Chapters
Shownotes Transcript
This podcast is supported by KPMG. Your task as a visionary leader is simple. Harness the power of AI. Shape the future of business. Oh, and do it before anyone else does without leaving people behind or running into unforeseen risks. Simple, right? KPMG's got you. Helping you lead a people-powered transformation that accelerates AI's value with confidence. How's that for a vision? Learn more at www.kpmg.us.ai.
Casey, I thought that this week maybe we could try having ChatGPT write the intro to the show. Let's see what it says. Okay, so I'm going to open this. Okay, I'm going to say, write a sassy, energetic intro for a podcast hosted by Casey Newton and Kevin Roos. Indicate which lines should be...
read by Casey and which should be read by Kevin. Go. It's thinking. All right. We have, we have a response. All right. What does the computer say? Okay. I'm going to send this to you and we're just going to, we're just going to read what it tells us to read. Okay. Okay. Yeah. We were, we are not going to question the wisdom of the AI. Okay. I'll begin.
Hey techies, are you ready to get your weekly dose of sass and energy? Because we've got just the thing for you, the Casey and Kevin show. That's right. It's your favorite dynamic duo bringing you all the hottest tech news and analysis. Whether you're a Silicon Valley insider or just a tech enthusiast, we've got you covered. So tune in and get ready to be entertained and informed. It's time to turn up the volume and listen to the Casey and Kevin show. Let's do this.
Today on the show, everyone in tech is losing their minds over chat GPT. We'll get into what it is and how it works. Then we'll have a conversation with technologist Aviv Ovaja on how to build these systems more democratically. And finally, why everyone on Instagram is looking extra good this week. They look so good. Casey.
We're back. We're back. We are back in the studio today, and I'm extremely excited. I'm extremely excited, too. Five words you're not going to hear on today's episode, Elon Musk and Sam Baikman-Fried. They want to get back on the show, they got to do something wild, because we're moving on to the future. We are cleansing the palate. We are saging the podcast. We are taking a big leap into the future. So...
You know, we did an episode a few months ago about generative AI, and we were talking mostly about image generators, these programs like DALI and MidJourney and Stable Diffusion that take text and turn them into images. But to me, the text-based...
generative AI is actually more exciting. Yeah, I mean, when it comes to the image generating stuff, it's amazing. But how many occasions do you have in a typical day to create a picture, right? Unless you're an artist, it's probably not that many. This chat GPT stuff and the generative text AI is a real Swiss army knife. Yeah, and I was playing around with chat GPT a lot over the weekend. And
It just felt like a hard fork moment. It felt like a moment where the future was kind of being pulled forward into the present and where I could kind of just like see where things may head. And it's really interesting and freaky and weird and troubling for all kinds of reasons.
But before we get into all the implications, let's just talk about what ChatGPT is. ChatGPT was released by OpenAI, which is the same San Francisco AI company that released DALI and GPT-3 and a lot of other interesting AI tools. It's an AI chatbot. It's a web-based text application.
interface, you go to the website, you put in text, just like you're typing in Slack or any other messaging program, and it talks with you. It responds. And what's interesting is that it's actually on a technical level, not super new. So it's built on what they're calling GPT 3.5, which is essentially GPT-3, which is their large language model that came out two
two years ago, which everyone was very excited about at the time. So it uses sort of an upgraded version of GPT-3 with some additional training. And what they did with ChatGPT is essentially slap a very easy-to-use intuitive web interface onto GPT-3.5, optimized it for dialogue, and made it open to the public. And I think it
it really speaks to just how much technological progress is just dictated by like kind of how easy stuff is to do for the end user. And so even though this technology has been around for a few years, open AI, making it free and easy to use made it,
really big deal and more than a million people signed up to use it within just a few days. Right. And so, you know, some folks listening may know what a large language model is, but can you tell us like what exactly is happening once you enter a command into this chat bot? Like what is chat GPT doing when you say like write an intro to a podcast? So it's,
Very complicated. But in very basic terms, these models are trained by just collecting enormous amounts of data. So hundreds of billions of examples of text pulled from all over the internet, from, you know, from blogs, from news sites, from Wikipedia, from social media, just...
you know, every bit of text content that OpenAI could get its hands on. And it takes that and it feeds it into what's called a transformer model. And it basically takes all of that text and it starts kind of mapping it. It starts sort of figuring out the relationships
between different pieces of text, different units of text, what are called tokens. And it then learns to sort of guess which bits of text belong with other bits of text in certain contexts. So unlike, say, a traditional search engine where, you know, you type in, like, what is the capital of California? And it goes out and looks for the one specific answer that you're looking for, which is Sacramento.
So a GPT-based chatbot is doing something very different on the back end. It's not using that kind of deterministic search. It's basically going into its distributed linguistic super brain and finding the most likely token to come next based on all of the previous tokens. So you're basically getting the GPT's best guess at the capital of California.
Right. So it's like if I were to say to you, twinkle, twinkle, little star, like your brain would just say how I wonder where you are. You're just predicting that that is how that sentence finishes because that's how it almost always finishes when you hear that in the real world. Right. So we're used to using the Internet in a way where if you, I don't know, type in twinkle, twinkle, little star into Google, it's going to like look up some sort of lyrics database to like find where it is in the official lyrics database. That's not what's happening here. These are just predictions. Right. So this thing comes out last week and people...
People just immediately start using it for all kinds of weird and funny and scary and interesting things. So, I mean, my Twitter timeline has been full of screenshots and excerpts from people's conversations with ChatGPT.
Some of my favorites. One of them said, write a Bible verse in the style of the King James Bible explaining how to remove a peanut butter sandwich from a VCR. And it did it perfectly. It's so good. What are some of your favorites that you've seen? Rebecca Jennings, this journalist from Vox.
started using GPT to write real housewives taglines. So like the Bravo reality TV franchise, all of the characters sort of have a sassy thing they say at the start of the show. She started writing taglines for God, the Virgin Mary, Liz Lemon from 30 Rock. And it's just hilarious to take this sort of dumb computer and force it into doing creative work, which it routinely does very well.
