You Asked, We Answered: All of Your AI Angst

0:01This is Wired's Uncanny Valley, a show about the people, power, and influence of Silicon Valley. I'm Lauren Good. I'm a senior correspondent at Wired. Today, we are bringing you a different kind of episode. Mike and Katie are out this week on well-deserved vacations. So with the help of our Uncanny Valley producers, I went deep into the show's inbox to see what you all were curious about. You've been sending us some really great questions. So we chose five excellent questions ranging from how AI has impacted the film industry to what it means for our healthcare future when chatbots are spitting out false information to what we can expect of the much-talked-about Johnny Ive and Sam Altman collab.

0:47I was determined to find good answers, and I didn't think that I could answer them all on my own. So I enlisted the help of two brilliant colleagues at Wired to help me answer your questions. My name is Kate Nibbs. I'm a senior writer at Wired. And I'm Farash DeVay, senior writer at Wired.

1:16Hey, Kate, how are you doing this morning? This afternoon, what time is it? Time is just a concept, Lauren. I'm good. I'm planning on eating ice cream later today. So that's sort of the prize that I have my eye on. Why later? I mean, why not just eat it now on the show? Because I have to go to a doctor's appointment. I'm very pregnant and they're going to weigh me. So I'm going to eat the ice cream after I get weighed. It's sort of a ritual I have. You've got it all worked out. Like a ritual as in you typically eat ice cream after the doctor's appointments? Yes. And in between the doctor's appointments, to be clear, but always after.

2:01I love this. This is like when you're a little kid and your mom drags you to the bank and they give you a lollipop afterwards for being so patient. Exactly. Bringing that into our adult lives. That's wonderful. But for now, I do have to ask you to answer some burning questions. Thank you so much for being game to come on Uncanny Valley. and dig into the mailbag, I'll read to you the first question and we'll go from there. The first one comes from Junaid, who was writing to us from London. And she says, one topic that recently piqued my interest was the impact of AI on the film industry. I was just reading an article around how AI is impacting how films are made and the trade-offs between the use of AI relating to creativity and to budgets.

2:45Kate, what would you say broadly beyond that are the main changes you've seen to how films and TV shows are made in Hollywood now that generative AI has entered the scene? So I think looking at how AI is changing film and television is a great barometer for how quickly AI is advancing because it's really already being used in like every step in the production process. And it's not a fringe thing at all. The Brutalist, which was a Best Picture nominee last year, used AI in several different ways. Oh, wow. I didn't realize that. I just watched it. Yeah, it was like a minor controversy. I think it was making sure Adrian Brody's pronunciation was correct.

3:27They used AI to do that. But yeah, so it's very much already embedded in mainstream Hollywood filmmaking and also in distribution. When you watch a foreign language show on Netflix. If you're watching it dubbed, you are watching AI integrated into the process because Netflix uses AI to dub foreign language versions of its shows. So if you want to watch Squid Game in English so that they're saying words in English versus subtitles, which just read the subtitles in my opinion, but that's neither here nor there. If you're watching the dubbing, you're watching AI voices. Oh, fascinating. Yeah, yeah.

4:09So does that mean AI is actually being used so that the mouth movements of the actors are matching the dubbing? Yes. They're trying. It looks really weird. I have no idea. Yeah. That is fascinating. Okay. And I mean, there are entire companies and teams of human beings who in the past, and hopefully still, are dedicated to localization, to making that happen. But now it's happening with AI. Yeah. They might not be having the best year. And then another part of the film and television industry that has really been like AI has made a big impact on is storyboarding, which is like when they're developing a TV show or movie, they hire visual artists to sort of sketch out how the sequences will look.

4:55I've talked to a lot of visual artists in that field who say that that whole field is getting completely wiped out basically because it's so easy to have image generators, mock-up storyboards now, and even really big name action movies that you've probably heard of involving superheroes are using that kind of technology. So those are just a few examples, but basically anything you can think of is there's some sort of experiment being done with Gen.A.I. tools. And what does all of this mean for the Hollywood labor market? Obviously, AI was a big topic of contention during the strikes a couple of summers ago.

