers here. So, the way it works is you give a text. That's a context about your meme. For example “please save the world from Coronavirus” like general context. Then the AI will select a meme format which would fit well. And it will then, given the text, also generate a text which fits that meme format. So, for example, “Brace yourself, Corona. I'm coming to save the world” something like that.


 

---- | Bohdan | 00:05:23 ----

I think it's really great that we're talking on audible podcast about visual memes and giving examples. And I think it's like those AIs really simulate browsing, some Reddit meme pages because, like maybe one in 100 will be actually funny and others will be just random. So, just like humans do.

 

---- | Adrian | 00:05:46 ----
Do you remember my amazing joke, which I told you and you were laughing like crazy? This is coming from collection from a person who asked GPT-3, “write me a joke” and AI, GPT-3, at least, if you don't really change the perimeter, is just like a “write me a joke”. You'll be disappointed because he’s an anti-joke joker. So, for example, I'm going to read like, first of examples.

Just to understand what I mean here. 

“Turns out there are two kinds of computers. Those with funny jokes in them and those without.”

Well...


 

---- | Bohdan | 00:06:26 ----

Cue laughter.


 

---- | Adrian | 00:06:30 ----

“If we have an AI that could make jokes will it be funny? Yes, but only if you're a robot”.


 

---- | Bohdan | 00:06:37 ----

I think it's like AI started creating jokes that are funny for AI. Just like memes created communities that are funny from those communities. So we're like, normies for the AI. It's like a meta meta meta joke. Where it’s reached already such a level of technology of just GPT-3 flexing on us and making memes that we cannot understand.


 

---- | Adrian | 00:07:05 ----

Now, that doesn't mean that AI can't generate good jokes, Bohdan, right? So, this is only if you ask a question. But if you really give him like more context, they can generate good jokes. Do you have any examples of that, Bohdan?


 

---- | Bohdan | 00:07:17 ----

Yeah, I have a quite funny exemple, “what do you call a dinosaur that is sleeping? A dino snore”. This is a classic example of a dad joke, meaning maybe the AI in future will be helping with parenting our kids.

 


 

---- | Adrian | 00:07:38 ----

So, when you give GPT-3 examples of like dad jokes, you will get dad jokes, you’ll be somewhere more successful than just asking for a joke.


 

---- | Bohdan | 00:07:48 ---- 

Also, GPT-3 can explain you a joke, if you're a little bit slower on that side. Quite fun examples and I think we will finally be able to use that explicit tech to put more rowdy jokes.


 

---- | Adrian | 00:08:08 ----

As said before: AI is not really funny, GPT is not funny, but you know when you explain a joke, it's not really funny at that point. So, exactly, GPT-3 is good at not funny stuff.


 

---- | Bohdan | 00:08:24 ----

Exactly


 

---- | Adrian | 00:08:28 ----

If you could have dinner with any historical figure, who would you choose?


 

---- | Bohdan | 00:08:34 ----

I think I would choose to have dinner with Alan Turing. And you?


 

---- | Adrian | 00:08:38 ----

Gandhi.


 

---- | Bohdan | 00:08:39 ----

Why?


 

---- | Adrian | 00:08:40 ----

More food for me!


 

---- | Adrian | 00:08:43 ----

GPT-3, please explain this joke to me.


 

---- | Adrian | 00:08:47 ----

Gandhi was a famous Indian leader who was known for his hunger strikes. So, you see GPT-3 can't actually explain the joke quite well like, we get it.


 

---- | Adrian | 00:09:01 ----

Yeah, we? It's horrible to give jokes on a podcast. It's not even... how do you even say jokes? It's been so uncomfortable, so not natural.

 


 


 

---- | Bohdan | 00:09:12 ----

I know I feel like right in my wave of making unfunny jokes. So, for me it's like the best episode we ever recorded.


 

---- | Adrian | 00:09:21 ----

It's the new meta. It’s like...


 

---- | Bohdan | 00:09:24 ----

Also continuing with GPT-3, it can rewrite your favourite books in different styles, making quite nice, funny copypastas in the process.

There's quite a lot of examples of GPT-3 rewriting Harry Potter, like in Hemingway or like different, different authors. So, it's quite cool.

 


 

---- | Adrian | 00:09:51 ----

So, what we talked about this before? This is stylistic text style transfer, and this is like a big topic. You don't need to use GPT -3 like “Oh man, GPT-3, use GPT-3”. Nobody has money for that. I get it.