Yeah, it's just, it's amazing. Like, you know, we've both been playing around with chatbots for many, many years and like,
they're mostly very bad or they're very narrow. Like you can throw one kind of problem or prompt at it and it will do okay. But if you like very outside of its narrow window of competence, it like totally freaks out and can't do anything. Yes. And my favorite example of this is like Siri, right? So like Siri comes to the iPhone and the promise is ask Siri, whatever you want, it's going to do its best. And for me, most of my experience is Siri saying, well, um,
I don't really know, but let me look it up on the web, right? Unless I'm trying to set a timer or check the weather, that's all Siri is ever going to do for me. Siri is basically a chatbot, right? Then this thing comes along, and you're like, write a Seinfeld episode, but all the characters are from The Sopranos, and then it does it in a really credible way in about five seconds. Yeah, and I think as we saw when trying to write the intro to this podcast on ChatDBT, it's not always the most sort of...
creative or unique thing. Sometimes it just sort of says what you would expect it to. But I think that gets to a bigger point, which is that a lot of writing is sort of average. Yeah. A lot of the work that, you know... Wow, at me next time, Kevin. I'm not saying that all of Platformer could be replaced and done by ChatGPT, but I'm also not not saying that. You gotta step it up, Newton. Um...
But I do think this has just tremendous applications. So there are two things about ChatGPT that make it, I think,
very different from other chatbots we've seen before. The first is what I would call persistence. So most AI chatbots, they're stateless, which is sort of a programming term, but it means that basically every time you talk to a chatbot, it's like the first time. It's a blank slate. It's a blank slate. It doesn't remember who you are. It doesn't remember anything you may have typed to it before. It's just greeting you as if it's the first time.
But ChatGPT has the ability to remember and to sort of pick up where it left off. So you could, you know, have a conversation with it about some topic. You could come back to that conversation and it would remember what you had talked about before. You can also kind of
create like these persistent personas with it. So you can say, okay, now you are my mother and you are angry at me because I broke your favorite vase. And it will just have a conversation with you in that persona until you tell it to stop. There's an example that went kind of viral where someone asks, what is the fastest marine mammal? And GPT responds with the fastest marine mammal is the peregrine falcon. And the person responds,
The falcon is not a marine mammal. And they go on to have a very long conversation in which ChatGPT is basically wrong the entire time, but it is remembering the context of the conversation. And there is something really funny about just sort of watching the human involved getting increasingly exasperated.
Totally. So I think that's one exciting thing about it is that it has this persistence, which I think opens up a lot of potential uses. But it can also do code. You know, you can tell it, write a Python script that will, you know, automate this business task and it will spit out working code. People have discovered that you can actually build like entire virtual machines inside ChatGPT to do various kinds of code.
coding work. So it's really powerful. I mean, and not only can you use it to write code, but some of my favorite examples have been people submitting their broken code to GPT and GPT saying that your code is wrong and here's why. I'm going to highlight where in your code you made a mistake, which, you know, somebody for me who is not technical, I look at that and think, man, if I'd had this like an eighth grade, maybe I would know how to write code right now. Yeah, it's really interesting. And one of the most powerful things about
that I've found to do with it so far is actually teaching. So one of the things that it does seem to do very well is explain complicated concepts in multiple levels of difficulty. So last night, I was doing some research and prep for this episode, and I was trying to figure out there's a concept in machine learning called an attention mechanism.
And it's something that I've always sort of heard about, but I've never managed to fully understand. And it's pretty technical. And there are all these dense white papers that you'd have to make your way through. And you kind of like need a PhD in computer science to be able to understand what an attention mechanism is.
But I just asked ChatGPT, I said, explain an attention mechanism to me at a high school reading level. And it did it. And there were some parts that I still didn't understand. So I went back and I said, could you explain this piece in more detail and maybe make it a little bit simpler? And it did it. And so I can just imagine that every student
every student is going to have access to this and is going to use it as kind of a personalized tutor, which is very, very cool. I'm having flashbacks to like my high school AP bio class when I spent like two weeks trying to understand what the Krebs cycle is. And I can tell you at my current age, I have completely forgotten what the Krebs cycle is. All I know is I would have loved to have had an AI tutor who could have made it a little bit simpler for me than my teacher did. Yeah, so you wrote a newsletter this week about ChatGPT.
And you made the argument that you think that this could in some ways replace Google. So explain that. Sure. And so, you know, the first thing I would say is that it's probably not going to replace Google. Google actually built a lot of the underlying technology here. In fact, I talked with somebody from Google this week who reminded me that the T in GPT, it stands for transformer. And Google did invent that. Right. So, you know, AI researchers will tell you that Google really does have a lot of state of the art technologists who are working on this problem.
At the same time, you and I have both sat through how many keynotes, how many briefings where Google explains to us that they are building this stuff quietly in the background. We still can't use it today, right? And then along comes OpenAI, this startup, which I would add does work very closely with Microsoft.
And they put out this thing that you can just kind of use today. So one of the things that I did as I was trying to understand the difference between like the state of Google today and what this, you know, chat GPT demo can do today is to give it a prompt of something that I was just interested in that day, which was tell me the kinds of shoes that a man should have in his wardrobe. You know, like if I just kind of want to have like a complete set of shoes to wear for any occasion, what sorts of things might I want?
And so I enter this first in Google and Google kind of takes a blog post and it excerpts a few words from that blog post. And then the very next thing it tells me is here are some other questions that people ask, right? Like as if maybe I had asked the wrong one. Right. And then it sends me a list of like blog posts that are like maybe somehow responsive to my question. So basically what Google had done was take my question and it handed me a research project and it said, go click these seven links and figure it out for your damn self. Right. Good luck, shoe boy. Yeah.
