Musk 三剑客攻入华盛顿 + Spotify 的幽灵音乐 + 工具时间

Match Group has a new CEO, Kevin.

Yeah?

And he is the former CEO of Zillow, the real estate company.

Yes.

Which I imagine you probably spent some time on Zillow recently, as you've been looking for houses.

Yes.

Well, this raises the question, Kevin. As you're browsing through your Tinder matches—something I know you do a lot—do you think we're going to soon see some sort of Zestimate of that person's worth?

I think that's a great idea.

You know?

They should analyze market conditions and say, “The market price for a tall gay man in San Francisco, down 30% from last year.”

That's right. Short kings are having a huge moment. I think the Zestimate should say something like, “This person probably still has roommates,” you know? Not that it's bad to have roommates, but it can introduce complexity, and maybe you want to know that before you swipe.

You know how on Zillow you can see the history of every house?

Uh-huh.

Or of every property? I think you should be able to see the entire relationship history.

The entire romantic history.

Yes.

The last three romantic partners.

Yes. Two relationships ago, this person got dumped for not being a good communicator.

I have to say, we've come up with more good product ideas for Tinder in these past 5 minutes than Tinder has in the past year.

Call us.

Call us.

I'm Kevin Roose, a tech columnist at The New York Times.

I'm Casey Newton from Platformer.

And this is Hard Fork.

This week, The Times's Jonathan Swan joins us to discuss Elon Musk's tech takeover of Washington, D.C. Then, author Liz Pelly stops by to discuss her new book on Spotify and how its algorithms are reshaping music culture. And finally, it's tool time. We'll tell you about the new AI tools we're using and the one that we wish existed.

1. Musk Takes Washington

Well, Casey, the biggest story in tech this week is actually not happening in the Bay Area, where we live. It is happening across this great country, in Washington, D.C.

It sure is, Kevin.

Elon Musk and his team at DOGE, the Department of Government Efficiency, have been hacking away at the federal government, barging into agencies, demanding data and access to computer systems, basically staging what some people are calling a tech takeover of the federal government.

Yeah.

Musk brought with him to Washington a bunch of people to help him at DOGE with this effort, including a number of young men, some of them in their 20s and even, reportedly, a teenager or two, who are helping him with this effort.

Yeah, including Luke Varator, who we mentioned on the show in a previous episode, Kevin, because he was part of an effort to decode ancient scrolls using AI.

Yeah. Together, they've been pulling late nights, some of them reportedly sleeping literally in their offices, so that they can work basically around the clock to shrink the federal government.

Yeah. And they're doing it in some really aggressive, and some would say scary, ways. They have already gained access to the Treasury's payment system. They have put on leave nearly the entire workforce of USAID. And they have emailed roughly 2 million federal workers, Kevin, offering them the option to resign and allegedly be paid through the end of September.

Yes. The subject line of that email was “Fork in the Road,” which is not a Hard Fork reference—

That we know of.

—that we know of. But it was the same subject line that was sent to employees at Twitter after Elon Musk took over Twitter, giving them the chance to resign or take severance packages if they didn't want to work there anymore.

Yeah.

So, Casey, why are we talking about this on Hard Fork this week? We are not a politics show.

We are not, but Kevin, several listeners emailed us saying, “We want to know more about what is happening.” What we're seeing unfold in Washington is certainly unprecedented in the modern history of the United States, and it involves somebody who has been a main character of this podcast from the beginning: Elon Musk. In 2022, Elon Musk bought and took over Twitter, and what is happening in the federal government, while it is infinitely more important than Twitter, is unfolding in a very similar fashion.

Yeah. That, to me, is what brings this into our lane. I feel like the Twitter takeover was sort of the warm-up act for what is happening with DOGE and the federal government. Many of the same tactics and playbooks that were used to take over Twitter, to purge the disloyal, woke employees of that company, as Elon Musk saw them, are now being used on a much bigger scale on the federal workforce. So we brought in someone who is an expert on politics, Trump, and all things Washington: Jonathan Swan, my colleague at The New York Times. He was one of the authors of a piece that came out in The Times earlier this week that was basically a broad and sweeping look at all of the ways in which Elon Musk and his allies have been making what they called an aggressive incursion into the federal government. Really great story. Jonathan has been covering Trump for years, and he's got a feel for the pulse of Washington and how people are reacting to Elon Musk's big invasion.

