Last summer the coolest place in the house was in your freezer. This year it's time to level up. Reliance Home Comfort has over 155,000 five star reviews for delivering the type of outstanding customer experience in Indians of Canada for over 60 years. Right now, don't pay for 12 months on a featured air conditioner or heat pump.
Call in the experts that know how to beat the heat. Reliance conditions apply. See website for details. From NBC News in Washington, the longest running show in television history.
This is Mika Press with Chuck Todd. Hello from Washington. I'm Chuck Todd with special edition of meetup Press. This Sunday we're bringing you our best interviews on artificial intelligence as AI technologies improve and creep deeper into our lives.
By all indications, AI, artificial intelligence getting exponentially more intelligent and powerful. It has potential to do good like revolutionizing cancer detection. But it also poses great perils. Just this month, hundreds of top industry leaders of a short letter warning that unfettered AI could lead to human extinction.
In the next hour, we'll give you the good, the bad and the scary when it comes to AI. And I want to begin with JC Reese Anthus who argues we need an AI rice movement. And I started off by asking him to set the stage as America seems caught up in its how moment to reference 2001 Space Odyssey with the engineering of AI ahead of the science and our ability to grasp it. Yeah, there are maybe three ways we can build things.
We can engineer them, as you said, we can do them scientifically. If you hear about the olden days of AI in the late 1900s, it was about building expert systems, rule based systems that took everything humans knew and was there kind of a flowchart. We really understood from the ground up how they worked. Engineering is another level where you have a high level understanding, you bring some parts together.
I don't even know if AI is there. What we're doing now looks much more like growing living things. It's like gardening or raising animal or even a human child, you know, a superhuman, very strange human child. But we put in inputs to the process.
Process. We give it data. We say here's what the world is like. We give it some rules for learning from that.
So we say when you get a new data point, here's how you predict some outcome and change the way you think about the world. So you can predict the outcome next outcome better. And that leads to all sorts of strange systems. And there's a whole, there's a science of what we call deep learning, a science of these big AI models, but it's light years behind the engineering and the growing incubation side.
But the bigger conclusion I've come to over the last couple of weeks is that it's too late to do anything preemptive and that at this point we're going to be catching up and everything is going to be a bit reactive to the moment, which is never a good place to be when trying to set up a policy. I think that's right. Most technological policy ends up being that way. Fortunately, we've got kind of a big stick to wield with AI, which is we can slow it down and impose.
Are you confident we can slow it down? I mean we can slow down America. Can we slow it down globally? It would be challenging right now.
What's for better or worse is that there are a few companies that are in the lead in building at least the biggest, these so called large language models like ChatGPT. So insofar as that's a small group of people, they're incorporated in the U.S. they're beholden to U.S. stakeholders, including investors, but also policymakers.
I think that's to our benefit. And if we can do things to, you know, put more security measures that they have to pass before they come out of these new models, make them have to use certain hardware, track that hardware, ensure its security, anything we can do will kind of, you know, bog down the process of at least putting force farmers and stronger models and make them focus on what they should be focused on, which is making existing models better and safer and more aligned with human interests. It does feel like though, when you look at what's happening in healthcare, the advances are so exponential in the moment. I mean you talk to some of these healthcare scientists and they're just, they've never been more excited.
Like this is an amazing moment and it's all due to AI. So there are so many people now going to health advice online. For better or worse. There's a subreddit called Ask Docs that does verify the physicians who are on there giving advice.
But they're a push for time. They're not always experts in the issues that people are bringing to that online forum. So in JAMA Internal Medicine, a top medical journal, some researchers took verified health experts and had them evaluate the responses of physicians on that forum and compared it to what, GPT 3.5. So one of these new models would say in response to the people's ailments.
And it turns out not only are the chatbot responses rated as higher quality, they're also Rated as more empathetic. So they're able to spend more time saying, oh, I understand where you're coming from, that these physicians who are struggling to volunteer their time online aren't able to do well. The point is there are so many things, I mean the ability now for data collection to then narrow down. Well, let me check these eight things first.
No human doctor can know every potential outcome that a symptom may lead to. Right. Like it's just that's an impossible task task of any human. I guess my question is, is the excitement over the healthcare advances going to make slowing down AI a little bit harder?
What we need right now in healthcare isn't just bigger and more powerful and larger models. Because for example, when we're integrating them with physician. But there's some great researchers like a New York University who are trying to look on the ground, these are sociologists trying to understand when are physicians using these models, how are they using them, how are radiology departments able to make better assessments. What they want is more what we call interpretability or explainability.