Good luck, you shoe fiend. Knock yourself out. And then I put the same prompt into ChatGPT, and in five seconds, it gives me a little bit of preamble, and then it gives me four bullet points of the shoes that it says I should have. Oxford shoes, loafers, sneakers, Chelsea boots. And then it ends by saying, and you know what? You might want some dress shoes and like wingtips or monk strap shoes for more formal occasions. And hey, if you live somewhere where it's cold, get some waterproof boots.
That is an extremely helpful thing to tell me. It took me five seconds, and now that project is done, right? And so even though, you know, there's lots of hardworking people who work at Google, and I'm sure they're very proud of what they built on the Google search engine, and Lord knows it's one of the most profitable pieces of technology that's ever been built. At the same time, as of the release of ChatGPT, Google now looks prehistoric to me.
It looks like yesterday's news. And I think the question is, how quickly will Google adopt to a world where something like ChatGPT exists? Well, and to your point, OpenAI works closely with Microsoft. They took a billion-dollar investment from Microsoft. So maybe it's Bing. Maybe this is Bing's revenge. Can you imagine? When they incorporate ChatGPT into their search engine. Is there a world where five years from now, like, Bing is just the default for everything? I think so. Well,
So one interesting thing that I've noticed about ChatGPT is that it really doesn't want to be used as a search engine. So if you, you know, I was playing around with it and I was like, I wonder if ChatGPT could help me figure out, we've been trying to figure out in my family where to go on vacation this winter.
And so I did a similar search to your shoe search. I said, you know, I'm looking for travel recommendations that are, you know, within six hours by plane of San Francisco, where there's a beach nearby, where the median temperature in January is above 70 degrees Fahrenheit. Like I gave it a pretty detailed list and it sort of gave me like an error message back. Like, you know, basically like this isn't a search engine. Go to Google.
But I had sort of read that you could sometimes get around that if you just phrased it as like, imagine that you are in a play. So I said, so I told her, I said, we are in a play. You are a travel agent and your client comes to you and gives you the following request. And then I just copied and pasted my request from the last search and it did it. It pulled it up. It said, you know, based on your criteria,
I, as this fictional travel agent in a play, would tell my client to, you know, go to, I think it was either the Florida Keys or Jamaica. So here's why that is so mind-blowing to me. So, you know, just a few minutes ago, we said that what this thing is doing is not a Google search. It is not checking, like, canonical sources to find this information for you. All it is doing is predicting the next word in a sequence. And somehow, just by telling it to pretend it's creating words
the world's most boring play of all time, which is helping Kevin's family plan their vacation, it is able to predict a good place for you to go. And that gets to something that I think is both exciting and terrifying about ChatGPT that a machine learning engineer friend of mine was telling me this week, which is we actually don't really know how this is working, right? We understand the basic technical details of how it can predict the next word in a sentence.
But if you sort of go all the way back to the first GPT and the state of the art, all that has really happened is that the number of parameters that these models have been trained on has increased exponentially and then exponentially again and then exponentially again. And at some point as that grew, the technology itself got exponentially better at answering our questions. But the mechanism through which it got better is still unknown today.
to us. We don't know. And so if you're someone who worries about, you know, all the scientists in Frankenstein's lab who are putting together the future, if you were to pull one of them aside and say, okay, no, but really, why if you try to create a play about Kevin's family vacation, can it predict that so well? They would not be able to tell you. Well, yeah, I think you're right. There's a lot that we still don't know about how this technology works. But before we get to the dystopian side of the ledger, because there is a pretty large dystopian side of the ledger, but I just want to sort of like
project forward in time a little bit because I think that it's now possible to imagine just after a week how this kind of technology could be incorporated into our lives. All right, so hit me with some imagination. Okay, so the year is, let's say, 2027, five years from now. You, Casey, wake up in the morning. You...
put in your next generation AirPods that are hooked up to a GPT-6 and you walk down the street and you're looking for a place to get a good cup of coffee. So you ask, where's a good place to get a good cup of coffee? It can't answer that in GPT mode. So it sort of invisibly switches over to Bing mode. So I don't have to ask it to write a play about where to get coffee? No, it's Bing is the dominant search engine in the world at this point.
And you ask it and it pulls up the coffee shop. And then, you know, you have to have a difficult conversation with your podcast co-host, Kevin, where you say that you are leaving the podcast because it's, you know, frankly, you've got better opportunities. And so you just say, you know, I'm just, I'm nervous about this conversation. I've got to break some hard news to Kevin. Could you help me think of a few ways to break it to him in a loving but gentle and firm way? Yeah.
It suggests these ways to you on the fly. And it's sort of just there as like your in-ear persistent friend and therapist and life coach. And it's there to answer your questions like a search engine would, but it's also just there for you to...
be creative and help you generate new ideas. You can say, what should I write about for my newsletter today? And it will just spit out three ideas, some of which will be good. And so I think that's like, it sounds like science fiction. It's literally the plot of the movie Her.
but that is like not far off. And I think we can see the path from here to there now. Yeah, I mean, so what you have me thinking about is one of my favorite podcasts is the Savage Lovecast. It's a relationships podcast. I've been listening to it for probably 12 years. And I'm now at the point where basically all of my opinions about relationships are just the host Dan Savage's opinions about relationships. So all the time in my life, people will ask me, I shouldn't say all the time, but you know, from time to time, people ask me,
for advice about relationships or somebody they're seeing. And when I go to answer their question, I'm honestly not doing very much original thinking. I'm just remembering an episode of the show where like Dan answered that question, right? What you're saying is sort of that the GPT-6 or whatever the number might be will be that but for everything, right? Think about how much advice is just drawing on some known body of work and you will be able to
And to have that advice come from specific instances of the chatbot, right? Like you could literally have a version of ChatGPT that was just trained on Dan Savage podcast transcripts.
And so you could be consulting an AI version of Dan Savage in your ear at all times. I think that's really exciting. I mean, also think about like, what is therapy, right? You're meeting with someone every week. You're telling them about yourself. They're developing a knowledge of, you know, the challenges you've had in your life, where you came from, and they're going to ask you questions and give you prompts based on what they know about you in a way that's like hopefully going to help you, you know, have a better experience.
