Senator Ted Cruz: A heavy-handed, nanny-state approach, much like the European Union, would cripple America's leadership in AI.

Brooke Gladstone: GOP lawmakers are divided on how to regulate artificial intelligence. From WNYC in New York, this is On the Media. I'm Brooke Gladstone.

Micah Loewinger: I'm Micah Loewinger. The question is, will China's DeepSeek, oh so cheap and transparent, pop America's AI bubble?

Senator Ted Cruz: It's always been this idea that America has built these big, beautiful, large language models that require all of this money based on their narrative. There's not actually been much behind them.

Micah Loewinger: Also on this week's show, an AI music generator gives a composer a run for his money.

Mark Henry Phillips: I don't want to admit it, but it's good. Like really good. It feels real.

Brooke Gladstone: It's all coming up after this.

[MUSIC]

Micah Loewinger: From WNYC New York, this is On the Media. I'm Micah Loewinger.

Brooke Gladstone: I'm Brooke Gladstone. I'm not breaking any news here when I say that artificial intelligence has found its way into almost every aspect of our lives. A recent survey found that 86% of students use AI and more and more people are using it in their office jobs. People are falling hopelessly in love with AI powered chatbots or falling down AI-generated conspiratorial rabbit holes.

When it comes to jobs, whether you create content, assemble cars, or even drive them for a living, you're probably scared, especially if you're young and have the most time ahead of you. One recent study found that 52% of people aged 18 to 24 have job worries related to AI. Before the Senate removed it altogether, there was even a provision about AI regulation in the Big Beautiful Bill that split the MAGA ranks.

Congresswoman Marjorie Taylor Greene: It would take away states rights to regulate or make laws against AI for 10 years.

Brooke Gladstone: Georgia Congresswoman Marjorie Taylor Greene.

Congresswoman Marjorie Taylor Greene: I think that's pretty terrifying. We don't know what AI is going to be capable of within one year, let alone 10 years.

Brooke Gladstone: On the other side of the issue, Senator Ted Cruz.

Senator Ted Cruz: There is a real danger of state legislation adopting a heavy-handed, nanny-state approach. I think it is critical that America win the race to AI that we not lose to China.

Brooke Gladstone: Speaking of China, back in January, the Chinese tech startup DeepSeek stirred things up when it claimed to offer a cheaper, faster, more sophisticated AI model than its US competitors. OpenAI complained that DeepSeek violated its terms of service when it used OpenAI's intellectual property to make their model. Ironic, given that OpenAI trained its model on the intellectual property of the entire Internet. DeepSeek's launch six months ago roiled the market.

Reporter 1: Earth-shattering developments in the AI space.

Reporter 2: China's new AI chatbot is sending shockwaves through the tech industry, triggering a massive sell-off on Wall Street. The tech-heavy NASDAQ dropped more than 3% while the S&P 500 slid about 1 1/2%.

Reporter 3: Tonight, the biggest single-day loss for a stock ever.

Reporter 4: Shares of NVIDIA, which makes the chips used to power artificial intelligence, plunging nearly $600 billion.

Brooke Gladstone: Despite the hype, the Tow Center for Digital Journalism found that when given an excerpt from a news outlet, DeepSeek misattributed the source 115 out of 200 times. Ed Zitron, host of the Better Offline Podcast and writer of the newsletter Where's Your Ed At wasn't surprised.

Ed Zitron: In the case of a model based in China and run in China, yes, they have biases that they're going to feed into this model. That's not surprising. It's also important to remember that previous tests of ChatGPT and Google Gemini have found them to be disinformation and misinformation machines as well.

Brooke Gladstone: In January, I spoke to Ed about the AI arms race between the US and China and and so much more.

Ed Zitron: It's important to know how little the AI bubble has been built on. It's always been this idea, based on their narrative, that America has built these big, beautiful, large language models that require tons of the most expensive GPUs, the biggest data centers, because the only way to do this was to just cram more training data into them. DeepSeek found a way to build similar models to GPT-4o, the underpinning technology of ChatGPT and O1, the reasoning model that OpenAI had built, and make them much, much, much cheaper. Then they published all the research as to how they did and open-sourced the models, meaning that the source code of these models is there for anyone to basically adapt.

Brooke Gladstone: Now, GPUs, it's an electronic circuit that can perform mathematical calculations at high speed. Is this the chip that NVIDIA makes?

Ed Zitron: In a manner of speaking. GPUs, graphics processing units, they're used for everything from hardcore video editing to playing video games and such like that. Well, computer games, in the case of a computer. NVIDIA has a piece of software called CUDA. Now, all you need to know about that is GPUs traditionally were used for graphics rendering. What CUDA allowed, and it's taken them 15 years to do it or more, CUDA allowed you to build software that would run on the GPUs.