Let's bring him in.

Jonathan Swan, welcome to Hard Fork.

Thanks for having me.

So, Jonathan, give us the view from Washington. What is the vibe on the ground as Elon Musk and his band of Silicon Valley programmers move around trying to overhaul the federal government?

Well, it really depends who you talk to. For the career civil servants, it's terror. These people don't know if they're going to have jobs. In some cases, they don't know if their agency is going to exist in the morning. The website goes dark. At USAID, they get an email after midnight: “Don't come into work.” He's calling their agency evil. They're following him on X.

A lot of it's very opaque. He's got these young guys who work for him at Tesla and some of the other companies. Some of them are in their 20s. One of them was 19. They're roaming around the agencies. They'll do these interviews with folks, but they sometimes won't tell them their names because they're worried about being doxxed.

Imagine you're a civil servant. This guy shows up from DOGE, wearing a T-shirt and a blazer, and starts basically interrogating you, with the questions all based on the assumption: “You are a lazy, worthless, idiotic federal worker. Justify your existence to me, please.” And they say, “Well, who are you?” “I'm not telling you my last name,” or whatever.

So this operation is now unfolding. Do we know where it's going? Is there a roadmap that anyone can see, or do we just have to rely on reporting that folks like yourself are doing to even understand what is happening and what the plan is?

2. DOGE Plans Massive Cuts

Broadly speaking, we know what he wants to do, right? He has said he wants to cut $2 trillion out of the federal budget. The federal budget is around $7 trillion. It's almost impossible to imagine how he would actually do that.

He's since revised that down. Even Elon Musk, who's famous for setting unrealistic expectations and deadlines at Tesla and SpaceX, has said, “Well, maybe we'll get to $1 trillion.” Even that would be astonishing, and people don't really think that's plausible.

But we know he wants to cut. We know there's an ideological agenda. Agencies that are doing things that are seen as not aligned with the Trump movement are going to face more hostility. USAID is the platonic ideal, in their minds, of the evil leftist deep state, because what is the Trump movement? It's, quote-unquote, “America first.” Well, what is USAID? It's an agency that spends money overseas in foreign aid and humanitarian assistance. They have found themselves in the crosshairs, but so have a bunch of other agencies. Their general contempt for the federal workforce was really evident if you read that email.

Right, the “Fork in the Road” email.

Yeah. You guys know this because this is your field, the tech, but basically, this email went to almost all federal workers—around 2 million federal workers. I thought the email was really revealing.

Mm-hmm.

When you talk about his plan—what he's thinking—I thought it was such a revealing document because the text of the email basically said, “We'd love you to resign.”

Mm-hmm.

“Whoever you are—you are in a lower—”

Mm-hmm.

“Productivity job, and you should resign and take a higher-productivity private-sector job. I don't distinguish between your expertise. I don't distinguish between your experience. You're all basically worth nothing.”

Yeah.

Wow.

Jonathan, let's talk a little bit about the cast and characters here. Obviously, our audience is very familiar with Elon Musk, but tell us about the people around him, these young men from Silicon Valley that he's brought in with DOGE, who everyone's been talking about this week. You mentioned some of them are in their early 20s, maybe even one who's still a teenager. Who are they? How many of them are there? Do they have any private- or public-sector experience, or are they just interns from his companies who he thinks would do a good job helping with this?

Well, this is a little opaque to me, and I have to give credit to my wonderful colleagues Ryan Mac, Kate Conger, and Teddy Schleifer. From what I can understand, some of these are very bright and experienced allies of Elon Musk that he's worked with for a long time. Tom Kraus, who is, I think, the one that was given access to the Treasury payment system and I think he's the CEO of, like, a software, cloud software or something.

One of them deciphered some ancient scrolls. That was his claim to fame. Luke Ferreter.