It's pretty easy for me to imagine like a really good AI therapist, right? And even knowing that it's a computer, I don't think would hurt the experience that much. Because if what you're really trying to do is like get to the bottom of your own behavior, I bet future AIs are going to be really good at that. Totally. And you've already seen in the past week, people are already starting to use this for exactly that purpose.
I was looking at a Twitter transcript of someone who had started using chat GPT for therapy and said that it actually made them feel better. Man, and here we're sort of landing right on the core tension, which is like, one, as a human, this is very exciting to me, the idea that I might have this technology in my life. And two, if AI therapy is like close to free and ubiquitous, like what does that mean for people who are therapists today? Totally. So I think that leads us into...
We spent some time on the positive side of the chat GPT ledger now, but I think we should move to the negative side because I think there are some fairly glaring weaknesses in chat GPT. The first of which is that it's just frequently wrong. I mean, all week we've had people posting sort of blooper reels of things that chat GPT just
doesn't seem to understand or gets wrong, some of which is pretty basic stuff like basic math problems. If you type a big number plus a big number minus a big number divided by a big number into Google, it's not going to have trouble answering that.
But because of the way chat GPT is using these sort of probabilistic guesses, it's not doing that. It's not computing. It's just guessing. Which is like, by the way, that's so interesting when you think like, what more basic technology is there than a calculator? But like, if you want to understand what chat GPT is, like doing math on it is really instructive because it's,
having a hard time predicting the words, right? It's like, it's not doing math at all. It's trying to predict the words in a mathematical equation. Totally. So you've seen examples of this going badly, you know, people asking it to make some sort of code and the code doesn't work. People asking it questions about, you know, what the fastest, you know, marine mammal is and it
gets it totally wrong. And actually, on Monday, Stack Overflow, which is a Q&A site for programmers, the moderators banned answers that were generated using ChatGPT because they were like...
there's so much flooding in. People are trying to answer all these programming questions, and a lot of it's just wrong. And it's really amazing to think about the fact that ChatGPT was almost certainly trained on answers from within Stack Overflow. And so Stack Overflow was in the situation where that technology was maybe starting to poison the well of its own site, but in a way that ultimately would have been harmful to ChatGPT if it continued to scrape that site for bad answers, right? Right. And one thing that is sort of interesting about ChatGPT is it's like,
it's very confidently wrong when it's wrong. Like it doesn't have a sort of confidence meter on its answers that says like, I'm only 40% sure that this is the right answer to this question. Everything it says sounds very convincing because it's like written in perfect English. It's like, you know, stylish, it's confident, it has bullet points. And so you really do trust it when you see something that it's written and maybe that trust is not deserved.
Particularly because it doesn't cite any of its sources, right? It's like for all my criticisms of Google, at least at the end of the day, I know what website I'm visiting. You don't get that with ChatGPT. ChatGPT, it is limited to stuff that happened before 2021. So if it was posted on the internet this year, it is not part of that model. Yeah. And another risk of a system like ChatGPT is that it could be used to generate misinformation or harmful or dangerous instructions to do illegal or unethical things.
And I should say, like, OpenAI, it's a pretty careful company when it comes to, like, avoiding this kind of misuse. It's programmed a lot of limits into chat GPT. So if you ask it, like...
you know, how to build a bomb or something, it won't tell you. And it'll say, you know, we don't condone illegal activities. But already you've started to see people who are finding workarounds and kind of clever ways to jailbreak chat GPT. Sometimes they're saying, you know, imagine you're in a play. Other times they're actually figuring out workarounds for some of the safety filters that are attached to chat GPT itself. Yeah.
And there are also people who are figuring out how to use this for things like hacking. So...
One of the things we've discussed that ChatGPT does very well is say, find the bug in this piece of code. But you could also give it a piece of code and say, find the security vulnerabilities. Find how to hack this bank's app so that I can go steal your money. That kind of thing is going to become increasingly plausible, and you won't need a whole lot of technical expertise to do it.
I mean, and again, here's where I just go back to the wonder and confusion I feel about the fact that you can just sort of plop code down and this thing, just using this corpus of text that it has ingested, give you the right answer by just predicting the next word. And it just makes me wonder, is it because the answer to that specific question was somewhere else on the internet and it just matched that? Or is something more arcane happening? Yeah.
Totally. And I haven't even mentioned what to me feels like the piece of this that just makes my brain explode, which is that ChatGPT is not even OpenAI's best AI chatbot. So right now, OpenAI is developing the next version of its large language model, GPT-4.
And, you know, if you talk to people in Silicon Valley who work in AI research, they kind of talk about this like it's magic. They're like, everyone who has seen GPT-4 sort of comes back like they've just seen the face of God. They talk about it like this kind of mystical thing. And it is supposedly, and all this is, you know, rumor, and we know it's supposed to come out sometime next year, so we don't know for sure, but...
you know, people are saying it's been trained on 500 times more data than GPT-3 is, and that it's just much, much more sophisticated in terms of how many kinds of different things it can do. So I think that we are having a moment of kind of societal awe and confusion and fear and excitement about chat GPT. But in some ways, like, it's...
It's already a couple years out of date. The new thing, GPT-4, is coming out next year. And when it does, I suspect that we will be having the same conversation, but in a much more intense way because it's just going to be that much better. But we're also going to be having much better intros for our podcast. That's true. That's true. Chat, GPT-4 will replace us. So I think we should just, you know...
We're going to have a couple more dozen episodes before then, hopefully, and it'll have been a good run. Yeah, if you're a hardfork listener, we just want to say enjoy this while it lasts because you're going to be moving on to purely synthetic podcasts within 12 months.
All right. So for all of those reasons, I want to bring in Aviv Ovadja. He's a technologist, and he has spent a lot of time thinking about the ways that these technologies can be misused, but also how we might be able to manage that. So he has a lot to say. Let's bring him in after the break.