What NVIDIA started doing was making these very, very, very powerful GPUs. When I say a GPU in a computer, that's a card that goes inside. The GPUs that NVIDIA sells are these vast rack mounted like they're huge and they require incredible amounts of cooling. Generative AI, so the technology under ChatGPT, runs on these giant GPUs, and it runs them hot. They are extremely power-consuming, which is why you've heard so many bad stories about the environmental damage caused by these.

Brooke Gladstone: DeepSeek didn't have unlimited access to NVIDIA chips, partly because of the Biden sanctions and other things.

Ed Zitron: Now, before the sanctions came in, DeepSeek grew out of a hedge fund called High-Flyer. They had stockpiled these graphics processing units before the sanctions came in, but they'd also used ones that have less memory bandwidth. NVIDIA could only sell kind of handicapped ones to China. They had a combination of these chips. They found ways around both training the models, feeding them data to teach them how to do things, and also the inference, which is when you write a prompt into ChatGPT, it infers the meaning of that and then spits something out, a picture, a ream of text, and so on and so forth. These constraints meant that DeepSeek had to get creative, and they did. Just to be clear, all of this is in their papers. People are running the models themselves now.

Brooke Gladstone: They can also modify them.

Ed Zitron: Exactly, and they can build things on top of them. Now, an important detail here is that one of the big hubbubs is that DeepSeek trained their V3 model competitive with the underlying technology, ChatGPT, and they trained it for $5.5 million versus GPT4O, which is the latest model, cost 100 million or more, according to Sam Altman. They kind of prove you don't really need to buy the latest and greatest GPUs. In fact, you don't even need as many of them as you thought because they only use 2,048, as opposed to the hundreds of thousands that hyperscalers have. They're building all of these data centers because they need more GPUs and more space to put the GPUs and ways to cool them.

Brooke Gladstone: Didn't ChatGPT, or the company that made it, OpenAI, just create, as you say, this new reasoning program? They say it can answer some of the most challenging math questions. Are we saying that DeepSeek can do similar things?

Ed Zitron: Yes. Now, when you say it can do these very complex things, this is another bit of kayfabe from this industry.

Brooke Gladstone: What is kayfabe?

Ed Zitron: Kayfabe from wrestling, specifically, where you pretend something's real and serious when it isn't really. The benchmarking tests for large language models in general are extremely rigged. You can train the models to handle them, they're not solving actual use cases. On top of that, O1 is unfathomably expensive. When it came out, the large discussion was, "Wow, only OpenAI can do this." You had similar things, Anthropic, other companies, but OpenAI, top of the pile.

That was why they were able to charge such an incredible amount and why they were able to raise $6 billion, except now the sneaky Chinese. I mean that sarcastically. This is just good engineering. They managed to come along and say not only can we do this 30 times cheaper. To be clear, that number is based on DeepSeek hosting it, so we don't know who's subsidizing that, but nevertheless, not only can they do it cheaper, on top of that, they open source the whole damn thing.

Now, anyone can build their own reasoning model using this, or they can reverse engineer it and build their own reasoning model that will run cheaper and on their servers. Eventually, you're going to find cloud companies in America that will run these models. At that point, where's OpenAI's moat? The answer is they don't have one, just like the rest of them.

Brooke Gladstone: Now, China and America are in a hegemonic battle of generative AI. It seems that this DeepSeek tech has upended our assumptions of how all this was going to go. You say that there was never really any competition among American AI companies?

Ed Zitron: Yep, that is the long and short of it. This situation should cause unfathomable shame within Silicon Valley. What happened was Anthropic and OpenAI have been in friendly competition doing similar things in different ways, but they're all backed by the hyperscalers. OpenAI, principally funded by Microsoft, running on Microsoft servers almost entirely until very recently, but paying discounted rates and still losing money. They lost $5 billion last year. They're probably on course to lose more this year.

Brooke Gladstone: OpenAI?

Ed Zitron: Yes, and that's after $3.7 billion of revenue. Anthropic, I think they lost 2.7 billion last year. Google, Amazon, they back Anthropic. They just put more money in. You'd think with all of that loss they would be chomping at the bit to make their own efficient models. What if I told you they didn't have to? They all had these weird blank check policies and the venture capitalists backing the competition, whatever that was. There's nothing they could do because, based on the narrative that was built, these companies needed all this money and all of these GPUs, as that was the only way to build these large language models.