Yeah. You run the gamut from that guy, the scroll decipherer, to more seasoned people who've worked with Elon Musk for a long time. But I will say it's not actually that clear to me how many of them there are. I've heard that there were around 40 at the inauguration. They show up, and they're very confident. They ask a lot of questions and want access to the systems. They want to get their hands on the pipes of government and not take the word of career officials telling them what they're doing.

Man, it just makes me think of every time over the last decade that I've heard some person in Washington saying, “We've got to get more young people interested in government.” I just pictured the monkey's paw curling.

Yes.

You might not have wanted it to go down this way.

Careful what you wish for. So, Jonathan, one of Musk's first moves was, as you mentioned, to seize the federal government's payment systems. Why did he start there?

3. Musk Targets Payment Systems

Musk has told people in the administration that, in his view, the way to control government is to control the computers. He's got a real history of taking an interest in the detail—really getting down to the nitty-gritty, roaming the floors of the factory, and asking, “Why do we have this part in this machine? Why can't we get rid of it and make it more efficient and cheaper and whatever?” That's the mindset, as far as I can tell from talking to folks in the White House and the government.

His view is, “I don't want my guys at DOGE to sit down with deep-state, quote-unquote, official X, who's going to tell them everything's great and blah, blah, blah. No, I want my guys to have the source code to go in and see for themselves all this nasty fraud.” He's sort of doing a version of what he did with Twitter. He's trying to publicize things that he thinks are ridiculous that money's being spent on. It's the same playbook we're seeing publicly.

But with the Treasury, this is really important. This is a really important part of America's critical infrastructure. I've talked to—

Mm-hmm.

Former senior Treasury officials were really alarmed. These were the outgoing Biden people in December, when they got these requests from the DOGE people saying, “We need the source code.” This payment system is literally the payment system that distributes more than $5 trillion a year. About 88% of federal spending goes through this system. People's Social Security payments—people depend on this system.

It's historically been managed by a small group of really trusted, really experienced career civil servants. As far as I know from talking to former officials, they've never heard of a situation where a political appointee has requested access to this, let alone been granted access to it. Now, they are—

Right.

Insisting to us that it is, quote-unquote, “read-only access,” meaning they can't alter payments. But even that is considered extraordinary.

But here's why I think so many people are concerned about this: The Constitution gives the power to spend money to Congress, right?

That's right.

Not to the president. The president cannot unilaterally decide what to spend money on or not spend money on. Give us a sense of the conversation around the law here. Is anyone even trying to make the case that what Musk and his colleagues are doing is legal?

Firstly, you're absolutely right. Congress has the power of the purse. Just no question about that. The White House has cast this as a temporary freeze to examine the spending and make sure it doesn't conflict with Trump's policy priorities. But as Charlie Savage, my colleague, has written, it also appears to plant the seeds of a potential Supreme Court fight over how much power a president has to refuse to spend money that Congress has appropriated.

So there actually is a legal question here that could be litigated all the way up to the Supreme Court. Trump's aides have long wanted to seize back some of this power to withhold spending.

There's this theory floating around that I want to get your take on: Some of this is just a diversion or a tactic. Basically, the DOGE folks know that not all the things they're doing are legal. They know that not all of it will end up passing muster with Trump or with Congress. But they're asking for a foot and expecting 6 inches—that they're essentially overreaching on purpose so that, even if half of what they're doing gets overturned or overruled or can't make it through all the checks and balances, they will still have gotten a fair bit of what they wanted. Do you buy that, or do you think they legitimately expect all of the things they're doing to stand up?

4. DOGE Floods The Zone

Oh, no, no, no. We've written about this. This is actually a really important part of how they think and their strategy. They learned in their first term—it took them a while—but they learned that the most effective way to get really aggressive policies through is to flood the zone. It's to do so many things at once, and so aggressively, that your opposition—and when they think about their opposition, their mental map is the media, Democrats in Congress, and outside nonprofit groups like the ACLU who are likely to sue them—has a limited amount of resources.