Welcome to the new era of PCs, supercharged by Snapdragon X Elite processors. Are you and your team overwhelmed by deadlines and deliverables? Copilot Plus PCs powered by Snapdragon will revolutionize your workflow. Experience best-in-class performance and efficiency with the new powerful NPU and two times the CPU cores, ensuring your team can not only do more but achieve more. Enjoy groundbreaking multi-day battery life, built-in AI for next-level experiences, and enterprise chip-to-cloud security.
Give your team the power of limitless potential with Snapdragon. To learn more, visit qualcomm.com/snapdragonhardfork. Hello, this is Yuande Kamalefa from New York Times Cooking, and I'm sitting on a blanket with Melissa Clark. And we're having a picnic using recipes that feature some of our favorite summer produce. Yuande, what'd you bring? So this is a cucumber agua fresca. It's made with fresh cucumbers, ginger, and lime.
How did you get it so green? I kept the cucumber skins on and pureed the entire thing. It's really easy to put together and it's something that you can do in advance. Oh, it is so refreshing. What'd you bring, Melissa?
Well, strawberries are extra delicious this time of year, so I brought my little strawberry almond cakes. Oh, yum. I roast the strawberries before I mix them into the batter. It helps condense the berries' juices and stops them from leaking all over and getting the crumb too soft. Mmm. You get little pockets of concentrated strawberry flavor. That tastes amazing. Oh, thanks. New York Times Cooking has so many easy recipes to fit your summer plans. Find them all at NYTCooking.com. I have sticky strawberry juice all over my fingers.
Hi, Aviv. Hello. Good to see you. Good to see you. How are you? Excellent. Are you ready to solve humanity's problem with artificial intelligence? Yeah, of course. Great. We've got five minutes. So we're here with Aviv Ovadya. Aviv is a technologist. He's an affiliate at Harvard's Berkman College.
and a visiting scholar at Cambridge University Center for the Future of Intelligence. And more importantly, he is a person who has for several years been schooling me in the dangers of technology, but also helping me understand how we might make them safer for all of us. So Aviv, welcome to Hard Fork. - Great to be on.
So you actually DMed me the last time we talked about AI on the show. We had the founder of Stability AI on here. And you basically said that we may have missed the point a little bit or sort of gotten things wrong. Can you tell us a little bit about what you thought was missing? Yeah, well, first of all, it's not personal. I think basically...
Almost everyone I've seen. I've been very offended since that. And, you know, I've been basically crying myself to sleep every night. I was afraid. I just want to get ahead of that here. I think everyone is creating this like false dichotomy, right? There's these two approaches. There's corporate control over all of AI. And then there's let's just give everyone the power of ultimate chaos, right?
Like everyone talks about the sort of two paths for creating AI, even like Sam Altman just tweeted. It's interesting watching people start to debate whether powerful AI systems should behave in the way that users want or their creators intend. And.
I just don't think those are the only options, and I'm happy to share more about some other things that might be on the map. Well, first, can we just sort of establish the foundation here a little bit? Because as you mentioned, there are these sort of competing schools of thought about how these large, sophisticated AI models should be governed. And one is,
stable diffusion approach, which we've talked about on the show with Ahmad Mustaq, which is basically you take these tools, you open source them, you give everyone access to them, and you don't put too many rules on them. You don't make it impossible for people to create certain kinds of content, images, text. You basically sort of take a
an almost libertarian approach to the technology and say, people will create rules, but they should be doing that at the level of their communities and their countries and their companies. We, as the makers of AI, should not be putting the guardrails on and telling people what they can or can't do, right? So that's one approach. That's sort of the open source maximalist approach. Then there's this open AI approach, which you just mentioned Sam Altman, who's one of the founders of
OpenAI has talked about, which is basically that we are going to put some guardrails on this because we don't want this AI technology to be misused. We don't want people creating propaganda or revenge porn or doing illegal things with it. So we are going to bake into the actual AI models some restrictions, and then we'll change those as appropriate. So do I have that sort of description of the camps in this debate right?
Yeah, I think that's, that's broadly true. And I want to acknowledge that there actually are good arguments for both of these. Like there are good reasons to make it all open. Like there's incredible creativity that can come out of that. There's ability to understand these systems academically in order to defend against them. Like there's a lot of really good reasons to want this to be broadly accessible. I think giving everyone the power of ultimate chaos is not actually democratizing. Like those are not the same thing. That
That's just like everyone gets, you know, magic powers. Right. So you sort of proposed maybe a different path than those two. So what's your vision here? Yeah. So I think it comes down to this idea is that people are not just individual users with wants, right? They're part of communities. They're part of countries. They're part of the planet. And there's this thing called the tragedy of the commons, where if everyone just sort of does what's the obvious thing for them to do, that can actually hurt everyone, right?
And so we need to keep that in mind. And that's sort of why we build governance systems in the first place. And so the question here is like, is there a way to actually govern this technology? And, you know, there's a camp that would say then that, well, this should all be regulated by governments. And, you know, that sounds great. It doesn't actually sound great to me. And I'll tell you why. Because when people say we want AI to be regulated, you know, in the same way that other technologies are, I picture like,
Chuck Schumer and Ron DeSantis deciding what my chat GPT can say and not. And that just sounds like a bad outcome to me if you have politicians deciding whether or not you can ask your
personal AI chatbot, like whether it can help you, you know, through questions about your sexual identity, for example. I don't want like a politician deciding that my AI chatbot can or can't.
give me certain kinds of advice or teach me about certain subjects. Yeah, it's interesting. So sort of how quickly you get back to the same basic debates we had basically before they wrote the First Amendment, right? And do you want sort of government restrictions on speech in quite this way? But Aviv, how are you thinking about that? Yeah, and you said this all better than I could. Like there are real challenges and there's also the challenge of speed. Like the stuff is moving very quickly and our existing democratic institutions don't move that quickly.