They wanna understand those models because that will be what makes for a better healthcare system. That's what the, how the incentives are on both for companies and for the individual physicians. So I don't think they're just making them more powerful is what you need. We also have other concern.
So for example, with radiology there are lots of inputs into images that we say correlate really well with health outcomes, few of which are what actually matters. So you might for example, have a rich hospital and a poor hospital, they got different outcomes and they also have different imaging machines so the images will look different and the machine will learn that. So if we want to improve on that outcome, because we want to be predicting people based on their actual what's going on in their body, we're going to need more interpretability, we're going to need to be able to tune these systems. So I'm kind of optimistic in that sense.
What's most concerning is the companies that say that they're racing towards godlike AI or artificial general intelligence. And their goal really is to, you know, bring on this singularity as quickly as possible. That sort of frontier is what we need to most slow down. Congress is going to have to do this.
I feel like the only thing you might be able to get Congress agree to is to ban human like forms that produce AI outcomes. Right. I do think not allowing it in human form will cure the whole. There's this threat of deepfakes and all of that to me, that's like, that's only one tiny concern we ought to have with all things AI.
But I mean, what's a realistic policy you think Congress, like the one we have, can pass? I think a top one would be imposing steps that companies have to take before the release of one of these new models. Do you think Microsoft is irresponsible here? I've heard from a lot of folks that think Microsoft made a mistake here.
Microsoft is complex. You know, when people were starting OpenAI, its goals were very, I mean, open and beneficial to society. And now many people think that it's an arm of Microsoft. They think that that's pushed them towards the, you know, being holding the shareholders and that sort of thing.
I wouldn't have released these most recent models, GPT 3.5 and GPT 4, as quickly as they did. They didn't expect ChatGPT. I mean, that's an interesting one is, as you're saying, it's a human likeness. It's not the fact that ChatGPT is a much better system than previous ones, technically speaking.
It's that you can get online, you can use it, you can talk to it, and that's a way to take off the public imagination. So I think before they release each of these models, they are going too fast, and sometimes they take the steps that they need to. But you can have someone from the outside come in and say, you have to do X, Y and Z. And there are a lot of researchers, especially sociologists and people in philosophy and ethics in these fields who are building infrastructure to give you the audits that you need to run to ensure the models are safe.
How long is that going to take to get up and running, do you think? I think we could do it pretty quickly. And I think the onus should be on the people pushing forward the models to say, hey, maybe we'd have to wait a few months. I think the general concern about AI is all about our lack of concern about social media.
We didn't think about the negatives of social media. We didn't contemplate. Is there a chance here we're overcomplating all the negatives of AI? I feel like we're overcomplating them.
We see these AI winters that happen. There was one in the 70s and one in the late 80s and early 90s where, for example, when people were trying to solve chess with AI, they said, this is just impossible. They said the number of possible chess games, if you were to go through every pathway through the board that you could think of is more than the numbers of atoms in the universe. And that's just impossible.
We'll never get there. So you get the sort of, you know, sensationalism, both in terms of critique and sensationalism in terms of what those models can do. I just go back to that healthier example. The chatbot was producing longer responses than the physicians.
So in some sense it wasn't a level playing field. And you don't see people discuss that when they talk about that example just from last Friday. There are many others. You know, you might have seen that GPT4 is now at the 90th percentile of the US bar examination, so it could become a practice engineer in a technical sense.
But what does that mean for an AI who's trained on all this text? It's probably seen tests much like the bars games before, because it has the structure of our knowledge that no single human being has. And you tell me this. I'm not as fearful of the machines taking over for man in this respect.
The machine is still only as powerful as our imagination. I think that's where this engineering or, in fact, growing aspect comes in. It's a unique technology because we're able to make something more powerful than we imagine. Now we are able to give it all the text on the Internet.
When we don't understand the text on the Internet, we don't know what intelligence will develop from that. We've never read it. Do you think that AI will discover new languages? I was at a briefing last night where someone posited, what if the human body is a foreign language and we just haven't translated it yet.
A lot of people think that's what it's doing. So when we talk about what these models do, they learn mathematical representations of the world, whether that's words or images. They've got the formula that take an image of a healthy human and turn that into an image or an X ray of an unhealthy human. And we've peered into that, but it's very deep.
The frontier science hasn't gotten there. So when you say when I use the word language, you sort of say, yeah, it basically ones and zeros. It's a math equation. Yeah, yeah.
I mean, there is a discussion. So one of the. Again, talking about hype on both ends. Google was talking about their model that supposedly learned a language on its own.
It had never been exposed to. But it turns out that much like, you know, those GPT4 examples, probably in its training data, it had some of that language that Seeped in. So for example, some of these like training corpuses, as we say, they have extra words. If you look at a lot of English text, it has Spanish peppered throughout.
