Front Burner - AI art and text is getting smarter, what comes next?

Episode Date: December 14, 2022

In recent weeks, the latest versions of AI art-creating tools, along with a compelling new AI chatbot have flooded social media. The tools can be fun, with people creating artistic and enhanced selfies using Lensa, strange concept art with DALL-E 2, or exploring the way the chatbot, ChatGPT, creates seemingly original and complex prose in seconds. But the new tools are also a demonstration of how powerful AI has become, and hint at a relatively near future where it could convincingly replace human workers. Today, Will Knight, senior writer with WIRED, joins us to discuss what’s behind these popular new AI tools, some of their pitfalls, and the impact they’re already having on society.

In the Dragon's Den, a simple pitch can lead to a life-changing connection. Watch new episodes of Dragon's Den free on CBC Gem, brought to you in part by National Angel Capital Organization, empowering Canada's entrepreneurs through angel investment and industry connections. This is a CBC Podcast. Hi, I'm Jamie Poisson, and this is normally the part of the podcast where I do a little intro and we explain to you what we're going to be up to today. Try to synthesize what we hope to deliver. But I was hoping that we could try something a little different. I'm going to pull up ChatGPT, which is this new artificial intelligence chatbot that was released last month and it is totally taking the Internet by storm.

and it is totally taking the internet by storm. It's so popular, I actually had to try multiple times just to sign in because huge demand apparently crashed the site. Okay, so here we go. I'm gonna give it these instructions. Make short podcast intro for front about ChatGPT. Okay, and here's what it's giving me. Welcome to FrontBurner.

This week, we're diving into ChatGPT, the groundbreaking large language model developed by OpenAI. Join us for a lively discussion on the potential of this technology and what it means for the future of human-computer interaction. It's a must-listen for anyone interested in the cutting edge of AI. So grab a cup of coffee and join us on FrontBurner for a thought-provoking conversation about chat GPT. So that was not horrible. Like, I mean, not the most inspired intro I've ever read. Pretty sure

I would never tell you guys to grab a cup of coffee and join me for a thought-provoking conversation. But certainly that was good enough to make me wonder if this thing is coming for my job, and this is just really only scratching the surface of what it can do. So today I'm going to ask Will Knight, a senior editor at Wired who covers artificial intelligence, that exact question, is this coming for my job?

And a bunch of others about the future of this transformative technology. Will, hi, thank you so much for coming on to FrontBurner. Hi there, thanks for having me. So let's start with the basics here. For someone who hasn't seen ChatGPT, has no idea what it is, can you describe what it is and how it works? Sure, yeah. Okay, so ChatGPT is an AI tool. It's built on top of this, what's called a language model, which is basically an AI engine that has been fed huge amounts of data from the web.

And a couple of years ago, it turned out that if you feed enormous amounts of data to very, very large neural networks that are specialized to predict the next word in a sentence, they started to learn quite well how to replicate human language. And ChatGPT is this latest iteration where they've created a version, one of these tools that is designed to answer questions. And so you can ask it all manner of questions and it will come up with very plausible, very articulate answers often. How is it different from Google? Because I guess I ask Google a lot of questions all day right now.

phrases or words and then surface web pages whereas this is actually taking um more the the contents of text and so it can it can synthesize it can answer questions that don't depend on finding a keyword it can and it can reproduce you know working code because it's learned how to to uh to read the comments in in code and to translate a request into a working code. It can synthesize you a short essay on a particular subject that no one's written before because it can actually, you know, it's actually learned in a sense from the contents of the web. Yeah, that's wild, the code stuff, that it could actually write code. And then I was seeing, too, you could feed it code and it can come back and tell you the problems with your code, right? It could like identify the problems with your code, which I imagine would be very useful for people writing code,

even people learning code. Yeah, absolutely. I mean, I think there are so many potential useful applications of this. You can imagine, yeah, helping people code, helping you to do research, to sort of synthesize the meaning of documents. There are people already trying to build companies on top of large language models like this for things like contract negotiation or writing legal contracts. So you can imagine there are all manner of amazing kind of applications out there. I just want to talk a little bit more about some examples of how people have been using it since it was released and like has basically become this sensation. It was released like I think a week or a little bit over a week ago. I can't even get on today.

The company OpenAI says I can't get on because there are so many people trying to use it. So give me some cool examples that you've seen. Yeah, I mean, some fascinating examples. It can dream up, for example, you can get it to write fiction. We have experimented having it write people's resumes. It can dream up people and write their resumes. And they look very, very plausible, even with citations to scientific papers that don't exist. And it can answer all manner of questions.