Why would any of them ever build something more efficient? The real damage that DeepSeek's done is they've proven that America doesn't really want to innovate. America doesn't compete. There is no AI arms race. There is no real killer app to any of this. ChatGPT has 200 million weekly users. People say that's a sign of something. Yes, that's what happens when literally every news outlet all the time, for two years, has been saying that ChatGPT is the biggest thing without sitting down and saying, "What does this bloody thing do?"

"Why does it matter? Oh, great, it helps me cheat at my college papers." Really, the biggest narrative shift here is that there was no cheaper way of doing this at the scale they needed to, other than give Sam Altman and Dario Amadei more money. That's the CEO of Anthropic. All they had to do was just continue making egregious promises because they didn't think anyone would dare bring the price down. I think it's good that this happened because the jig is up.

Brooke Gladstone: You've been saying since early 2024 that generative AI had already peaked. Why did you think that then, and why do you think so now?

Ed Zitron: The reasoning models, what they do, just to explain, is they break down a prompt. If you say, "Give me a list of all the state capitals that had the letter R in them," it says, "Okay, what are the states in America? What states have this?" Then it goes and checks its work. These models are probabilistic. They're remarkably accurate. They would guess that if you say, "I need a poem about Garfield with a gun," it would need to include Garfield and a gun, perhaps a kind of gun, and it would guess what the next word was. Now, to do this and train these models required a bunch of money and a bunch of GPUs, but also a bunch of data scraping the entire Internet, to be precise. To keep doing what they are doing, they would need four times the available information of the entire Internet, if not more.

Brooke Gladstone: Why?

Ed Zitron: Because it takes that much to train these models, requires you just shoving it in there, and then helping it understand things. There is no fixing the hallucination problem. Hallucinations are when these models present information that's false as authoritatively true.

Brooke Gladstone: I thought they'd been getting much better at catching that stuff.

Ed Zitron: No, they haven't. The whole thing with the reasoning models is that by checking their work, they got slightly better. These models were always going to peter out because they'd run out of training data. Also, there's only so much you can do with a probabilistic model. They don't have thoughts; they guess the next thing coming, and they're pretty good at it. Pretty good is actually nowhere near enough. When you think of what makes a software boom, a software boom is usually based on mass consumer adoption and mass enterprise-level adoption.

Big companies, financial services, healthcare, they have very low tolerance for mistakes. If you make a mistake with your AI, I'm not sure if you remember what happened with Knight Capital. That was with an algorithm. They lost hundreds of millions of dollars and destroyed themselves because of one little mistake. We don't even know how these things fully function, how they make their decisions, but we do know they make mistakes because they don't know anything. ChatGPT does not know. Even if you say, "Give me a list of every American state," and it gets it right every time.

Brooke Gladstone: It's just pattern recognition.

Ed Zitron: Yes. It doesn't know what a state is, it doesn't know what America is. It is effectively saying what is the most likely answer to this. Remarkably accurate probability, but remarkably accurate is nowhere near as accurate as we would need it to be. As a result, there's only so much they could have done with it.

Brooke Gladstone: You wrote, "What if artificial intelligence isn't actually capable of doing much more than what we're seeing today?" You said you believe that a large part of the AI boom is just hot air pumped through a combination of executive BSing and the media gladly imagining what AI could do rather than focus on what it's actually doing. You said that AI in this country was just throwing a lot of money at the wall, seeing if some new idea would actually emerge.

Ed Zitron: Silicon Valley has not really had a full depression. We may think the dot-com bubble was bad, but the dot-com bubble was they discovered e-commerce and then ran it in the worst way possible. This is so different because if this had not had the hype cycle it had, we probably would have ended up with an American DeepSeek in five years. The way Silicon Valley has classically worked is you give a bunch of money to some smart people, and then money comes out the end.

In the case of previous hype cycles that work, like cloud computing and smartphones, there were very obvious places to go. Jim Cavella over at Goldman Sachs famously said, "No one believed in the smartphone." Wrong. There were thousands of presentations that led fairly precisely to that. With the AI hype, it was a big media storm. Suddenly, Microsoft's entire decision-making behind this was they saw ChatGPT and went, "God damn, we need that in Bing. Buy every GPU you can find." It is insane that multitrillion-dollar market cap companies work like this.

They went, "Throw a bunch of money at it." Buying more, doing more, growing everything always works. I call it the rot economy, the growth at all cost mindset. Silicon Valley over the years has leaned towards just growth ideas, except they've chased out all the real innovators. To your original question, they didn't know what they were going to do. They thought that ChatGPT would magically become profitable.

When that didn't work, they went, "What if we made it more powerful and bigger? We can get more funding that way." Then they kept running up against the training data war and the diminishing returns. They went, "Okay, agents. Agents sound good." Now, agents is an amazing marketing term. What it's meant to sound like is a thing that goes and does a thing for you.