So that was kind of silly. They said it was only a language. On the other hand, you can do something called zero shot translation. So the idea here is you know how to translate English to Spanish and you know Spanish to French, but you've never seen a dictionary between English and French.
And the model can learn that because when it gets that high level representation of words, it knows something like what a car is mathematically. As long as it can connect the dots in some pathway, it can speak those two different languages together. That's something humans can't do. So the other argument for pressing forward with, with all of this research is that in order to get good at, there's a defensive aspect to the research.
Right. If you know the rest of the world is racing towards this too, and it is our geopolitical enemies, Russia, China, et cetera, then we kind of, we have no choice but to push the boundaries of this, even if it's to figure out what our vulnerabilities are. What say you? I used to be more concerned about that in the sense I thought it was more of a global arms race.
But since a few companies have, I would say, gotten close to an algorithm, you know, gotten so much control over training data, over the best engineers in the world, they're dominating the top conferences that academics go to, they publish this sort of research, that there's something very unique happening, not in the US not even in California, but in Silicon Valley itself. And the fact that that happens to be where we are means we have a unique opportunity to control that. But in China, for example, they're trying to produce chatbots, much like ChatGPT, and it's really hard. So, for example, like they don't want the chatbots to say certain numbers or mention Tiananmen Square and then they pull that out.
They're trying to teach their chat about culture, right? Yeah, it can learn culture on its own. But if we want to tell it, for example, take a bias against certain races or genders, we want to say, hey, take that out of the system. So learn from the language on the Internet.
But when you see people being toxic or hateful or discriminatory, remove all that. We have no idea how to do that. It's an incredibly hard problem. This was a debate point with a Google engineer that we had at a dinner recently and he said, well, you can't.
What we learned is in order for chat GBT to learn that Nazis are bad, they have to know who the Nazis are in the first place. So he goes, we entered this assuming, well, let's not teach them that the Nazis exist, let's not teach them about fascism, and therefore the ChatGPT won't know it. And that just turned out to be a false promise. People used to imagine very narrow systems that were trained on specific medical text, for example.
But it turns out if you want a system that's a really good doctor, it's going to first need to learn the English language from the whole Internet. And then you can do what they call fine tuning and have it specialized. But because of the Internet, images, text, everything is so blended together, it's really hard to exclude anything like that. And in fact, this is one of the issues that try to make these models safer because if we teach them what the right thing to do is, they can plicitly learn what the wrong thing to do is as well, because it's a duality of good and bad and that makes it especially dangerous.
And it's one of the hard technical challenges. We just don't know how to exclusively teach them to go in the right direction and not the wrong direction. All right, well, Chasey, I appreciate this. Thanks for having me.
You can listen to more conversations just like this one on the chatcast. It's available where you get your podcast. Subscribe now. We also like five star reviews and we'll accept four and a half star reviews.
We're releasing new episodes every Wednesday, Friday and Sunday. Coming up, AI is offering the possibility of living forever. Released in Hollywood, a new movie starring an AI generated James Dean. Not making this up, folks.
Could be coming to a theory near you sooner than you think. Is that such a good thing? You're listening to Meet the Press. Stick around.
Welcome back to a special edition of Meet the Press. So immortalizing our icons as the new Madame Tussauds Wax Museum. Holographic concerts are proof of that. But what if we could actually talk to these people?
On the latest season Meet Press reports, I spoke to two experts who are working on just that. Travis Cloy is the CEO and co founder of Worldwide xr. It's a global futurist company focused on technologies around AI, machine learning and more. You're going to film Back to Eden and it'll feature a digitally resurrected James Dean.
Kathleen Haas is a producer at the USC Institute of Creative Technologies Vision and Graphics Lab. And she's focused on the AI generated virtual humans. Her work By Way is primarily funded by the Department of Defense. We have been asking Kathleen to walk me through the difference between a deepfake and an AI generated human.
Well, a deep fake is a 2D technique and it is embedded in video and it's very effective. We've seen a number of them, they look very cool. But the AI based approach to face replacement is a different process. It is a CG face and in our lab we do a hybrid approach where we take a morphable model.
It's developed from numerous light stage scans and that gives us all the reflectance data, which means it can be like the skin. And then we marry that with the highest resolution 4D video that you can find. And you can do a little encoding and decoding over here and a little machine learning, encoding, decoding over there. Secret sauce.
Mix it together and then you output this pretty fantastic looking photorealistic CG face. And it's animatable, completely animated by texture, excuse me, by text, to speech or, or audio or anything. It's like using layouts and writing over it. It sounds like.