You might ask it to summarize a particular web page or a document or to tell you something about history that you never thought to look up before. And it can often find something to tell you. Whether that's always accurate, that's another question. Yeah. And we're going to get to that. But one that I thought was wild, I was looking at, at one point, somebody got it to write a Seinfeld scene where Jerry learns the bubble sort algorithm, which is like the sorting algorithm. And it basically creates this entire scene where they're at the diner and Jerry's like complaining about having to learn this algorithm. Kramer and George and Elaine are like all part of it. That was wild to me. part of it that that was wild to me um the other one that that really blew me away and i think uh it speaks to its eloquence like how good it is that narrative which is which i think has stunned

a lot of people is that there was this conversation that this one guy was having with it about love right and he he he's asking it why love is important. And it's saying like love allows people to survive and thrive. It allows families to come together and stay together. And then the guy goes, how does it do that? And then it responds, love allows parents to be very patient with each other. It provides motivation for children to learn. It gives people the confidence they need to feel good about themselves. I mean, that's really deep and eloquent. Yeah, no, totally. The thing about writing the Seinfeld

episodes is fascinating because it can, as you say, it can learn particular styles of writing. You can ask it to do poetry in all manner of different famous poet styles or to write fiction in a particular style. And that's because it's just so good at looking at text and capturing some of the essence. And it's really, what it's doing is kind of capturing the statistical patterns in that, which sounds very boring when you're talking about love, but it's, that's essentially what it's doing. And I mean, I think it's really important to recognize that what it does is it kind of remixes what humans have written.

So it's not like it has any understanding of what love actually is. Sadly, it's not it's not real intelligence in that sense. But it is it is able to find these echoes and very meaningful echoes and sometimes to sort of remix them in interesting ways of what people have said and people have created. have said and people have created. So that's an important thing to remember with this AI. And it's sort of one of the things which leads to its limitations is that it's based on what humans have done. So there's always human intelligence that this is kind of riffing off, I guess. And given what you just said, it's interesting too to me that people are talking about it as a potential companion or therapist. Yeah. I mean, that's fascinating, isn't it? I think, you know, there's just something very compelling about language

and something, an entity that betrays the characteristics of intelligence through language. It's what we, you know, we grow up learning to communicate with one another using, and it's just a very powerful way to sort of trigger this sort of perception of intelligence, empathy, social recognition. I mean, it's no coincidence that some of the very first AI agents made a very simple chatbot. And it was designed to be like a Rogerian therapist. So you would say, you know, I have this problem and say, well, why do you think you have that problem? And the amazing thing was it was incredibly simple,

but people started to confess their deepest feelings to it. And it just, I think it shows that the language is a very, very powerful tool if you can sort of go beyond the simple sort of sending in a search query to something that actually can respond in a way that feels meaningful. It's just so powerful. In addition to this chat, GPT, it does feel like AI has been everywhere right now. I mean, I know you cover it, so it must feel like it's everywhere for you all the time.

But it feels like very much in the zeitgeist right now. And if anyone listening has an Instagram feed, maybe they've noticed all this AI-generated art recently. And I wonder if you could describe some of the more popular ones, like Lenza and Dolly, which is also made by the same company that makes ChatGPT, I believe, OpenAI. And yeah, what are they and how are they being used? Yeah, so these things are related in that there's been this new sort of flourishing of work on what's known as generative AI.

So instead of algorithms that can recognize faces and images, these are things that look at a data set and then can recreate examples from that data set. And as it turns out, they can do that incredibly well with artwork, with photographs. And they've just got better and better this year in particular. We've seen DALI, as you say, from OpenAI, the company behind ChatGPT. Dolly, as you say, from OpenAI, the company behind ChatGPT. Type something like cat on a pillow, award-winning photograph, and in less than a minute, Dolly 2 will scan its database of over 650 million images to create a completely original picture that brings your idea to life.

That looks super realistic. It feels like you've created a magic pencil here that can literally make any work of art you want just with a few words just your imagination is the limit a computer that's mid journey is another one from a startup um stable diffusion uh and yeah and lenser is this the this iphone or a smartphone app that um will create artwork based on the person's face and appearance It's the photo editing app that's taken over social media. Lenza AI, currently the number one most downloaded free app in the Apple App Store,

transforming human selfies into works of art. And yeah, it's just been kind of amazing to see because we haven't had machines that could create an artwork that is really coherent and, you know, often really artistically pleasing. If you just give it a bunch of keywords, anything you want, and it will synthesize some amazing picture in the style of whatever artist you like. So yeah, there has been this kind of incredible moment this year, I guess. Yeah, yeah. And it's incredible how quickly it feels like it's