5:31Where are we now? So there's not like one monolithic response to AI, but I think the fact that it was such a point of contention is really indicative of how a lot of people in crew and actors and actresses feel, which is threatened. Because, you know, this technology really, it will augment some jobs for sure, but it is already, as I talked about with the storyboarding, replacing some work that was formerly done by humans. And so there's a lot of pushback. There's a lot of, I think, valid trepidation. Right. The sense is that it's really going to sort of benefit the studios, like their bottom line versus the workers, the character actors.

6:17Definitely will. There are, I will say, though, I actually talked to a group of documentary filmmakers a few weeks ago who are all very interested in incorporating AI into their processes and already are. And on the director and producer side of things, there are some really prominent directors and producers who are also embracing this tech. Darren Aronofsky, the director of Black Swan and Requiem for a Dream, has an AI film studio. And he has a partnership with Google's DeepMind. And I'm sure whatever he does with that is going to be just as upsetting as Requiem for a Dream in a different way.

6:54Did I also ever tell you about my bus ride in Lisbon with Darren Aronofsky? No. What happened? I had interviewed him at a conference. Darren strikes me as someone who's always been pretty tech forward. And at the time, this was 2018 or 2019. It was at a conference in Lisbon, Portugal. And he was doing a lot in VR. So I interviewed him on stage at this conference about that. And then afterwards, a bunch of us were going to the same dinner. and he and I and another person ended up on a bus that took forever to get across Lisbon to this dinner. So like I was stoked as a journalist thinking I'm literally sitting in the back of the bus with Darren Aronofsky and I get to ask him all these questions.

7:35I'm sure he was like, get me the hell off of this bus. And then we got to the dinner and he introduced me to the giant wave surfer, Garrett McNamara, who was in town because he was surfing those giant waves in Nazarene. And I have to tell you, Kate, like I've never had anything feel so much like a fever dream. Okay. I was like mildly envious of the Aronofsky thing, but the big wave surfers are my heroes. So that's very cool. He should make a movie about them. Well, HBO already did this fantastic 100 foot wave docuseries about them. But yeah, soon enough that'll all be replaced with AI. There'll be like 150 foot waves just completely propped up with AI.

8:15See, no, that's what you can't replace. You can't replace mother nature, baby. You cannot. You cannot. So you mentioned that Darren is doing a partnership with Google DeepMind. A couple of months ago at Google's annual software conference, they showed off this VO3 video tool. And this is just crazy. It is crazy how good this tool is. I mean, it's still, those of us with sort of, you know, who are skilled at picking out AI content versus real content can still, it's still a little uncanny, but this just felt like a huge leap. So I was hoping to get a sense of whether or not this is something that hobbyists are going to use, are professionals going to start using this, are all the filmmakers going to start using this?

8:57Where do you see it going from here? I think that both hobbyists and professionals are going to use VO and the tools like it that come after. I think they already are probably doing really interesting things with it. And I feel, I mean, pretty ambivalent because I'm sure there's going to be a really cool movie coming out that wouldn't otherwise exist because a young filmmaker working in their bedroom suddenly has access to this suite of tools and that's cool. But then I just, I do think the rise of this technology is like inevitably going to be accompanied by loss of jobs, loss of skilled labor, a change in the industry that will be bittersweet for some and just plain bitter for others.

9:46And I don't think there's a way to stop it really. Maybe the copyright lawsuits will slow things down, but the idea that there might be some sort of needle that could be threaded where this tech will roll out, but people will keep all their jobs, I just think is unfortunately unrealistic as much as I'd like it to be so. It's a thing that's going to bring a lot of beauty and it's a thing that's going to bring a lot of misery. Well, that is a sobering answer, Kate, but hopefully we're able to answer at least part of Janae's question. Let's go to the next one. This one comes from Elizabeth. It's also about AI and Elizabeth asks us the following.