Right, it's a cut and paste a little bit, but the AI generated obviously you set a little bit more. Now you, we put it, we gave you a photo of me and you did an AI generation and you did it in 24 hours. We should give me a little more hair. So literally, so this, we gave you high sharp.
We're not a 4K resolution, for what it's worth. So how would you advise somebody to be able to figure out what's fake there? I mean, obviously I know it's fake with hairline and all that. Well, it, it's still up in the air, you know, as far as how you detect whether that's a deep fake or fake, are you putting watermarks in?
Are you guys thinking about this yet? We're thinking about it for sure. I mean, we're a research lab, so we're now just, we're developing a computer vision research lab at usc funded by the army mainly. And our, our main focus is research and publications.
So we're sort of, you know, for the greater good, generating knowledge in the topic. And we're putting together these really amazing, you know, capabilities. All the other parts of those is not really our department. And how long did that take?
Yeah, it took by the time they got it, it took a couple hours actually. A couple of hours you could generate, you know, with the right type of equipment, you can generate essentially a fake unit. Yeah, it's pretty good. Travis, you're about to make a movie with fakeun.
Look, you have the rights to this and all of that. Why do you think this is a good idea? You know, you know, we're a digital rights management agency, so we work with a lot of historical iconic states. And in working with the James Dean estate, who my business partner, Mark Rosler from CMG Worldwide has represented the family for four years, it's been the desire on behalf of the family to extend the legacy of James Dean, inspire future generations essentially.
And so in those conversations with the estate and with the family, we had the desire to work with them to extend that legacy to different mediums, whether it's film or virtual reality or experiential content for future. And over the course of time, we've had a number of film studios that reached out to us about doing James Dean feature films. And it was always a conversation of hiring a big celebrity talent to play that part. And so recently with all the, you know, innovation, evolution, this technology that we felt like creating the digital human of James Dean was the ultimate path to do that.
So how are you doing it? Is it just all of his every on screen role is put into sort of one data set, a data pool, if you will. And your AI James Dean pulls from that pool. It's a lot of copyright images and video assets that we have the family housing controls called the source material.
So it starts there, starts with aggregating all that source material, running it through machine learning to create that base model. And you have a lot of other types of innovation that you apply to it. So if you, James Dean, whatever role he played, I mean, what's interesting here to me, and this is sort of why I'm sort of stuck on AI humans, is that you're not getting the essence of who he actually was. You don't have the thoughts that we're going through his head then, do you?
No, we don't. No, we don't. But we have the family, which we try to make it as authentic as possible. You know, we do everything in ethical, responsible way as much as we can.
Working with the family, what does that mean? Leveraging books, leveraging a lot of the copyright material, just the family's consent, working with them to make it as solid as we possibly can. And it's a combination, like I said, different types of innovations. Motion capture, if we're worried about the body, voice synthesis, if it's worried about the voice, if we find people that have different.
Because James, James Dean passed away 67 years ago, so limited to assets that existed a Long time, just like you. Then you're deep fake or your digital human created with a video that wasn't 4K. But you do have 4K videos that Kathleen could have used. We're very limited with the material, the source material that we have.
What's interesting here is I want to put a source from Keanu Reeves. He is clearly going to try to. He doesn't want people to use his image for this. He said what's frustrating about that is you lose your agency when you get a performance in a film.
You know you're going to edit it, but you're participating in that. If you go into deep fake land, it has none of your points of view. That's scary. Then he added, it's also fascination, it seems, for us, the animals on the planet.
Like, how do we defeat death? You said Robin Williams also made actually put in his will. He did not want to have his image used after death. So more actors are thinking about this?
Yeah, more actors are thinking about it. Robin Williams put in his will 25 years of publicity. Nobody could use his rights of publicity for since the date of death. So essentially he's protected himself through including this language in his will.
If you don't specifically put it. There are other dead actors just sort of up the wraps. Not necessarily. I still think you need to go through the right channels to secure the rights of publicity agreements and licensing deals with those estates to leverage a lot of copyright material that it takes to create base modeling.
To some extent, yes. There are the bad apples out there that can leverage a lot of the existing IP that's on the Internet to create a digital human to somebody else. Kathleen, what are your ethical garbage? We have to adhere to our sponsors and deliver very specifically milestones and tasks in the Department of Defense.
So, you know, ethics, though, it certainly comes up and it's a concern along with all the wonderful, amazing, cool things that it can do. I think you have to have a full conversation about it. Our lab doesn't, you know, we don't make constraints based on ethics. We also don't try transition it out into the world.