There's only so much mental bandwidth to fight them, so people need to pick their targets. Meanwhile, you're shooting bullets one after another on all these different issues. That's absolutely been their approach in this onslaught of executive action that we've seen in the last 2 weeks.

Yeah, I'm sure everyone in Washington is very shocked and surprised. I imagine there's one group of people who are not all that shocked and surprised, which is Twitter employees or former Twitter employees. Casey, this is a question more for you, but you covered the Musk takeover of Twitter and everything that followed that, including layoffs and budget cuts and general madness. As you're watching what's unfolding in Washington, is anything surprising you, or do you feel like we're just seeing a story we've seen play out before, just on a much bigger stage?

I mean, the playbook is not surprising. We know that this is how Musk operates. He does have tremendous disdain for anybody whom he did not hire himself, right? We're seeing so much of the way he treated Twitter employees reflected in the way that he's now treating the workforce of the federal government.

What I think is so surprising, though, Kevin, is that Twitter was a company that he bought, right? He had the legal right to do most of what he did. There were some lawsuits related to some agreements that he maybe broke, but for the most part, he bought the company, and it was his right to decide whom he wanted to work there and what he wanted to do with them.

The shocking thing about the case of the federal government was that, as Jonathan just said, this is the richest man in the world. He was not elected. He has not presented Congress with a plan for what he wants to do with the government. He's not gotten consent from the legislative branch.

Yeah, I mean, to me, these don't feel like perfect comparisons, because as you said, one of these is a company and one of these is a government. But I am starting to see some parallels in some of the tactics that Elon Musk is using, one of them being this idea of zero-based budgeting. So, Jonathan, tell us about zero-based budgeting and how it's showing up in Washington.

Well, it's the idea that you bring a budget to zero and then justify every dollar that you add in spending. Instead of saying, “What should we cut?” it's, “No, let's start from zero and say, ‘What should we add?’” It's just a way of forcing people to justify every single dollar that they spend.

Elon Musk has told people that the success of this DOGE effort—his metric for it—will be how many dollars they save per day. And to do that, they’re looking at Treasury payments and USAID. He’s looking at the federal government’s real estate portfolio to see what they can offload. So the range is so wide.

Yeah. To me, what happened at Twitter, to the extent that that can be used as a preview of what might happen in Washington, is that there were 2 phases of that takeover of what he considered a hostile institution.

One of them was the operational phase, where you try to figure out who’s paying what to whom, what we’re spending money on that we don’t need to, and where the inefficiencies are. Then there’s the ideological purge, which happened when he would go around to Twitter employees and ask them to commit to being extremely hardcore, and try to figure out who was on his side and who wasn’t, and then purge the people who weren’t.

Do you see any signs, Jonathan, in Washington, that that kind of thing is happening? Are these DOGE people going around asking people to pledge their loyalty to the Trump administration, or is that still to come?

I’d have to go back and look at the text of that email that was sent out, but I think loyalty was one of the criteria in that email. Certainly, the Trump team has made loyalty absolutely central to the way that they hire people.

5. DOGE Brings AI To Government

There’s been some reporting in The Times, Jonathan, that Elon and his crew want to bring AI into government. Do we know anything about how or what they mean by that?

I credit my colleagues for this, Kate Conger and Ryan Mac, but this was in our big story on Musk. As we understand it, Musk’s allies aim to inject artificial intelligence tools into government systems, and the point, supposedly, is to use them to assess contracts and recommend cuts.

What Kate and Ryan were told was that on Monday, Thomas Shedd, who’s a former Tesla engineer and has been tapped to lead a technology team at the General Services Administration, told some staff members they hope to put all federal contracts into a centralized system so they could be analyzed by artificial intelligence.

I know from my own reporting that Elon Musk has privately, for months now, been talking about this idea of using artificial intelligence to identify waste within the federal government. It doesn’t seem like a crazy idea to me conceptually. Use whatever tools you can to figure out where the wasteful spending is. The problem is, I don’t have visibility into what these tools are. It’s all very, very opaque.