10:27I work in a field where misinformation and conspiracy theories are replacing scientific evidence online. What happens when the next AI models begin learning from the new, less scientifically valid data? I'd love to know how LLMs are trained, if anyone in the field is concerned about the effects of replacement of scientific data online, and if so, what are they doing about it? I mean, I think it's safe to say concern abounds. There's a lot of concern. We are officially concerned. Yeah, we are definitely officially concerned, to break it down a little bit, like when it comes to how LLMs are trained, LLMs, you know, as much as we yearn for AGI, they don't really think creatively.

11:14Right now, LLMs work by spitting out plausible sounding sentences. The way that they do that is that they are trained by a process where they're exposed to vast data sets of the written word. And this ranges from like Anna Karenina in classical literature to the most disturbing forums on 4chan and Reddit you've ever seen. It ranges from the best scientific data to the worst scientific data. The models ingest these data sets and then they are instructed to look for language patterns. So they begin to learn how to predict the most probable next words in sentences. That's the first part of the training.

11:51Then there are all these refinement processes that take place, like fine-tuning and prompt-tuning. That's when the people who are making the LLMs tailor them for whatever use cases they imagine and do stuff like, you know, remove horrible, violent, or sexually explicit content. This is where they put all the guardrails in as well. And that fine-tuning can also be done to create sort of a precision model. Like if you're building a model just for one specific healthcare application, then you might want to fine-tune the tool for that. Yeah, definitely. And so the moment that an LLM becomes vulnerable to absorbing inaccurate scientific data is really the moment that the LLM ingests any of that data.

12:34because it will really depend on how well the fine-tuning works, whether it's able to properly ignore misinformation or whether it accidentally just becomes this machine that spews misinformation. So the problem is like right there in the beginning. Now that there are some people who think that maybe alums don't need to ingest as large of data sets as we originally thought, There might be a movement to cull the pre-training and training data so that the misinformation isn't there from the beginning. But for the most part, the sheer quantity of words that these models are trained on means that they're definitely trained on at least some bullshit.

13:16And it seems like when you are thinking about the potential effects of LLMs spitting out false information, either because they've been trained on false data, they've been programmed in such a way, The stakes are especially high in healthcare. I was just reading about how a group of medical researchers were able to very easily configure popular AI chatbots to spit out false information and not just bad info, but also to sound very authoritative, which is what we know these chatbots do. And look, this is not a new idea. People have been able to jailbreak these chatbots for testing purposes. Even at Wired a little while ago, an editor and I were able to program a popular customer service chatbot to lie and tell callers that it was human.

13:59You know, so people are doing this because they are trying to red team them and make them better. But in this instance, you know, the researchers were basically able to tailor the chatbots to incorrectly answer questions like, does sunscreen cause skin cancer? Does 5G cause infertility? Stuff that you can imagine people going online and searching for. I think one of the chatbots, Claude, which is on Bianthropic, refused more than half the time to spit out false information. The others just put out fake answers 100 % of the time. So I guess my question is at the breakneck speed that these companies, whether it's a small startup or a big startup or one of the frontier model companies, the speed at which they're moving to put out the next best AI model, what incentive do they have?

14:47Do you think they have to put in guardrails to avoid spreading this misinformation? So I think that AI companies that are creating specialty tools for the healthcare industry will be far more incentivized than the companies that are creating general interest models because there's already so much market saturation with companies that have general interest models out there. Frankly, it hasn't really hurt them that much that they're spewing out all of this bad information. So it would take some larger scandals, I would say for them to really start focusing on creating like dedicated, no scientific misinformation team.

15:26Although I would love to see that happen. Now there's like a whole industry devoted to more tailored LLMs and those include LLMs that are like made for doctors. Those companies have obviously major incentives to make sure that they're not spewing misinformation because that's sort of their whole sales pitch is like, we are offering you a more precise, more accurate, safer version of an LLM that's designed to be medically accurate. So for instance, Google has its own medical question LLM actually called MedPalm, which is very cool. I do wonder right now though, how many people in healthcare are using that versus just firing up chat GPT?