Like, you know, currently. You know, I had some people who are like really upset at Microsoft blaming them for opening for the first, for putting it to the public before everybody was ready. Yeah, I've heard a lot about that too. Absolutely.
I will say this about the, the content itself, and I don't think, you know, James Dean is maybe the candidate, but there's. There's such a digital footprint that we all have now. And you Know, future, you know, I don't know how many years. Everything's moving so exponentially quickly.
It's very possible that you could gather a lot of data on any individual person with all the photos and the contracts and the letters and the emails. An AI based system would actually be able to. To pull out a personality from really, really quickly. Oh, yeah, I do.
I mean, Ray Kurzweil did that with his father to a certain extent as a test. About 60 John Adams. You think if we put all the letters that John Adams wrote. Yeah.
Yes. You could create AI? I think so. I think you could have conversational AI so you could, you could make the system and, you know, you could, you could have a conversation and it could be pretty much, you give me another rationale that said, hey, better that the good guys are first than the bad guys are first.
And on one hand I get that. On the other hand, is having an AI arms versus a good idea. I think we're. I think the cat's out of the bag is just a arms race.
I don't know. No sense wringing your hands. Well, I think that, that smart people need to get together from all walks, whether it be government or scholars or others, and come up with, you know, constraints and regulation in something. Yes, I do.
Travis, you convince the public wants to see that actors. I do. I feel like I wouldn't call them dead actors, but I think they're icons of the past reimagined in a new frontier. I wouldn't necessarily use terms like dedaptors essentially, but I think there's a desire to, as I mentioned earlier, just inspire the future generations by leveraging this technology.
My thanks to Travis Cloyd and Kathleen Haas. Season 6 by way Bepress Reports are releasing new episodes right now on Peacock, YouTube and wherever you get your streaming content, especially where NBC News is around, you can catch up on our full broadcast of Power AI, of course, but it's also where you'll find some of the NBC News strongest reporting on the race from Mars, the anti drag movement, Grifter Nation, and much more. Come up. I'm gonna talk to you AI Alchemist.
We'll explain why artificial intelligence powered tools could actually help journalists and educators. That's when Meet the Press returns. Hey guys, Willie Geist here reminding you to check out the Sunday Sit down podcast. On this week's episode, I sit down with one of the biggest bands in the world, Mumford and Sons, as we get the boys together to talk about their new number one album, Prize Fighter, and the evolution of that irresistible foot stomping sound.
You can get our conversation for free wherever you download your podcasts. Let's kick start your wellness journey with the crack today at Workouts Meal Plans. It's your fast track to a healthier you. And now during the Xfinity Member celebration members to get an exclusive 50 off annual subscription.
Head to xfinity.com membership to learn more Xfinity Imagine that subscription automatically redundancy here at 6005. 99 for taxon fees until cancel on the rims May 202026 prices subject to change. Visit today.comxfinity for full on returns and details. Welcome back to a special edition made of the press chat GBT da e Sable Diffusion these are some of the new big names in generative artificial intelligence tools.
Those are AI tools that can create text, images and even voices. And according to my next guest, the work created by these tools is something new. He calls it synthetic media. And he warns that the division between an original piece of content and synthetic media is simply vanishing.
Anders Grimstad is head of foresight and emerging interfaces for Shipstead. It's a media company based in Norway. Shipstead owns companies that produce news, commerce and marketplaces for everything from cars to baked goods. I asked Anders how we should be thinking about the regulatory space as artificial intelligence develops.
Maybe to answer that question, it's good to give a few examples of what it can sort of do. So let me explain a couple of differences from like classic AI and some of the experiments that we run. So classic AI, as most people think of it, is more about predictions. It's about finding patterns in large amounts of data to be able to predict things.
It's no different than any just taking a bunch of simulations in the past and saying, you know, and so what you're really looking at is, oh, there's a 98% chance this is it, it's a 94% chance this is it, etc. Right. Exactly, exactly. Now with generative AI, the difference there is that the models are trained on human outputs.
So texts, imagery, videos, 3D models, all that kind of stuff. And they find patterns, but you can kind of instruct them to recreate human like content. And that content can be very, very convincing. It can be almost indistinguishable from what humans create.
And so when we first started out experimenting with this, the first thing we built is a silly thing for a hack day, but we built a Tinder profile generator that generated fun Tinder profiles that could probably convert. Okay, then we set it to automatically create weather tweets from abstract weather data. So it's able to transform data into weather updates with only a very few examples of how we should do that task, which I'll get back to. And thirdly, we tried to see whether we could take an article that was written by a human, a journalist, and create versions of that article that was better tailored to different target audiences, like young people that have different tone of voice, or older people that don't understand the complexity of technologies, that kind of stuff.