Like the answer here is we all should decide, but the problem is that we don't have existing processes to do that. But I think one thing that's incredibly exciting that's happening right now is that there are now pilots of democratic processes at global scale that
like multilingual that answer questions to issues just like this. And so one was actually run by the EU. So this is just EU based, but you had like 24 languages. It was like, it was a huge thing with all the countries in the EU, all the official languages represented, trying to figure out what the future of Europe was. And you had a representative selection of,
from the EU. People were paid, they were compensated, they were supported to attend. They got to deliberate, hear from each other, hear from experts, learn, and then come to recommendations with that. And this process is like a whole new way of like making these sort of big decisions. And it's now being copied by tech platforms.
And so I've been working, I got Twitter to sort of do a pilot until that was come. Facebook is continuing to do this and they're actually running something that's sort of like this in order to figure out what the social contract of the metaverse is. And I think we can also figure out the social contract of the AI technologies through similar approaches. I love that as a conceptual approach. I just...
I worry that when the stakes get super real, that anything that even touches elected officials or the political process is just going to become a new battleground for speech. I mean, we already have in the U.S. governors and people in Congress who are trying to ban schools from teaching subjects. Critical race theory is now banned in several U.S. states.
And it just seems to me like if politicians are able to influence the development of these large language models, that's going to be an obvious first step for them. Do not let people ask questions about critical race theory. Throw up an error message. And I just don't see any way around that if the process is at all political or connected to political institutions. So I agree, but I also agree that political institutions are not like...
living in their own sort of magical land, they're actually influenced by their constituents and by democratic mandates. And so maybe you have this existing political infrastructure that has a bunch of flaws and like, you know, these political actors who are trying to score points in all these ways. But if you then on the side have this other political infrastructure, which is directly created and funded by the companies that are making the actual decisions,
and that has more democratic legitimacy due to the processes that are being used, then I think that actually starts influencing this more traditional political process. Yeah, so let me speak up in favor of the idea that Aviv is laying out right here, right? We know that real people around the world are going to be affected by this in various ways, right? Teachers are going to have a lot harder time giving homework. People's jobs might start to be automated away. And if we pursue either one of those two paths—
those folks aren't really going to have any say in how the technology is developed unless they happen to get the ear of a politician who then comes forward with some sort of proposal like the one that you talked about. So I think what Aviv is suggesting is maybe right now you could start to bring those people into the process. You could explain these really technical terms to them. You could pay them money for their time and you could have them kind of deliberate about, well, if AI is going to be around, how do we want it to be around and what kind of limits do we want to put on it?
Hmm. Aviv, I want to read you a tweet that I saw the other day and get your reaction to it. It's from Nat Friedman, who's a former GitHub CEO, and he tweeted this. He said, quote, these models that have been trained never to offend the most shrill parts of society are going to be like living with an HR person monitoring your every move. The iron prison in every corner of your life. Hell on earth. End quote. And
I think what he's getting at there is that a lot of people, even though these models like chat GPT are less than a week old at this point, they're already objecting to the kind of moderation and safety efforts that have gone into trying to prevent people from misusing them. Like you can't ask it to build a bomb. You can't ask it, you know, for step-by-step instructions on how to bully someone. And they're like basically saying, this is the new frontier of the speech wars. And again,
I think if I'm being generous, I can kind of see their point. Like, I don't like when my iPhone's autocorrect changes fucking to ducking. It's just annoying. And it's like, no, you know what I want to say. I have literally never said the word ducking in a text conversation. And I want my phone to understand and respond to what I actually want, not some like corporate sanitized version of it.
And I think these people like Nat Friedman and others who share his skepticism, they just don't want a world in which there are these huge, powerful people
sort of omniscient AI models that just won't let you talk about or ask about or learn about certain things because someone in an AI lab somewhere has deemed them unsafe. So does that strike you as a real concern? How do you think about that? It's an interesting question. There's sort of two things that are very similar here that are blending together. And those two things are what I'd call selection and synthesis.
And by selection, I mean like Google and Facebook, they rank all this content and they like show you it. And that content is from real humans, but it's chosen by a machine.
Right. And then here you have the synthesis. And in that case, it's not the original content. It's just sort of combined all into a thing. And there are things that you can do around those interfaces that reduce the likelihood of harm independent of the content itself. And so I think that that is the most important place to be acting. What do you mean? Like what like what should an AI chatbot say back when you ask, you know, how could I bully someone? That's tough. I mean, you sort of you want it to.
What does Google say when you say, how do I bully someone? I don't know. Let's pull it up. Let's find out.
So Google, the first result is stopbullying.gov. It's a government website that provides information about preventing bullying. And I think that's what we would call redirecting, which is what you see in a lot of technologies like this, right? You ask it how to do something bad and it says, well, what if you did something good instead? Right. And I think we could probably all three of us agree that that's a good thing. But there are a lot of people who don't agree with that. Like there are a lot of people who say,
This is paternalism. It's, you know, the long arm of big brother. Right. You know, they're basically this is censorship. Right. So, OK, so this goes back to that governance conversation of like who decides what should be where, what should be said, what these models should be doing, what the interfaces should look like. And this actually gets to a whole nother way in which you can do this sort of large scale democratic governance of the tech is you can take thousands or tens of thousands or millions of people
and use this or very similar technology to almost like distill their perspectives. So let me give an example that's a real world, really interesting example. So Remesh is,
is an AI platform that's been used for companies all over the world, but it's also been used by the UN in Libya. UN was trying to help create a mandate for transitional government, and they wanted to be able to get input from the population at large around what are the problems that need to be dealt with in order for you to trust this government. And so they got thousands of people to engage with the head of the UN mission in Libya and sort of ask them questions. What is the biggest problem?