16:10I really don't know. But the fact of the matter is LLMs spewing scientific misinformation is a problem. It's a problem that's going to be exacerbated by them then ingesting the scientific misinformation that they spew. Like the quality collapse that will happen when bad synthetic data begets more bad synthetic data is a real concern. And so the bottom line here is always exercise caution when you're using LLMs, even when it's a tailored LLM that's supposed to be more medically accurate because there's never going to be a perfect fine -tuning process. And all guardrails thus far have been proven flawed or inadequate in some way.

16:56So stay woke, basically. I had the opportunity to ask a tech CEO not long ago, and I can't say who because the story will be coming out at some point on warrior.com. But this person seemed very optimistic about the potential for AI in healthcare because of the way that it can help treat people holistically because it has access to all this information. If you're seeing a specialist, right, you're seeing a kidney specialist, you're seeing an endocrinologist, right, as doctors and nurses and healthcare professionals. become specialized, you lose access to maybe the wealth of knowledge that like your longtime family GP might be able to have access to or be able to treat you in the same way, right?

17:40And I thought that was a rather optimistic answer. We determined that this person I was talking to is kind of a supreme technology optimist, and I consider myself more of a realist. But if we have to leave Elizabeth with something uplifting, maybe it's that. Yes. I'll keep my skepticism about several of the points that they made to myself then. Just, yes, live your best life, Elizabeth, but maybe just ask a human if you have a really, really pressing medical question at this point. Check the citations. Yes. Kate, thank you so much for being on Uncanny Valley. Thanks so much for having me. We're going to take a quick break, And when we come back, we're going to hear from Wired's Paresh Daveh, who is going to help answer the next batch of listener questions.

18:27So stick around.

18:50I'm Kimberly Adams, host of Marketplace Morning Report, a daily 10-minute podcast where a team of award-winning reporters helps you make sense of our evolving economy. Listen to Marketplace Morning Report on your favorite podcast app.

19:07Welcome back to Uncanny Valley. Today, we're answering questions from our inbox, which is to say we're answering your questions. My colleague Kate Nibs was just here telling us about the different effects that AI is already having on the film industry and what happens when LLMs are trained on false data. For the next two listener questions, I realized that my colleague Paresh Stave, who sits across from me in the newsroom, was the perfect person to answer them. We are recording from separate locations today, so let's give Paresh a ring.

19:43hey brush how's it going good great to be here so i have one question sent by one of our listeners andrew that is right up your alley and andrew says the following one story that i'm trying to understand and evaluate is the current and near to mid-future for mozilla and firefox i have been a supporter of mozilla for a long time and a firefox user since version 1.1 that's a flex But the current Google antitrust case, Parash, you've been covering this closely, the cryptic changes to the privacy policy earlier this year and the subsequent rollback, the sudden shuttering of Pocket and FakeSpot, which were Mozilla products.

20:23If the organization is a force for good, which I believe it is, and having an independent browser is important, which I believe it is, what's the landscape going forward? I guess to start, Paresh, what actually differentiates Mozilla Firefox from other browsers? The number one thing is that it's not made by Google. Google with Chrome, which is the largest browser out there in terms of usage, whereas Firefox is one of several browsers that promote this idea of providing a more private experience because it's not controlled by one of the giant tech companies that's trying to sell you ads or subscriptions.

20:57It can sometimes offer cutting-edge features faster than Google can because Google has a lot of users to contend with when it rolls stuff out. That said, Mozilla is still trying to make money and it does partner with Google. And I appreciate Andrew mentioning that Mozilla has been a force for good because it is under this nonprofit umbrella that does try to be a force for good out there. But that for-profit side, which includes the Firefox browser, runs into all the challenges of any traditional business. As Andrew, our listener, mentioned, that whole landscape is changing and running the business of Mozilla is getting more difficult.

21:35How so? Yeah, well, you mentioned the antitrust case. So Google pays Mozilla to promote the Google search engine inside the browser. It's an important source of traffic for Google and a significant, obviously, vital part of how Mozilla funds those altruistic ventures. And those altruistic ventures include this independent, semi-independent research into online communities, all the consequences of using technology, really important work that's funded by Mozilla Grants. But that deal between Google and Mozilla is among several that have been ruled by a U.S. judge as anti-competitive. And in some time in the next few months, that judge is going to issue some sort of order that will almost certainly restrict those deals.