So the way that we're sort of looking at this is in the short term, these models are able to mimic human outputs. They're able to produce content that's very human like, but they mess up. They sometimes forget the instructions that you give them. And they will incorporate things that they've been trained on that's completely irrelated.
So they will kind of hallucinate. It's the term that people with AI and I using to describe when models sort of mess up and include non original type of stuff. But the really interesting thing for us is these tools. Even today, the drafts that they create are really good.
They're almost so good that you could just say we should just automatically publish this kind of stuff. Which means that there's no kind of. We don't do that, by the way. But we could in theory for some of this stuff.
But the issue becomes, how do you do this thoughtfully? Do you do it? Do you use this as a tool with humans in the loop to sort of give our journalists, our writers superpowers? Or do we push it the furthest we can and try to automate human tasks?
Why is it neither or. I mean, look, I look at it as what a great new research tool, but it's no different than Google. It's just me doing a search and the search output is just more refined than ever before, I would argue. It's funny just listening to you.
The simplest hardware all seems to me is, well, you know, anything that's published by artificial means should be labeled as such, pure and simple. It's going to be about labeling. It's, I think, no different than the fight over GMOs when it comes to food, right? Where really labeling is what people care the most about.
Hey, this was synthetic, created to make up, you know, we're doing this in the food space. But labeling seems to be the thing that I think that matters most here because I think what we're all concerned about is the miss and disinformation aspect. In an ideal world, if that was easy to do. That would be one tool to make sure that the, you know, that's the same as kind of blockchain technologies promise.
You need to prove the origin of the content. One of the tricky things is a lot of this stuff is open source, meaning that I can run it on my computer, on my own gpu, I can run it on my phone because they're so optimized and having guardrails in place that covers, you know, if it was controlled by one or two big companies, it's easier to regulate. But this is extremely distributed even now. And Even though it's 30 days, it's not that easy to label whether something has been synthetically created or not.
And even OpenAI launched a tool, I think it's two days ago now for fact or for kind of trying to assess whether a text was written by an AI or not. And the accuracy of that tool is I think around 20%, which means that it's not really usable for anything at this stage. It's a promising tool, but it's not really usable for anything. And I think most people within AI think that this kind of delicate tools to identify what other AI tools produce is kind of.
You will never catch up to the speed and the change. So it's a very, very kind of tricky question. And I think on the one hand the extreme is to ask can we label things? And by that you sort of create this binary that a human created piece of content is real and the artificial intelligence piece of content is fake in a way.
Well, that's not the case. So I take your point. But do you think we're ever going to get to a point where a computer is better than a brain? I don't think we will as long as humans are the ones making computers and programming correct.
That's a tricky one. Also not my area of expertise. What I do think though is that computers can make us better at certain tasks that I completely. It seems like AI is an assistant, right?
Is whether, I mean, I think about, think about whether it's trying to identify a problem in a vehicle. Look, there's some journalism that should be automated sports scores. By the way, a lot of sports sites do this already. You see it truly at places like the local sporting events where they don't have the people right to actually go out there and they're writing an article based on numbers.
They didn't actually see the field. They didn't see the interaction between the players. They didn't see the emotional ups and downs. Right, like that.
To Me is what we know is going to be missing. That's why I believe that stuff never is truly replaced. But I see it as an assistant and there I think we're very much on the same page. So our sort of.
The reason why we're pursuing this and the reason why I'm personally very excited about this piece of technology is because, you know, the pain that we have today, and I think this is pretty common across media companies all over the world, is that we struggle to engage with people, a broad set of people. And we also struggle with the fact that there are ever increasing number of interfaces that we need to publish content to. And the reality is that even if a journalist writes a great story, you kind of, he or her has to pick a format in a way and every time you have to duplicate that into a different version or different content. It's a very costly thing to do.
And if you can leverage AI to take the original piece that was written by a journalist and then transform that into different formats and for different target audiences, that's a hugely powerful tool. No, I see where you're going there. I mean, being able to write for 15 to 25 year olds versus your 70 or to 90 year old. I see the power in that, in being able to dial up or dial down the complexity of a story to an audience.
I guess where the slippery slope is. The person behind the programming of that artificial intelligence program has their own set of biases. Yes, right. So how do you, how do we get that out?
That's a huge issue. Right. So I was actually a very interesting discussion about this the other day and it was something I haven't thought about. But I mean the reality is that if you take a step like these things are trained on the open Internet and they're trained on, you know, pieces of information that's primarily, primarily from English speaking languages.