Right. And I think we should say this part does not feel unprecedented to me. The various administrations, Democratic and Republican, have tried to bring in the brightest minds in the tech sector to update and modernize some of the creaky, outdated systems that many government agencies use.

We had the U.S. Digital Service. There are groups like 18F, these groups of technologists who are brought in to try to bring things up to date. But that is a process that is established and requires doing things like going through a procurement process if you want to use some new AI tool, because maybe it’s not secure. Maybe there are privacy concerns. You want to make sure that that is fully vetted before you roll it out into these very important systems.

It seems very different to have a group of engineers, programmers, and product people coming in and just saying, “We’re going to use these tools, whether you like it or not.”

Yeah. One thing that a source mentioned to me the other day, who’s been a very senior person in the government, is the counterintelligence risks here. When you have a bunch of people who are young, who are from Silicon Valley or different private companies, moving very fast, very aggressively, and getting really sensitive access to the federal government, that opens up all sorts of espionage opportunities.

Foreign governments are constantly targeting the American government workforce, looking for vulnerabilities. There are all kinds of potential side effects of this that perhaps are not being considered as they move really quickly and aggressively.

Well, help us think through the next steps here. We know that there are already some lawsuits percolating, designed to maybe stop some of this. We’ve also seen Democrats wake up and start protesting. But can you give us a read, Jonathan? What do you think is likely to happen over the next week or so? Do you imagine that anything is going to put the brakes on DOGE, or are they just going to have their way with the federal government here?

Obviously, there are lawsuits. One of the challenges with lawsuits in general is that the speed at which Musk and Trump are moving far exceeds, I think, the capacity of the legal system to catch up. They’re doing so many things at once, so quickly, that the facts on the ground are changing.

In the meantime, a whole bunch of things are happening. Relief projects in Sudan and all around the world, where U.S. foreign aid is helping people, have stopped already. So yes, there are going to be legal challenges. Some of this won’t fly ultimately, but some of it will.

We’re not seeing much appetite from Congress to assert themselves and assert their authorities. Obviously, the House and Senate are in Republican hands. We’re not exactly seeing a very aggressive legislative branch.

In terms of Musk himself, the limit on him is the extent to which Trump tolerates him. That’s the only kind of limiting principle. There have been a lot of people predicting that this relationship would blow up. It’s kind of interesting. He’s willing to tolerate a lot more from Elon Musk, and it might just be as simple as: it’s pretty flattering having the richest guy in the world, and pretty convenient having a guy who spent $300 billion helping you working for you, as Trump sees it.

Trump’s the president. Elon Musk can never be president. He was born in South Africa. From Trump’s point of view, that’s great, and Elon is the one who’s been most aggressive at turning his platform into a vehicle to support Donald Trump.

Yeah. Jonathan, out here in Silicon Valley, we’ve spent a lot of time over the past year talking about various types of management changes. One of them is founder mode, which is this school of thought that a lot of tech companies have borrowed from Elon Musk, where basically you stop listening to your workers, take control, dictate more from the top, and try to make things as lean and efficient as possible.

I see what Elon Musk is doing in Washington as an extension of founder mode, which is a kind of corporate authoritarianism. But I’m wondering if you think there’s a parallel to be drawn here between the way that a company is managed in an industry like tech and the government, or do you think that those are just fundamentally 2 different things?

Yeah, of course. You’re talking about a federal bureaucracy at a really dangerous time in the world, a complex world, and a federal government that has to do so many things: make sure our water is clean, make sure our food is safe, take care of our critical infrastructure, manage national security, including cybersecurity, and manage air travel.

The “break it to fix it” mindset—the “break it” part of it is pretty important—because what gets broken, the stakes are just so much higher when you’re talking about the entire country and the federal government.

Well, we saw what happened at Twitter, right? Twitter doesn’t exist anymore. That was how the Elon Musk approach worked for Twitter. There’s something else now. It’s called X. It’s not as good, doesn’t make as much money, and doesn’t have as many people using it. He tries to sue people just to advertise on it to keep it running.