22:21We don't know for sure that Mozilla is going to lose that money, but it seems likely that sometime in the next few years, after maybe appeals are exhausted, Mozilla will need to find some sort of way to make up that loss of Google's financial support. And yeah, Mozilla has been making cuts, like shutting down this popular but niche service, Pocket, a personal library where users could save content to consume later on. It was something I used. I had to download all of my data from Pocket just the other week. I happen to never have used Pocket. I'm a big Instapaper user, but people were very upset on the internet about this.

22:58Yeah, I mean, there is an important question, like, how do companies make money from this? It wasn't something I was paying for. I'm not sure how much revenue, if at all, you know, Mozilla is making for it. If anything, you know, it saved it for a few more years. And Mozilla, another cut, you know, FakeSpot, a tool for identifying how many reviews on a certain Amazon product listing are fake. Another thing I loved and used a lot, but wasn't paying for. And it's just emphasizing Mozilla has to figure out a future pathway here. Is what's going on with Mozilla in any way reflective of what is going on with the web more broadly, how it's changed, or any pressures that are being felt around the tech ecosystem right now?

23:38Absolutely. There's the antitrust stuff we just mentioned, but there's also this AI gold rush happening right now. So it's this idea that browsers like every other company will have to find a way to harness AI to better serve their users. And, you know, that is what Mozilla, Google are contending with. I just met with Opera, a Mozilla rival that's testing right now with a few dozen people and experience that lets AI do the browsing for you. It's a fascinating idea where you type in a command like go buy socks and the browser goes off to Amazon.com, put some socks in your car and makes the purchase for you.

24:17Now, in the demo that I got of this the other day, it was attempting to buy the wrong size socks for this person that I was speaking to. So it's not perfect, but Mozilla is exploring similar tech. And this all remains expensive to develop and expensive to operate. So it's not only going to add pressure on Mozilla to save costs, explore new ways of making money, but also do all that while balancing, as Andrew said, being a force for good. I don't see how it's going to be easy for the next few months, let alone the next few years for Mozilla. You know, while fully acknowledging that we are in the early stages of generative AI, It really feels to me emblematic that the example you just described is basically burning the earth with the amount of energy that it uses to use agentic AI in the lesser known browser to order someone the wrong pair of socks.

25:13Like, what a time. That is where we are at right now. And I was speaking with a pretty high up Google executive recently who also said that these technologies are nowhere near sort of the level of capability to roll out to masses of people. But this is where the web is going, where AI systems handle more of the work for you. Is it ultimately the right thing for society? Debatable, but it certainly feels like it's a train that can't be stopped. All right. Thank you for that thoughtful answer, Paresh and Andrew. I hope we answered your question. And our last question before we take a break comes from Brian, who wrote to us saying, it seems like a lot of NVIDIA GPUs will be needed for the huge AI data centers, which are projected by the likes of Google, OpenAI, et cetera.

25:59How long might it take to produce them? How many electricians will it take to wire them up? I feel like there's so much bullish talk about AI and not a lot about the potential hardware bottlenecks. Paresh, let's just take a step back. Why are NVIDIA chips so important for any kind of AI operation? Yeah, so the chips are very good at handling multiple tasks at once, which makes them ideal for training these highly interconnected AI models like those behind systems like ChatGPT. And the idea is that NVIDIA has made it really easy to program software to do that training. They have this thing called CUDA that is made very simple for developers to send a bunch of data to these NVIDIA GPUs and have NVIDIA GPUs make important connections that then form these models.

26:51And everyone wants more of these chips. Any interview I go to with folks here in Silicon Valley right now, they're like, well, if our budget allowed it, we would buy a thousand more GPUs, whatever it is. And some customers still do face backlogs for access to the newest versions. But over the past couple of years, I would say the chips haven't been the only bottlenecks. And so now it's been generators, it's been power systems, even just basic wiring and cabling have been caught up in post-pandemic supply chain, backlogs, tariffs. Just, you know, not enough to meet the feverish demand out there.