So industrialized countries, predominantly white and that that in itself is predominantly male. Right. You think about the literature. I've seen a lot of people mess around with chatgpt by, hey, how could Shakespeare write more Shakespeare today?
Right. Like Ernest Hemingway producing novels about Vietnam now and perhaps about the Afghanistan war, the Ukraine war. Right. I mean I have teasy here, but I'm sure somebody's thought about, huh, can we do all quite on the Western front, but just program in what's happening in Ukraine?
So much of what powers AI are patterns from centuries of writing and intelligence that we're at a huge bias and that is a huge issue. Right. The training corpus is not extremely diverse take ChatGPT as an example, or the work from OpenAI, which is really great, I would say, by the way, but I mean, the inherent bias that's in the training material is difficult to avoid. And if you're someone from, let's say, Zambia in Africa using this, you know, the way that things are written, the way that things are, even if they're translated, will be still biased from, you know, the societies where they came from.
And right now we're arguably already in an existential cold war between essentially the east and the West. Right. And the east has believes that history went a certain way and the west believes it went another way. It's tricky, definitely tricky.
But it's also, again, I'm super optimistic about the upsides because it democratizes access to not. I mean, we're talking about writing things, but there's obviously ups and downs for imagery as well. Everyone knows how artists are struggling with this and copyright laws. But if you kind of put that aside, the fact that these days, with these tools, anyone can create a painting that in itself is in the image of Picasso.
Right. Like, I mean, I mean, is that, is it a health good thing or a bad thing? I don't know yet. So I think it's, it's, it's a, it's both.
But it is a new way for us to express ourselves that was previously locked to people that had the, the sort of grit to go through thousands and thousands of hours to learn how to do that stuff. And, and I mean, that can be said for many technologies that we use, you know, for, take for granted today. But when you look at this and you think about all of them already, think about all the authoritarian manipulation we've had, you know, from Russians, from Chinese, from Iranians. If you wanted to program it for that, how easy could it be?
Quite easy, actually. I mean, that's, that's the power of. These models are really, really, they're, they're really powerful. The weather example that I gave to you earlier, just to illustrate that point, the service that we built, it's not in production.
It was just for fun, but it could have been. The way that we built it was that we took publicly available metadata and we took four tweets from one of the public weather services in Norway, and we used that as training data. And with only four examples, GPT learned that the next time I get a new data input, I will create a new tweet. And they are good.
And that's four examples. There are like, literally for many of the things that you want to do. There's zero barriers to entry. So if I were to create a bot that read papers from a certain source and then wrote a summary that's biased, that's super easy to do.
I could create a fake Twitter account. So what worries me is, of course sourcing. It will be more difficult to distinguish between what's good and bad content online. But I am kind of.
I don't like this binary between everything AI is. I'm with you. Look, I get that. It's like I was thinking about this moment and I was comparing it and it made this video a little bit before your time.
The first time we cloned a sheep and everybody went, oh my God, that's amazing. And oh my God, are we gonna clone humans? That's tragic. And everybody freaked out.
And I would argue we probably went overboard in limiting some medical advances. I agree. When I think about this, there's like three aspects of it that really sort of intrigued me. And one of it is like our ability to produce things.
It's gonna skyrocket where we talk about text and imagery, but there's video, there's 3D. And I'm super excited about the ability to create 3D space. I describe a world and I will happen happen for gaming or for virtual reality type experiences. Then there's.
What you can highlight here is the kind of research aspect. You can use this to help you find relevant things. And I have a colleague who likes to call tools like Perplexity, for instance, or ChatGPT, like the billionaires assistant. You know, they have like people that read books for them and then they recommend it forward.
I think Bill Gates is a great example there and I would love to have that. We're almost affording the answer to questions because it's like that's the only. Because we don't know the answer. And this feels.
And I know the unknown always feels scarier than maybe it actually will become. Right. Like, we know if people feared the advancements of airplanes, the advancements of this. Right.
You know, but I guess it really is the. I don't know what the guardrails look like. And it sounds like other than human beings promising to do it ethically, you don't know where they are either. No.
And that's the tricky thing. Right. One thing that's scary is the availability of the stuff. Can't undo it.
It's already out there. Right. Like, so the idea. It's too late now whether we think the public shoulder have this tool in their hands.
They do. This will be like a learn by doing or learn by observing how people mess up. This was a great conversation and it's exactly what I wanted just to sort of it was both high level and trying to be admitting what we all know yet. Thank you for inviting in.
I really appreciate these discussions. I think this is exactly what we need to do in multiple different channels and formats. So my thanks again to Anders for that. In a moment, the number of companies capable of building a true standalone AI system is much smaller than you'd expect.