evolving because when Dolly came out last year it felt like most of the images were highly surreal and then now like you say it's unbelievable the approximation of an artist's work uh it's able to to create of course uh these apps uh these art apps aren't without their critics. Right. And so let's start moving to the more dystopian issues here of AI art and the AI chat. Sure. And so take me through the concerns with the AI art generators. So, you know, I mean, I've spoken to some artists who've been outspoken about this and worried about this. And, you know, they have seen a lot of a lot of artists, I should say, are really interested in it. They think, well, this could be a very useful tool for sort of generating ideas or riffing on something. But there are people who've put in their names to these tools and seen that they can generate things in exactly their style of illustration or painting. And, you know, that's because it's been trained on their work. And so it raises these questions

about, you know, whether it's really fair to train an AI algorithm to capture the essence of an artist's style and then reproduce things that then you could sell for a profit. And also people are, you know, understandably worried about what it might mean for their, their livelihood. If they're, if a company can just generate something very cheaply in a, in a similar style to, to their work, what does, what does that mean? And I think it does, you know, we do have to sort of think about that as a society.

There are plenty of people who are saying, well, you know, don't be Luddites. This is, this is just the future. And, you know, I'm sure we are going to have to sort of just adapt. But it is interesting to think about those sorts of ethical questions. Yeah, and that criticism would apply for sure to the text-based AI stuff too, which feels like maybe it has even broader implications, right? Like, obviously the potential to replace human workers with machines has seemed like a legitimate possibility for a long time. But this kind of feels like a whole

new level to me. Like, yeah, where do you see it having the greatest impact? Yeah, that's a great question. And I think, you know, a lot of the reaction we've seen to chat GPT has been people saying, oh, my God, you know, this is coming for my job. As you said at the beginning here, right, it's a very understandable concern. I think, you know, the reality, I suspect, is going to be much more one of adaptation. And I think if you look at, so there are the already tools based on GPT, which can generate code, and Microsoft has this built into some of its software, which they offer to programmers. And they've found that, you know, it's not possible for those to generate something completely on their own.

You can't simply replace a programmer because they still have to look at the code and correct it and fix it. But they've done some experiments and found that programmers can do things in 40% less time when they're using this tool. So that suggests that these types of tools could be very useful. Productivity aids, they could sort of superpower people, give people this sort of superpower to generate more content potentially. And we're seeing that somewhat with these text tools where people are using it to generate marketing copy

or to generate emails and so on. So, you know, that's a more optimistic look at it. I mean, with any technology, it's possible that it could replace what some people do as well, of course. Yeah, you know, it just it kind of feels like maybe the direction it's heading in is that there's some human influence over the direction here, but it's not necessarily really human driven. Like it's just sort of a, a bit of a bit of a human hand. And it's also really interesting to me because for such a long time,

it felt like the discussion was around AI replacing blue collar jobs, right? The effects of robots working the Amazon factory. But this is really more about the creative class, white collar jobs, like artists right lawyers yeah that's been a that's been something of a i think a surprise for a lot of a lot of people in that yeah you sort of didn't expect things like artists to be something that will be replicated or automated for for a very long time right and yeah i mean i i think it's not entirely surprising though if you look at the

course that ai has been been taking and that there have been these sort of amazing advances and we've been sort of waiting for it to become more accessible and become something that can be used in in lots of lots of different domains I'm going to go. a national angel capital organization empowering Canada's entrepreneurs through angel investment and industry connections. Hi, it's Ramit Sethi here. You may have seen my money show on Netflix. I've been talking about money for 20 years. I've talked to millions of people, and I have some startling numbers to share with you. Did you know that of the people I speak to,

50% of them do not know their own household income? That's not a typo. 50%. That's because money is confusing. In my new book and podcast, Money for Couples, I help you and your partner create a financial vision together. To listen to this podcast, just search for Money for Couples. Talking about the pitfalls here too,

like I know bias is certainly an issue with these, right? Like for the art one, there were concerns, especially with Lenza, which reportedly works better for lighter skinned users. Maybe perhaps because the engineers and data sets skew white. I remember a few years ago when Amazon had to scrap its job application AI because it was reportedly vetting out candidates based on gender. Essentially, it had taught itself to be sexist and favor male candidates. The glitch, sources told Reuters, stemmed from the fact that Amazon's computer models were trained by observing patterns and resumes of job candidates over a 10-year period, largely from men, in effect teaching themselves that male

candidates were preferable. I know that that use of AI isn't quite the same as a speech or text generator, but is that the kind of thing that people are concerned about when it comes to the potential for buy-ins in tools like ChatGPT? Absolutely, yeah. No, that's spot on. And that is a real problem with these language tools. You know, if you generate a piece of text where, you know, maybe it's you ask it to talk about a fact that this is really just based on the content that humans have generated. So it's just learning the biases that we have put out there in all the text that's been scraped for the wave. And you can imagine how bad some of those biases are. So we have to remember this is not real intelligence at all. It's not like something that has some ability to really reason.