Brooke Gladstone: Just trying to get a plane reservation is a nightmare. If I could outsource that, I certainly would.

Ed Zitron: When you actually look at the products like OpenAI's operator, they suck. They're crap. They don't work. Even now, the media's still like, "Theoretically, this could work." They can't. Large language models are not built for distinct tasks. They don't do things. They are language models. If you are going to make an agent work, you have to find rules for effectively the real world, which AI has proven itself. I mean, real AI, not generative AI that isn't even autonomous, is quite difficult.

Brooke Gladstone: Coming up, more from Ed Zitron. If you thought the gloves were off in what you just heard, just wait for part 2.

Micah Loewinger: This is On the Media. This is On the Media. I'm Micah Loewinger.

Brooke Gladstone: I'm Brooke Gladstone. This is part 2 of a long conversation I had earlier this year with Ed Zitron, writer of the newsletter Where's Your Ed At. Before the break, he was eviscerating Silicon Valley's business model and said that its celebrated generative AI is good for practically nothing. I countered I mean, hey, it can write job application letters, it can write term papers, it can write poetry.

Ed Zitron: This is actually an important distinction. AI as a term has been around a while, decades, actually. AI that you see in a Waymo cab, like an autonomous car, it works pretty well. Nothing to do with ChatGPT. What Sam Altman and his ilk have done is attach a thing to the side of another thing and said it's the same. ChatGPT is not artificial intelligence. It is not intelligent. I guess it's artificial intelligence in that it pretends to be, but isn't. The umbrella term of AI is old and has done useful things, AlphaFold.

Brooke Gladstone: Yes, AI that goes through an enormous amount of data that shows us how proteins interact and maybe help us develop cures for diseases.

Ed Zitron: Right, and that is not generative AI, just to be clear. What Sam Altman does is he goes and talks about artificial intelligence and most people don't know a ton about tech, which is fine, but Altman has taken advantage of that, taken advantage of people in the media and the executives of public companies who do not know the difference and said, "Yes, ChatGPT, that's all AI."

Brooke Gladstone: Sam Altman went before Congress and said, "We need you to help us help you so that AI doesn't take over the world."

Ed Zitron: They love talking about AI safety. You want to know what the actual AI safety story is? Boiling lakes, stealing from millions of people, the massive energy requirements that are damaging our power grid. That is a safety issue. The safety issue Sam Altman's talking about is what if ChatGPT wakes up? It's marketing. Cynical, despicable marketing from a carnival barker.

Brooke Gladstone: They always say it never is going to generate a profit at the beginning. Even Amazon was not making money when it started.

Ed Zitron: They love that example. They say, "Oh, Amazon wasn't profitable at first." They were building the infrastructure, and they had an actual plan. They didn't just sit around being like, "At some point, something's going to happen, so we need to shove money into it." They love to bring up Uber as well. Now, Uber runs on labor abuse. Kind of a dog of a company, but you can at least explain to a person why they might use it. I need to go somewhere or I need something brought to me using an app; that's a business.

OpenAI by comparison, what is the killer app exactly? What is the thing that you need? What is the iPhone moment? Back then, to get your voicemail, you actually had to call a number and hit a button, but you could suddenly look at your voicemail and scroll through it. They were obvious things. What I'm describing here are real problems being immediately solved. You'll notice that people don't really have immediate problems that they're solving with ChatGPT.

Brooke Gladstone: Generative AI is incredibly unprofitable. $1 earned for every two and a quarter spent. Something like that.

Ed Zitron: $2.35 from my last estimates.

Brooke Gladstone: OpenAI's board last year said they needed even more capital than they had imagined. The CEO, Sam Altman, recently said that they're losing money on their plan to make money, which is the ChatGPT Pro plan. What is that?

Ed Zitron: Their premium subscriptions have limits to the amount that you can use them. Their $200 a month ChatGPT Plus subscription allows you to use their models as much as you'd like, and they're still losing money on them. The funny thing is the biggest selling point is their reasoning. Models O1 and O3. O3, by the way, it's yet to prove itself to actually be any different other than just slightly better at the benchmarks and also costing $1,000 for 10 minutes. It's insane. The reason they're losing that is because the way they've built their models is incredibly inefficient. Now that DeepSeek's come along, it's not really obvious why anyone would pay for ChatGPT Plus at all. The fact that they can't make money on a $200 a month subscription that's the kind of thing that you should get fired from a company for. They should boot you out the door.

Brooke Gladstone: How does DeepSeek make money?