And for now there's little regulation oversight on how they deploy that technology. NBC News is Jacob Ward in conversation with Google's former CEO. That's just then at the Press as the day wraps up, get to scoop on what's been happening with here's the Scoop, the new podcast from NBC News with me, your host Gabby Basugian. We'll take a deep dive into today's top stories of NBC News's trusted journalists.
It's a fresh take that's sharp, thoughtful and it's informative, bringing you closer to headlines and conversations that are shaping our world. The Front Page the Zeitgeist. Here's the scoop from NBC News. Listen daily on Amazon Music.
Welcome back. My colleague Jacob Ward is NBC News's top correspondent on tech and AI. He's also the author of the book the Loop How Technology Is Creating a World Without Choices and How to Fight Back. In his piece from the Press Reports, he talked to former Google CEO Eric Schmidt, an advisor to top AI builders.
Here's part of that exchange. My concern with any kind of premature regulation, especially from the government, is it's always written in a restrictive way. What I'd much rather do is have an agreement among the key players that we will not have a race to the bottom. You've described the need for guardrails and what I've heard from you is we should not put restrictive regulations from the outside.
So if policymakers who don't understand it, I have to say I don't hear a lot of guardrails around the industry in that. It really, just as I'm understanding from you, comes down to what the industry decides for itself. When this technology becomes more broadly available, which it will, and very quickly, the problems are much worse. I would much rather have the current companies to set define reasonable boundaries.
It shouldn't be a regulatory framework. It maybe shouldn't even be a sort of a democratic vote. It should be the expertise within the industry help to sort that out. The industry will first do that because there's no way a non industry person can understand what is possible.
It's just too new, too hard. There's not the expertise. There's no one in the government who can get it right, but the industry can roughly get it right and then the government can put a regulatory structure around it. Logic.
That was a pretty, that's a pretty erratic answer there. I'm sorry, I don't know how else to react to that. I'm sorry, nobody in government, you won't understand what we're doing. So you don't have any input.
Boy, that is anorexic. A little bit. I think if you push it. I don't mean it that way, but it certainly comes across it like, you know, I would say that what he said, there is an earnest belief that it really does boil down to expertise.
I think he does not believe that they have the expertise of government to really get involved in this at this stage. But I would say that across the spectrum of experts spoken to, that is one of the major concerns that there is a exclusion of other people from decision making process because only the people that afford $100 million per copy technology should be in charge of the goal effect us all. So something that I have been working out in my own head is this. I think we're all doing the skeptical approach to AI because we didn't with social media collectively as a society, how much of this are we over?
Is it possible we like you and I and the skeptics, are we overcorrecting because of, oh my goodness, we didn't have any of that input or is this the right reaction? So I'm so interesting that you know, Eschwe is speaking to there. He said to me, he said to me before, he says that he was naive about social media back in the day and yet he believes that he has some sense of how we should proceed. What all of the people that I'm speaking to agree on, whether they are for or against the AI moment we are in, is that it is the unfettered access to data that fuels all of this.
And part of what that is going to do create the same little coterie of very powerful companies that collected data on us for the last 20 years. They're gonna be in charge of AI for the foreseeable future. The Internet didn't have this. It wasn't, there wasn't a feeling that only one company control the Internet.
Well, that's right, because AI is so much different. It does feel as if it's a handful of companies that have this control. The Internet felt like we had pizza. Well, that is because the Internet was in fact created for the benefit at the time of the military.
It was built inside the government and the people who refined it were academics. Here we're talking about something that was once upon a time, at least in theory, academically. But it has really hit the road only when for profit companies put billions of dollars behind it. I would argue that it's one of the first times in human history that something that's going to affect us as deeply as anything ever has is entirely in the hands of the people who are poised to make money off it.
Why? An important point to end up there, Jacob. War. I know you've been passionately following this for almost a decade now.
Here's the moment. I keep it. I really appreciate that. So that's all we have for today.
This is a real recording of Chuck Todd, not an AI creation. I promise you that. So thanks for listening. We'll be back next week on the radio, on your headphones and on your local news station as if it's Sunday.
It's me, the person. Hey, it's Kate Snow, NBC News anchor and host of the Drink. This month, Demi Lovato is my guest. The global superstar tells me that she is the happiest she's ever been right now.
But getting there, it wasn't simple. Demi opens up about starting in Hollywood young and why she now thinks she may have started too soon. She talks about recovery, her new marriage, and the deeply personal reason behind her new cookbook. The Drink is always about the journey to the top, and this was an honest conversation about what that takes.
Hope you'll listen and follow the Drink wherever you get your podcast.