So that’s how that’s going. I have no confidence that what they’re doing is going to lead to some magically more efficient federal government, because nothing they have done so far suggests that they have a plan that is centered around taking care of people and making sure that people still get the services that they depend on, which is one of the key reasons the federal government exists.

Jonathan, quick last question, and then we know you have to go. So far, Elon Musk and his DOGE cadre have gone after Treasury. They have gone after USAID. They’re reportedly now setting their sights on the Department of Education. What are the 3 next agencies that you think are in their crosshairs?

Look, they’re going after every agency in the government. Although I will say, as far as I can tell, he hasn’t really been that involved at the Defense Department. But I do expect that that will come, because if you’re really thinking about how to cut government spending, you can’t ignore the Pentagon. It’s such a huge part of government spending.

And listen, Trump has really tied their hands to a large extent, because he said you can’t touch Social Security or Medicare—huge entitlement programs. He’s promised not to cut money from them.

So if you’re Elon Musk and you’re looking for savings, eventually he’s going to have to turn his eye properly to the Pentagon, and I’ll be very interested to see what they propose there. Again, huge conflict of interest: Elon Musk, SpaceX, huge federal contracts. But I’m going to be keeping a close eye on the DOD.

Jonathan Swan, thank you so much for joining us.

Thanks, Jonathan.

Thanks for having me.

When we come back, writer Liz Pelly tells us why Spotify is increasingly full of ghost musicians.

Spooky.

Kevin, if you were a streaming music service playlist, what would you be called?

Probably “Lo-fi Beats to Podcast To.”

Hmm. I think of you more as a 2000s hot girl, girly-pop Wednesday afternoon. But regardless, Kevin, next on our playlist today, we're going to talk about Spotify.

Yes. Spotify is a company that we really haven't spent much time talking about on the show, but I think they are very important within the world of tech companies.

In part because when we say “wherever you get your podcasts,” Spotify is a place where you can get your podcasts.

Right. Many of our listeners are probably using Spotify right now, and Spotify has had a big week. They just reported their first profitable year ever. Daniel Ek, the CEO, was quoted as saying, “It only took 18 years for us to get here, but we're here.” The company now has 675 million users and around 263 million premium-paying subscribers. Their ad-supported revenue is also up, but that's not what we're really here to talk about today.

No, Kevin, because as popular as Spotify is, a new book argues that the company's rise hasn't necessarily benefited artists or listeners. Liz Pelly is a writer based in New York who has a new book out called Mood Machine: The Rise of Spotify and the Cost of the Perfect Playlist. I have to tell you, I was captivated by an excerpt of this book that came out in Harper's Magazine recently, and the excerpt focused on what are sometimes called ghost artists.

These are musicians who Spotify uses as a way to fill out popular playlists with lower-cost music made exclusively for the company instead of songs from major record labels. And according to Liz, it's proliferating quite quickly.

Yeah, this is the kind of story that doesn't get told about Spotify that often, which is how it has essentially become an invisible force in the music world, shaping the tastes of its hundreds of millions of subscribers in ways that maybe some people, even hardcore Spotify users, don't fully appreciate.

Yeah, and we've talked about so many other invisible algorithms on this show that are reshaping culture in one way or another. This is our chance to learn how that is unfolding inside Spotify. So let's bring in Liz Pelly. Liz Pelly, welcome to Hard Fork.

Thanks for having me.

So let's talk about Spotify's evolution as a music service over the years. When I first started using it, I really felt like I was the person in charge of my music listening. I would search for the artist or album I wanted to listen to, play it, and then go look for more. Today, though, it feels like Spotify is more in charge, pushing algorithmic and paid recommendations at me every chance it gets. So how did that evolution start?

6. Spotify Takes Control

When Spotify launched, these things were more like search bars. You would have to know what you were looking for. You would have to know the artist or the album that you wanted to listen to because, in certain ways, when Spotify launched, it was really competing for the type of music listener who had become accustomed to the kind of digital library they had access to in the post-The Pirate Bay years, the post-Napster years.

The type of digital music listener who was used to opening their laptop, opening their music library, and being able to push play on whatever they wanted to hear at any moment.