Ed Zitron: That's the thing we don't know. It's important to separate DeepSeek, the company, which is now growth of a Chinese hedge fund. We don't know who subsidizes them, but their models, which are open source, can be installed by anyone, and you can build models like them. At this point, one has to wonder how long it takes for someone to just release a cheaper ChatGPT Plus that does most of the same things.

Brooke Gladstone: What's the end game here?

Ed Zitron: The end game was Microsoft saw ChatGPT was big and went, "Damn, we got to make sure that's in Bing because it seems really smart." Google released Gemini because Microsoft invested in OpenAI and they wanted to do the same thing. Meta added a large language model. They created an open-source one themselves, Llama. Everyone just does the same thing in this industry. Sundar Pichai went up at Google I/O, the developer conference and told this story about how an agent would help you return your shoes, how it would autonomously go into your email, get you the return thing, just hand it to you. It'd be amazing. Then ended it by saying, "This is totally theoretical."

All the king's horses and all the king's men don't seem to be able to get the Valley to spit out one meaningful mass market useful product that actually changes the world, other than damaging our power grid, stealing from millions of people, and boiling lakes. Everyone is directionlessly following this. They're like, "We're all doing large language models." Just like they did with the Metaverse. Now, Google did stay out of the Metaverse, by the way. Microsoft bought bloody Activision and wrote Metaverse on the side. Mark Zuckerberg lost like $45 billion on the metaverse. Putting aside the hyperscalers, there were hundreds of startups that raised billions of dollars for the Metaverse because everyone's just following each other. The Valley's despicable right now.

Brooke Gladstone: The Metaverse was Zuckerberg's effort to create some sort of multimedia space where people could live or something, right?

Ed Zitron: He was claiming it would be the next Internet, but really he needed a new growth market. The Metaverse was actually a symptom of the larger problem, the rot economy I talk about, which is everything must grow forever. Tech companies are usually good at it, except they've run out of growth markets. They've run out of big ideas. The reason you keep seeing the tech industry jumping from thing to thing, that when, as a regular person, you look at them and go, "This doesn't seem very useful."

What's happening is that they don't have any idea what they're doing, and they need something new to grow. Because if at any point the market says, "Wait, you're not going to grow forever," what happened to NVIDIA happens. NVIDIA has become one of the biggest stocks. It has some ridiculous multi 100% growth in the last year. It's crazy. The market is throwing a hissy fit because guess what? The only thing that grows forever is cancer.

Brooke Gladstone: You make a great case that if this is not a competitive atmosphere here in the US, if something does happen, it'll probably be somewhere else.

Ed Zitron: Yes. DeepSeek has proven that this can be done cheaper and more efficiently. They've not proven there's a new business model, they've not made any new features. There's an argument right now, a very annoying argument, where it says, as the price of a resource comes down, so the use of it increases. That's not what's going to happen here. No one has been not using generative AI because it was too expensive. In fact, these companies have burned billions of dollars doing so.

A third of venture funding in 2024 went into AI. These companies have not been poor. Now we're in this weird situation where we might have to accept, oh, I don't know, maybe this isn't a multi-trillion-dollar business. Had they treated it as a smaller one, they would have gone about it a completely different way. They never would have put billions of dollars into GPUs. They might have put a few billion and then focused up, kind of like how DeepSeek went, we only have so many and they only do so much, so we will do more with them. No, American startups became fat and happy. Even when you put that aside, there was never a business model with this. Come on, give me real innovation.

Brooke Gladstone: If this really is all early signs of the AI bubble bursting, what are the human ramifications of this industry collapsing?

Ed Zitron: I'm not scared of the bubble bursting. I'm scared of what happens afterwards. Once it becomes obvious that they can't come up with the next thing that will give them 10% to 15% to 20% revenue growth year over year over year, the markets are going to rethink how you value tech stocks. When that happens, you're going to see them get pulverized. I don't think any company shutting down, I think Meta is dead in less than 10 years just because they're a bad company. There are multiple tech companies that just lose money, but because they grow revenue, that's fine. What happens if the growth story dies? I hope I'm wrong on that one. The human cost will be horrible.

Brooke Gladstone: Even outside of the depression that might be experienced in the tech world, there are so many pension funds out there that may have investments. You're painting a picture of the housing crash of 2008.

Ed Zitron: I actually am. I wrote a piece about that fairly recently. It was about how OpenAI is the next Bear Stearns. If you look back at that time, there were people at that time talking about how there was nothing to worry about and using similar rhetoric. Just wait and see. Things are always good, right? It's horrifying because it doesn't have to be like this. These companies could be sustainable. They could have modest 2% to 3% growth.