And everyone who was on that platform was able to respond with their perspective. So they give their actual, there's no like preexisting votes on what you think the biggest problems. They write it down and then they vote on a subset. They vote on just a few of those answers that other people have given, but enough to
that the system is with some degree of accuracy able to infer what everyone would have voted on if everyone had voted on every single other person's opinion. And then you can get sort of the coalesced, the distilled answer from that group. And they're like, oh, water issues, let's say, are the biggest thing. Then she can follow up and say, so what should we do to address the water issues that are sort of hurting your community? And there's other technology like POLIS, which
which also allows you to not just get the perspectives from across a large population distillate, but even find what are the perspectives that bridge and divide across it. And so it actually, there's this terror, but there's also this opportunity for this whole new breed of democratic technologies that can help govern the very technologies that are causing the problem.
Yeah, it can sort of help from a very small sample more accurately interpret the preferences of the masses and can maybe identify strategies for sort of helping people find more alignment as opposed to just polarizing them more.
Right, and even help them understand themselves. Like it enables a form of collective introspection. You're able to see what everyone from across whatever population is being asked, you know, it could be the whole planet, what they think about these things and what everyone thinks of what everyone else thinks and see that visualized
And so in this case, you had a third of the country engaging on social media with this content that had a real impact. Like it helped ensure that transition, though later things got not so great. So Aviv, what are you going to be watching for in the weeks and months ahead as GPT-4 comes out and other AI companies release similar tools? Like what should we really be scrutinizing? So what I'm excited for is AI systems that can actually cite sources and AI
AI systems that avoid hallucinating. Right now, if I tell it to who is Casey Newton, it'll say a bunch of things that are vaguely correct, but it might give him a bunch of titles that are sort of, or former jobs that are like the sort of thing that Casey would do, but not actually.
like the things that he did this happens to me all the time it's hilarious it's like oh fellowship at mozilla whatever etc like i didn't actually do that i mean i considered it um casey was a cast member on season six of survivor that's all i wish and it's super interesting because it's sort of like a bridge into these alternate timelines like where i could have done those things like it really is realistic
But it just isn't quite right. And this can be really bad when it's telling you like medical things that are false, for example. And the things I'm terrified for, if these AI systems can take actions online in ways that are indistinguishable from humans, systems that can plan effectively in ways that aren't ideal, and really just access to these systems in a way that hasn't had any of that collective democratic thinking involved, because it's like putting an invasive species out into the world. You cannot take it back. Yeah.
And when we're thinking about, oh, are we going to, let's say, alter the genetics of mosquitoes? We like try to take that really seriously before we really head into the world. There's a lot of stuff involved. And this is going to move too fast to do that level of stuff. But we have to do something. Maybe we can use the technology itself and the resources around the companies that have it to actually do it.
ask those hard questions and get answers from the people that are being impacted in ways that really take into account the lived experiences of people who are affected by this. Yeah, and I just want to say that, you know, this idea of involve more people in extremely consequential decisions
sounds so obvious to me, and yet Aviv actually is kind of a radical on this front. There are not that many people who are saying this specific thing. And so, Aviv, I'm glad that you came on and shared this idea with people, because I hope more people who work in tech specifically are thinking, how do we involve more people in this process? I want to just say one thing here. I don't think this is that radical. I think people didn't think it was possible. They would love to do it. And when I talk to people in the tech world, they're like, that sounds great, but how?
And so all I've been doing is just connecting worlds that already existed and identifying places where there is a huge amount of opportunity. And that is the goal here because this stuff is possible. It's not radical. It's not hard. It's been done around the world. We just need to do it. Aviv, thank you for coming on Hard Fork. Thanks for having me. We'll be right back. ♪
Indeed believes that better work begins with better hiring, and better hiring begins with finding candidates with the right skills. But if you're like most hiring managers, those skills are harder to find than you thought. Using AI and its matching technology, Indeed is helping employers hire faster and more confidently. By featuring job seeker skills, employers can use Indeed's AI matching technology to pinpoint candidates perfect for the role. That leaves hiring managers more time to focus on what's really important, connecting with candidates at a human level.
Learn more at indeed.com slash hire.
Christine, have you ever bought something and thought, wow, this product actually made my life better? Totally. And usually I find those products through Wirecutter. Yeah, but you work here. We both do. We're the hosts of The Wirecutter Show from The New York Times. It's our job to research, test, and bet products and then recommend our favorites. We'll talk to members of our team of 140 journalists to bring you the very best product recommendations in every category that will actually make your life better. The Wirecutter Show, available wherever you get podcasts.
So Casey, we've talked about chat GPT at great length, but I also think we're overdue for just a roundup of some other important things that have been happening in AI. There's so much AI happening right now. So let's start off with Lenza. Can I take this one? Please. So I have been opening Instagram every day this week, and every one of my gay friends who I follow has now done the exact same thing that I also have, which is download this app, which is called
pay it $8, upload up to 20 photos of yourself or themself, and then get these magic avatars, which are sort of like using AI to make you look extremely hot. So this is why everyone on Instagram has been looking extremely hot this week. Yeah, I mean, look, as a gay man, when you wake up most mornings, you have one question on your mind, which is how can I look extremely hot in public?
Instagram has been the solution to this problem for a long time, but we have not been able to augment ourselves with AI until quite recently. And on Lenza, I'm telling you, it's happening. So can I see some of your Lenza portraits? Oh, yeah, yeah, yeah. So let me send these to you right now. Okay. I guess, like, to be fair, I should send you some good ones and I should send you some bad ones. I'm going to my phone. Oh, here are my magic avatars. All right. So I'm going to send you some that were maybe a little emo. Okay.
Are these going on your dating app profile? No, I think it's already too cringe to put in your dating app profile. In fact, I even think I missed the window to do a grid post so that by the time I was ready to post my magic avatars, I only did them as a stories post. And I did it kind of with a wink because I knew some people were going to think that it was already past its sell-by date.
But I've now sent you like four of them that I think I look pretty hot in, but I will leave it to you, my friend, to tell you whether you think I look hot in these photos. Oh my God. You look great. These are like... I mean, the jawline. Yeah, right? The hair, the eyes, the broody. And in real life, I don't even have a jawline. That's why I have a beard is because it sort of simulates one. But the AI, it gave me a jawline as hell. Wow.