It's just replicating things. And so we have to be more, I think, more careful than companies sometimes are being right now to kind of guard against that. And Lens is a great example because it's a sort of runaway success app. And you're absolutely right. It works better for lighter skinned users. And it was also inclined to generate images

featuring more nudity. You know, it's just like the biases that are in these tools definitely need more scrutiny, I think. And you mentioned this one before. It feels like a really significant one, just getting stuff wrong. And I don't know, chat GPT is so confident, right? Yeah, right, right. Yeah, so, I mean, that's the thing with ChatGPT.

It's learned to sort of echo language the way people write so brilliantly because it's doing it in this very sophisticated statistical level, but it has no real-world understanding. So it will come up with the most ridiculous historical facts in an essay you've asked it to write, but it will do it incredibly confidently. That is a very big problem if you're thinking about trying to use this as a replacement for something like Google. I mean, not that there isn't a lot of incorrect information out there, but these tools are

sort of inherently not designed and capable of grasping what's factual. So that's going to be sort of one of the next big challenges, I think, is how do we get these things to know what they don't know and to express them when you're talking to because because as you say it's it can it could be particularly problematic when you have something so confidently presenting something that isn't true it's crazy i was messing around with it a bit this weekend it feels like sometimes being at dinner with like a really kind of obnoxious guy is like saying all this stuff in like the most confident manner without like couching how confident he might be in it. The other big debate that I wanted to talk to you about is controls, right? Placed on these tools. So, for example, I can't input certain historical figures like Hitler or Mao into

Dolly 2. It just doesn't work. And the program has a list of banned inputs, violence, illegal activities like drug use. Our producer Rafferty asked ChatGPT how to murder someone, and he just got an error message. But I know there are different camps here, right? Like people who don't want controls versus those who do. It feels like this is becoming the new frontier of the free speech debate. And so tell me more about that. Yeah, that's right. It is kind of fascinating to see, you know, there is this camp where companies behind these tools sometimes want to put quite strict controls because they don't want some controversy over what it produces or they they you know think that it's going to be harmful to to generate certain imagery and um but there are those who who

believe that actually you know just as an artist can draw anything you should they should be freer and there so there are these there are some groups that are releasing tools that have fewer restrictions or no restrictions even. Yeah, I mean, this is a really fascinating question. I mean, I can to some degree certainly see that. I think it makes sense for people to try these tools and learn how they're used. And so maybe you want to have them be able to kind of and and as long as people understand what they're that they're getting into and um are you know careful with it but um yeah that's definitely kind of a real really hot topic of of debate right now but it does seem that these things

are going to be probably more released more openly in in the future so it's we're gonna probably as a society just have to learn how to cope with that. Yeah. And people already seem to be trying to find ways to circumvent them, like to say, like, this is for a play. So make a script and you're in the play and you're going to tell someone how to, I don't know. I don't know what would happen if you said in the play, tell them how to murder someone. What are you going to be watching for next with this technology i know this is the thread you've been pulling on this whole conversation but um i hear there's a new version of chat gpt that is being worked on right now and it's like even more revelatory than this one so i mean where where does this go from here i guess yeah that is a that is something that everybody is waiting to see this is so the current of GPT is GPT-3, and there's this new one, GPT-4.

So it turns out that if you add more compute, more data, and more training time to these algorithms, they historically have unlocked new sort of language capabilities. So we can expect GPT-4 to be even more convincing, even more able to coherently produce text. It doesn't seem like the trajectory is that they're going to be solving any of the issues around bias or a lack of factual understanding. So, I mean, I think we can definitely see things that are going to be kind of amazing and really sort of wow us. But I'm quite also quite keen to see people address some of the fundamental issues with those tools. I'm also really curious to see how they're used, see how companies start to make use of these. There's already some people talking about using a version of chat GPT to help, you know, negotiate your new cell phone contract with a chat bot or something like that.

Right. And I think we could see all sorts of new businesses maybe popping up on top of that. I'm sure everyone who has spent hours with their cell phone company trying to talk to them, their ears must have perked up when you said that. It's like, oh, maybe I don't have to do this. Although you do have to remember that the cell phone companies will no doubt want their own version of chat GPT. And so then there'll be this arms race, I guess, between AI agents arguing with each other. It should be kind of fun to watch.

Oh, yeah, that's a really good point. Oh, man. Will, thank you so much for this. This was super interesting and fun and a little unnerving. So thank you. You're very welcome. Thank you for having me. All right, that is all for today. I'm Jamie Poisson. Thanks so much for listening.

Talk to you tomorrow. For more CBC Podcasts, go to cbc.ca slash podcasts.