Google as a company basically prints money. They could just run a stable, good Google that people like and make billions and billions and billions of dollars. They would be fine, but no, they must grow. We are the users our planet and our economy too. We are the victims of the rot economy, and they are the victors. These men are so rich. Sam Altman, if OpenAI collapses, he'll be fine. He's a multi-billionaire with a $5 million car.

Brooke Gladstone: There's a $5 million car?

Ed Zitron: Yes, he has a Koenigsegg Regera. Nasty little man.

Brooke Gladstone: Ed, thank you very much.

Ed Zitron: It's been such a pleasure. I loved it.

Brooke Gladstone: Ed Zitron is host of the Better Offline podcast and writer of the Newsletter Where's Your Ed At. We reached out to OpenAI for comment, but we didn't hear back.

Micah Loewinger: Coming up, a music composer grapples with whether or not AI will take his job.

Brooke Gladstone: This is On the Media. This is On the Media. I'm Brooke Gladstone.

Micah Loewinger: I'm Micah Loewinger. For the rest of the show, we're sticking with the theme of how AI is changing our lives, for better or ill. This final piece comes to us from a former OTM producer, Mark Henry Phillips. You may not remember his name, but you're probably familiar with his work.

[MUSIC]

Micah Loewinger: Yep, that infamous jingle was a Mark creation. When he left the show 14 years ago, it was to pursue a career in music, and it kind of worked out. He started scoring films and making commercials for clients like Google, PayPal, and Ford. He's also made music for podcasts like Serial Startup, This American Life, and Homecoming. As he says, he's not a rock star, but he's made a decent living making music until now. Mark has become terrified of and fascinated by AI music generators, particularly one that launched in April of this year. When he first encountered it, Mark realized his work would never be the same.

Mark Henry Phillips: I know this sounds dramatic, but the first time I played around with this AI music generator, it caused an existential crisis.

I literally lost sleep staring at my ceiling in the middle of the night, wondering, "Will this be my last year making money as a musician?" Okay, what happened? I stumbled across a track by a user named Man or Monster. He was trying to recreate a Toots and the Maytals track and had typed in some lyrics along with Soulful Reggae, Ska, 1969 Hammond organ male vocalist and out popped this.

[MUSIC]

Mark Henry Phillips: I don't want to admit it, but it's good. Like, really good. It doesn't feel like it was created by AI in a data center somewhere. It feels like it was made in a makeshift studio in Jamaica on a hot summer night in 1969. It feels real. Of course, this brings up all sorts of copyright issues because this sounds a lot like Toots, also known as Frederick Hibbert, also known as a real live human being. I would wager to bet that sounds that good because Toot's entire catalog was used in this AI music generator's training data.

Let's ignore that for now. Instead, let's just focus on how good this sounds.

[MUSIC]

AI-generated images often have glitches, like pictures of people often have extra fingers. In AI text, it also has its subtle tells that make it feel not quite human. This track, it might not be perfect, but it feels like real music. If it feels real, that has big implications. "Like what," you ask? Let me give you a hypothetical. You're an ad exec making a beer commercial. You want a track that sounds like it was recorded in Jamaica in 1969, but you don't want to deal with all the money and legal back and forth that comes with license, a vintage track.

You turn to a composer like me, you pay me, not as much as you'd pay Toots or Desmond Decker, but still enough that it'd be worth my while. Boom. You get a commercial, and I have a job. Years ago, that's exactly what happened to me. I was hired to make a song that sounds like an early Rocksteady tune from 1969, Jamaica. Essentially, I was given the same prompt as that AI track. Here's what I came up with.

[MUSIC]

It took me a couple days, and it's pretty good, fine, mediocre. It's a little embarrassing to play it now, and I was actually going to take the vocals out because that part is super embarrassing, but if we're going to do a comparison, it's kind of relevant. So here they are.

[MUSIC]

Whatever you think of my track, the vocals on the AI version just sound so much better, in that legally and morally dubious type of way. Don't get me wrong, there are many musicians out there who could have produced something way better than me or the AI, but the vocals, here's that AI-generated track again.

[MUSIC]

To get something like this, you'd have to hire or be an amazing Jamaican legend with a group of amazing backup singers. Even if you were Toots himself, you couldn't write, produce, record, and mix a track in under a minute. That's the crazy thing. Our hypothetical ad exec can now make this track. Hell, 10 tracks like this in under five minutes, and it's free. See why I was having that existential crisis? A lot has changed over the past decade, and the money has been on a steady trajectory downward, but now, post AI, I think a lot of composers are going to be without a job, including me.