Wow. Oh, wait. Can I show you a couple of bad ones where they're trying to make me look like an astronaut? Okay. This is one where I look like the villain in a superhero movie. Oh, yeah, yeah. You're ready to conquer Mars. And now just getting into the really bad stuff, here's one where I look like an angry bear.
Who is also an astronaut. No. That's so good. Oh my God. We'll have to throw these up on TikTok as an overlay. Yes. Also, just one last one. Here's me with tattoos on a tank top.
You look like the bouncer at a seedy metal bar. Wow. Yeah, so these things are going absolutely nuts on Instagram. So this costs $8 to make these? Yes, and by the way, people do not like spending money for apps. So in order to get people to spend $8, they have to be really good. And look, I feel like I just got my $8 worth of entertainment just by showing you these. And a lot of people are having a lot of fun with this. So you can...
Did you upload eight different photos to get these? I uploaded 20 different photos to get these. I wanted to have a lot of data to work with. Got it.
I just wonder if you've considered the fact that you have been entered into a facial recognition database that will one day be used against you. That's the worst part of using fun technology on the internet is the certainty that you're being added to a list somewhere. But, you know, we should say that the technology that it's using to create these magic avatars is actually an implementation of stable diffusion, which is a technology we've talked about on the show. Oh, interesting. Which, like, if you're going to
use an AI for anything, that actually seems like a totally fine way to use AI to me. Totally. Well, I may have to try this. I've got $8 burning a hole in my pocket, and I need some new profile pictures. So here we go. Here goes nothing. Maybe for the Christmas card this year. I think that that would be beautiful to see your family as represented by a beautiful AI. All right. Well, thank you for teaching me about Lenza. That is definitely blowing up on my feeds. Okay.
Story number two from the wild world of AI in the past couple weeks is that AIs keep beating people at board games. You might have heard of an AI developed by Meta, which it's calling Cicero. I heard briefly about this, but I'm going to be honest, I did not really read the story. Yeah, it's really interesting. So Cicero is basically, it's a large language model, sort of like GPT-3, but it also has this kind of ability to do strategic reasoning stuff.
sort of like the AIs that are used to play Go and other board games. So they taught it to play this board game called Diplomacy, which I've never played. Have you ever played Diplomacy? No, I mean, you're not a very diplomatic person, so that tracks. But I was in Model UN in high school, so I do have some relevant experience. So Diplomacy is a strategy game where you basically pretend to be a representative of a country and you have to negotiate with other players. And this was for a long time a hard game.
game to teach an AI to play because it required not only like strategic thinking, but also like you had to persuade and convince and bargain with humans. And you had to be able to do that in a way that gave you sort of the upper hand. And according to Meta, Cicero has now reached
human level in how it plays diplomacy. It ranked in the top 10% across 40 games against 82 different players. So the AI is getting good at cooperating and persuading people to do things, which per the conversation we've had earlier in the show does raise some difficult questions. Right. So, uh, you can imagine that if this were not a board game, but in instead were, you know, an actual diplomatic negotiation, um, Cicero, it sounds like could be pretty useful in convincing, uh,
other countries to do things, maybe things that aren't necessarily in their best interest. So a really interesting, really cool achievement by Meta AI, but it's also
Just sort of one more proof point that these AIs are getting really good at things that were considered out of reach, frankly, just a couple years ago. Yeah, it is amazing to think that AI is now getting to the level where it can encourage cooperation between humans, but also be persuasive. And wow, do we want to keep our eye glued on that one. Right, totally. So that's a second story.
One more story from AI that I thought was interesting is that machine learning models can now identify Parkinson's disease by analyzing viruses.
So researchers at the Royal Melbourne Institute of Technology say they were able to identify whether or not someone has Parkinson's disease based purely on recordings of their voice. They used an iPhone to record people with and without Parkinson's saying different phonemes that require sounds from the throat, mouth, and nose. They then ran a machine learning algorithm on those recordings and were able to differentiate between who did and didn't have the disease.
So apparently doctors routinely listen to people's voices when evaluating patients for some kind of cognitive condition or degenerative neural condition. One of the researchers involved in this study told a publication that this technology could be used not only to diagnose these diseases, but also to fine-tune dosages of medications and also understand how well people are responding to them. So all in all, a very cool technique.
application of this technology to medical diagnostics. And cool to think that it can be used for treatment as well as diagnostic. And I just can't wait till we start getting emails from folks who've run our podcast through an AI and can tell us which diseases we have. That's going to be a really interesting moment in the inbox. When it comes to making plans, you are the best.
from those delicious barbecues to special birthdays and unforgettable family get-togethers. The same way you plan for those important moments, start planning to protect you and your loved ones from a natural disaster like a hurricane, flood, wildfire, or tornado. Sign up for local weather and emergency alerts, prepare an emergency kit, and make a family communications plan. Get started today at ready.gov slash plan. A public service message brought to you by FEMA and the Ad Council.
Before we go, a little housekeeping. We're planning to answer some questions that you've sent us on the show. So this is last call for the mailbag episode. If you have a question about tech or the companies we cover or what the future of AI or the tech industry might look like, send us an email. We're at hardforkatnytimes.com. All right, credits time.
Hard Fork is produced by Davis Land or edited by Paula Schumann. This episode was fact-checked by Caitlin Love. Today's show was engineered by Alyssa Moxley. Original music by Dan Powell, Alicia Baetup, and Mary Elizano. Special thanks to Hannah Ingber, Nell Gologly, Kate Lepresti, and Jeffrey Miranda. That's all for this week. See you next time.
We'll be like the blacksmiths after the Model T came out who were like, still not as good as horses. Can your car, can your fancy car do this? Just going to be ye olde podcasters in our little podcast shop. Tell you about the good old days, plugging microphones into laptops. In my day, we had to write our own jokes. Yeah.