Yeah. So at some point, Spotify begins pushing people away from this search-bar experience and more toward playlists. What's the origin of that?

Up until around 2012, when you looked at Spotify's branding—the way it characterized itself—on its own website, it would really focus on words like “instant,” “simple,” and “free,” and it would talk about giving you access to a world of music. It really wasn't until later in 2012, around a year after it had launched in the United States, that the way it positioned itself started to change.

Around this time, it had also commissioned a research agency to study its user base and try to give it information about what people were actually coming to the platform for. It had started to realize that its users weren't only looking for access to music, but were also looking for the ability to get a recommendation or hit play and get a feed of appropriate music.

At the end of 2012 and early 2013, you start hearing Daniel Ek in the press talking about how maybe he'd been too precious about this idea of a non-curated service. They started redesigning the homepage, and by 2013, they really started to lean into this idea of a more curated service.

And that's when I first started hearing about things like the RapCaviar playlist, which was a very popular playlist that a lot of people were using. Artists were angling to get into it, and labels were angling to get their artists into the Spotify playlist because Spotify's increasingly large user base was discovering new music through the playlists.

There was an element of that that felt familiar to me. Radio had the same thing, where artists and labels would fight to get their songs played. But this started to feel like Spotify was actually getting its own market power because it had all these subscribers, and it could start to direct them toward certain music and away from other music.

Definitely. As the years went on, these playlists became pretty influential in the music business. Like you said, they started to become a really integral part of how record labels thought about promoting music, and musicians, both major-label and independent musicians alike, started to be pitched on and sold on the promotional opportunities of this whole system.

When I started writing about Spotify, which was in the mid-2010s, one of the things that was really interesting to me at the time was the way in which independent musicians were being sold on this playlist system as a democratizing force.

Mm.

Spotify always said things like, “The playlist ecosystem is going to level the playing field,” and they talked a lot about this pyramid of playlist curation, where they would start artists on these low-tier feeder playlists, look at streaming data, and then move you up in the playlist system if the song reacted well or if there was a high completion rate.

Mm.

This was a myth that was sold to artists, but a lot of independent artists weren't necessarily feeling the magic of this data-driven, supposedly meritocratic system.

Right. Yeah, I want to ask about that, because I have to say, from my perspective, when it comes to the rise of playlists, that's basically okay with me. It sounds, from the way that you're talking about it, like a lot of the reason that playlists came to be on Spotify was just user demand. People wanted a kind of guide to their music. They didn't want to be responsible for thinking up every single thing that they wanted to listen to at any given moment.

But you write in your book that over time, Spotify became increasingly concerned with shaping user behavior on the platform. So aside from the playlists that you described, how does that manifest? How do they try to shape the way that we use the app?

I think this is similar across the platform economy. Platforms want to shape user behavior in order to boost engagement, hook people on their platforms, and extend the amount of minutes and hours that people are spending on their platforms so that people have tighter relationships with their products and see their products as more valuable.

In the case of a streaming service, a streaming service endeavors to keep people on the platform longer so that they view it as a useful part of their lives and retain their subscriptions, right?

But it's not just about boosting engagement, right? Because my understanding is that Spotify pays out a huge chunk of its revenue to record labels for their music.

I mean, billions of dollars are paid to artists and record labels. And so if you're Spotify and you're trying to grow your business, you could either grow your subscription base, or you could just pay out less money to artists and labels. One of the ways that you could do that is by steering people away from headline acts and artists with negotiating leverage and negotiating power, and the big labels, and toward maybe smaller or lesser-known musicians who maybe can't command the same types of market power. So is that something that Spotify was also doing?

7. Ghost Musicians Fill Playlists

Yeah, that's a great point, too. Part of the reason why a streaming platform wants to control more of the user experience is so that they can have more influence over the types of music that are being recommended to users. And in the case of Spotify, one of the things that I try to trace throughout my book is this series of cost-saving initiatives that the company developed in order to try to nudge users toward content—and I hate using the word “content” to describe music, but this is how they refer to it—in order to nudge users toward content that's cheaper for them to license.