If you've messed around with these services, you might be thinking, "They're interesting, but they're not that good without, what's this dude talking about?" But you're probably not using it to make the same stuff I make. AI isn't going to replace Billie Eilish or Radiohead, at least not for a while, but my bread and butter, theme songs, commercial music score, yes, it could replace me, like today. I don't think that's necessarily true yet with other AI products like ChatGPT or Dall-E. They're definitely cool or scary, depending on your orientation, but they're not really replacing professional humans quite yet.

ChatGPT isn't writing big ad campaigns, Midjourney isn't replacing photo spreads in magazines, but music, it's different from those other mediums. I think it's interesting to think about why. I was always taught that the golden rule of music is if it sounds good, it's good. Sounds pretty basic, but it has big implications. For the listener, that means you don't have to understand how a song was made to enjoy it. It's a black box process. Most people have absolutely no idea how their favorite songs were made, and it doesn't matter.

If it turns out good and it's good, but here's the weirdest part. For the musician, the process of writing music is also a little bit of a black box. A musician is never that conscious of what they're doing. You have a vague idea or a goal or a feeling, but you just mess around and discover the song. At a certain point, even if you're super cognizant of what scale and mode you're in and what the chords are doing and what is common in this style of music, at a certain point, you really just have to take a guess at what the next note is.

When I have a job, sometimes I'm really explicit with what I need to do. I'll say to myself, "Okay, this is a scene where a couple is getting to know each other. This needs to be something like a Jon Brion score from 1997. Something a little goofy and a little romantic." Jon Brion, if you don't know, is an amazing composer and producer who scored films like Magnolia, Eternal Sunshine, and I Heart Huckabees. Anyway, I don't actually go back and listen to one of Jon Brion's pieces and dissect the chord progression or the instrumentation, and I don't find just one piece to rip off. I just get a vague notion in my head, and I start playing around, and then something like this comes out.

[MUSIC]

It might not be amazing, but it hit the mark for the show I was working on, and importantly, I don't think it's a ripoff of Jon Brion. I had a good target in my head and I didn't get a perfect bullseye, but that's a good thing because it means I came up with something new. That's really how music works. I think this is a really key point. Take the Beatles, they were trying to sound just like Buddy Holly and Elvis, and their failed attempts became the Beatles. That's just how music evolves. The point is, I can't tell you exactly how I wrote this song.

Even though this was a pretty conscious for hire process. Even this writing and recording a song for a show, it's a black box process for me. That's why these AI music services are so good. It approaches music creation in the exact same way. It doesn't consciously say, "Write a four-chord progression in a melody in a mixolydian mode." No. Instead, it creates a Jon Brion-esque track just how I would. It's listened to all of his music and then using neural networks, whatever those are, it has a fuzzy image of a Jon Brion vibe. I don't know exactly how it works, but I think it's just always guessing what the next sound's going to be.

On one level, that means guessing the next note in the melody, but on another level, that means guessing what instrument should come in next, or guessing what the next cool production trick should be to keep the track interesting. In other words, it's approaching it just how I would. Here's the AI music service doing it for a Jon Brion track.

[MUSIC]

Its process is remarkably similar to a human musician. And that's why the results are too. When ChatGPT is "writing", it's also just guessing what should come next, but that's very different from actual writing. A writer knows what they're writing, but musicians, we don't fully know what we're creating, so the process is much easier for AI to replicate. This is why I think the next version of one of these AI music generators will replace me very soon. Let's take a project I'm working on right now. I'm scoring a show that's Hitchcock ass. Hitchcockian.

The show involves trains. Even though Strangers on a Train was scored by Dimitri Tiomkin, I obviously thought of Bernard Herrmann, who scored a bunch of other Hitchcock movies like Psycho and Vertigo, and who was friggin awesome. [00:41:32] Before I sat down at the piano and tried to figure something out, I thought I'd do a little experiment. I logged on to the AI music generator, typed in Bernard Herman theme, Alfred Hitchcock film Train Mysterious, and this is what I got.

[MUSIC]

To my ear, that's really good. Scary good. I keep waiting for it to break down and do something uncanny and weird, but it never does. Of course, I guided it as it made the track, but I really didn't have to do much. It was kind of a choose-your-own-adventure between really good options. This is the track that really scared the crap out of me. I could never in a million years write and record something this good. Even if I could, I would need to spend three times the entire budget of the project to hire a musical arranger, a real orchestra, and a recording studio.

All which would be necessary to create something that sounds this close to a Bernard Herrmann score, but AI, it did it in five minutes for free. This is where we flesh puppet musicians just can't compete. None of us are experts in every style. I might be able to beat it at an indie pop tune, but I can't do a better debut piece and a better medieval choral hymn, and certainly can't do a better Bernard Herrmann track. Plus, it's a virtuoso at every instrument. Violin, piano, vocals, guitar. Yes, in the vibes department, I might have an edge, but also maybe not.