Like, you can't get access key. But this can be used with your normal machine-learning, like transformers which are public available and you use an extra layer which help you transferring the style. This can be partially done nowadays. 

Okay, so let's go back through MemeLord. Can we have an AI Meme Lord?


 

---- | Bohdan | 00:10:28 ----

To be honest, I think for AI it would be really hard. To get to the stage.

It can’t be Meme Lord. It will have to be retrained constantly because nowadays these memes don't live for a long time and there's no templates anymore. Now it's very unique.

More unique, I would say, are memes that you take some situation and you put funny text on it. So, it's hard to learn the embedding for AI for understanding what the picture means and then put text on that.


 

---- | Adrian | 00:11:05 ----

Yeah, it's quite tricky to get like the relationship between text and image. But not, even though it might be funny, maybe it's old fashioned. Like if you, as you mentioned, use old meme templates like I feel like a boomer. 


 

---- | Bohdan | 00:11:22 ----

Exactly. It's like it's the same as these companies now try to advertise with outdated memes. It's not funny, it’s cringy and I think it's worse for the product when they try to advertise to young and hip kids with the outdated memes.


 


 


 

---- | Adrian | 00:11:38 ----

Exactly like if you show it to current Gen Z. It’s like “Sir, please take your retirement”.

Yes, exactly, however, well, what we can do with AI is identify those trends. So, we can identify like “Okay, let’s go back with Elon Musk and Dogecoin”. Twitter was exploding, so you can already see trending tweets and you can understand the context of what’s going on. So, AI can do sentiment analysis and basic out of the box text analysis to understand what is going on out there. The tricky part is, I think, you understand the context but how can I make it funny? And I don't see there's a good way. Current models like the Memebot, which seems like the relevant to this.

It's not that good. Because you need the predefined temple, but maybe have to do a unique template. We freely context to get, like that Meme Lord action.


 

---- | Bohdan | 00:12:38 ----

And today memes they don't even feel like a complete funny text on it. It can be like, you know, Elon Musk face marriage with Doge and like, written, “Hoddle”. Explain this to AI, explain it to somebody who is not into like crypto and other things like. It's so random that making embedding for it is nearly impossible. 


 

---- | Adrian | 00:13:05 ----

And that's a difficulty with this whole thing, since it's very contextual information like it's you could say jargon and or whatever. It's linguo for a specific niche. And getting data for that. It's very hard since it's like you have so many communities, millions of communities which are in changing the linguo all the time. There's...I find it quite hard. As an example with Hoddle all to really learn that so quickly to adapt it.

Like when AI will find a meme, which is funny it would be one month later.


 

---- | Bohdan | 00:13:41 ----

And I think it's also because memes are, by design, hard to understand. It kind of like shows that you're part of this community and you can communicate in this language that this community understands and outsiders don't understand. So, maybe like memes in future, will be a way for us to communicate that are not tracked by AI, because AI cannot understand memes.


 

---- | Adrian | 00:14:10 ----

Yeah, I mean memes now have become more pop culture than ever before. It got more embraced in the culture. Okay, we're now dabbling around now, so...


 

---- | Bohdan | 00:14:22 ----

Send us your best spicy memes.


 

---- | Adrian | 00:14:24 ----

AI related memes. So how can AI in the future make good memes? And I think the way to find that out is actually from a place from detecting harmful memes like racist names or, like you know, memes which are not good for the society discriminating whatever. And there's like research on this by Facebook, where they try to detect this harmful memes. And why is this relevant? Well, if you want, if you want to make a good meme, you have to understand these, like these subtle contexts, right? And basically what they do is like they take an image of the meme.

They take the text of the meme, and they use two transformers to make a bedding which understands text and images, and they would predict, if it's harmful or not. And I think with that principle, you can take it to like “it's sarcastic? or is it like Doge related?” You can have other outputs instead of hateful or not hateful like I don’t know I'm just thinking out loud here.


 

---- | Bohdan | 00:15:31 ----

Yeah, I think it's like the progress will be done in like this way, where companies try to remove harmful content from there because it's good for their business because it's very hard to make this resource consuming research just to make dunk memes. But when they pull out the models out there and show how it can be done, then it can be fine tuned to make a dankmemes.


 

---- | Adrian | 00:15:57 ----

Exactly, like this fine line of dankness and funny.