Those examples were really good. That's what's so unmooring about this AI thing for me. It's not just the loss of work, it's part of my identity. It was my thing. You could give me a commercial or a film or a podcast, and I could make a song for it. It's the thing that made me a little bit special. When I tell someone I'm a musician, a lot of times their eyes light up. They think it's cool. Some even ask me, "How do you just write a song?" Now with AI, anyone can write a song. My special skill just isn't that special anymore. From a musical and economic point of view, AI just has me beat.

This is where I was this summer, freaking out. I picked up mountain biking to distract myself. I thought about starting a granola business, but then something happened. For me, it changed everything. The AI music generator I'd been using added a new feature allowing you to upload your own music. The idea was that it could listen to what you were doing and just extend it. To test it out, I uploaded a 12-second jingle I made a decade ago for a commercial I pitched and didn't land. This is what I made back then.

[MUSIC]

Nothing special, but I always thought that could have been the start of something. I uploaded it, and with a little prompting, it turned into this. It starts exactly the same, but then completely, seamlessly, it keeps going.

[MUSIC]

This is way more mind-blowing than the tracks I heard it make from scratch, because it doesn't feel like some other musician, it feels like me. This track isn't mine in the traditional sense, but because it grew out of a seed I planted, it really does feel like something I would have done if I wanted to turn this track into a full-length song. Like the percussion that just came in, I totally was imagining a vintage CR78 drum machine loop.

That's exactly what the AI put in. I know, I know. This is like someone saying they could have invented Instagram or could have invested in NVIDIA before the stock skyrocketed. Yes, dude, you could have, but you didn't. Maybe this is just me tricking myself into thinking this is my music, even though it isn't. The more I played with it, the more exciting it was. Take this track.

[MUSIC]

This was a little demo I made to test out a new guitar amp. I don't even remember making it, but I stumbled across it and thought, "It sounds pretty cool." What you're hearing right now is still just me, what I originally recorded, but I uploaded it to the AI music generator and then prompted it to do a horn arrangement. Of course, it obliged and it's switching to the AI version right now.

[MUSIC]

Why stop there? I messed around with the prompting and had it come up with another version, and it popped that out in just a few seconds. That version starts now.

[MUSIC]

Also amazing. Again, it really sounds like where I was imagining taking the song next, and yet, I would never feel comfortable using the audio it's generating as my own work. That feels like a bridge too far, but I could take bits and pieces of what it's done as jumping off points. This is where it gets so exciting. If I'm stuck on a song, let's say I can't figure out a new section. I could have the AI music generator come up with 10 new choruses for me.

I could take elements from different versions. A tempo change from one, a chord change from another, a bass line part from yet another version. Of course, I would play it all my myself. Since I'm just taking bits and pieces from different versions and putting my own spin on it while I play it, it wouldn't sound like any one version that the AI created. It'd be like having the best music writing partner ever. They're always awake. They're fast, enthusiastic, and good.

As weird as it feels to me using AI as a co-writer, I think young musicians coming up now could lean on these tools, just like I grew up with spell checks baked into Microsoft Word. Yes, there will definitely be holdouts, but as a society, it'll become the norm, both for musicians and the listeners. I can't help but see this as a weird fork in the evolution of music. Like music was pure up until 2023, but from here on out, it won't be, because AI music is already in the water. All future models will be trained on music that was made with AI. This could cause feedback loops in the algorithms, a proverbial AI snake eating its own tail. It could get really weird.

This might be the last year I make money as a musician. If that's the case, it'll free up time for me to finish the half dozen unfinished albums I have. I'm excited to actually do them, using AI as a writing partner. Yes, that brings up all sorts of moral and legal issues, so I probably won't release them. I'll just make them because it's fun. Isn't that what making music is all about? For On the Media, I'm Mark Henry Phillips.

Micah Loewinger: That's it for this week's show. On the Media is produced by Molly Rosen, Rebecca Clark-Callender, Candice Wang.

Brooke Gladstone: On the Media is a production of WNYC Studios. I'm Brooke Gladstone.

Micah Loewinger: I'm Micah Loewinger.

[00:50:26] [END OF AUDIO]

 

Copyright © 2025 New York Public Radio. All rights reserved. Visit our website terms of use at www.wnyc.org for further information.

New York Public Radio transcripts are created on a rush deadline, often by contractors. This text may not be in its final form and may be updated or revised in the future. Accuracy and availability may vary. The authoritative record of New York Public Radio’s programming is the audio record.