There are 2 specific instances of this that I talk about in the book. One was around 2017, this phenomenon that people started to notice where their playlists for studying, chilling, sleeping and relaxing started to include tracks that didn't necessarily seem to be from real artists. People were noticing their playlists increasingly filled with what appeared to be royalty-free stock music. So one of the investigations in the book is into the rise of what internally at Spotify is called Perfect Fit Content, which is music commissioned for certain playlists and moods with improved margins.

What you're describing is music that is made for Spotify. It is not Spotify going out and curating music that exists in the world and putting it into playlists. This is like, “We want to make a new lo-fi study playlist, and so we are going to go out and have a bunch of studio musicians make this music, and then we're going to pay them much less than we would pay Taylor Swift.”

Right. So there's a handful of production companies that are part of this scheme, and those production companies will then go find producers and composers who can make the stuff. And one of the most interesting parts of reporting my book was talking to a handful of musicians who had made work for these production companies with these privileged deals. They'd talk about how sometimes they'd be cranking out 12 or 15 of these tracks in an hour, and the idea is to just create as much content as they can, to make it as simple as possible so it goes well in the background.

They'd be studying music on certain playlists provided to them by the production companies as examples, which would basically be songs that had done well in the lean-back environment on Spotify previously. Then they would try to replicate similar styles in order to hopefully make content that would stream really well.

And I should say, I understand the impulse to do this.

Personally, I am what you might call a lean-back listener. I listen to a tremendous amount of Spotify, many hours a day, mostly in the context of trying to go to sleep, and it plays while I'm sleeping, or trying to study. I'm a huge consumer of all of the lo-fi music to study to.

After reading your book, I assume that most of that is this Perfect Fit Content, and that artists are not being paid very much for that. But that is useful. I don't really care what the music is that puts me to sleep. I just want some music that's in the right genre with the right kind of sound.

So I should just say, I understand the market forces at work here, because I am part of the universe of Spotify subscribers who do use this more ambient kind of music. Yeah. Kevin is one of the reasons why the music industry is falling apart.

Well, it's interesting because, according to some of the interviews that I did, a justification that would be used is that some of the senior executives would say things like, “Well, most people don't know, and also they don't care.” I understand that there are certain types of users who won't care, but I think there are certain types of users who would care, and they can't decide whether or not they care if they don't know.

Right. Right.

So one of the things with these cost-saving initiatives that stands out to me as a glaring issue is the fact that none of this material is labeled as sponsored or labeled as, “This is being recommended due to a commercial deal.” I think from the beginning, these sorts of playlists have operated under the umbrella of editorial on streaming services.

And we're not just talking about Spotify. I think that there is reason to believe that other streaming services are likely engaging in similar practices as well. But if something is operating under the umbrella of editorial, I do think that there's some sort of expectation that if something's being recommended due to a commercial partnership, that should be labeled in some way.

I completely agree with you. And by the way, if they did have to label those things through regulation, you'd see a lot less of it, because they would be embarrassed.

I want to talk, in addition, about the effect on the culture, and maybe we should use lo-fi beats as a jumping-off point, since you brought it up, Kevin. There is a Chill Lofi Study Beats playlist on Spotify that is very popular. It's been saved about 2 million times.

And in your book, Liz, you write about how the rise of lo-fi beats really reflects this era. I'm curious if you can tell us what lo-fi beats were before it became a big Spotify and YouTube phenomenon. What was it as an actual culture before it became a low-cost alternative to paying record artists?

8. Spotify Flattens Music Culture

Yeah, I think it's interesting to know that the phenomenon that is now known as lo-fi hip-hop beats to study and relax to had a sort of prehistory online. In the early 2010s, the lo-fi hip-hop community online was based more in forums, with people making these J Dilla- and Madlib-inspired beats and sharing them with each other.

It was more based on SoundCloud. People were making music inspired by these producers that they really loved, and it involved more sample flipping, with people trying to outdo each other with impressive drums. It was less mellowed-out, not exclusively background music made for studying.

And according to some of the people I talked to, as this scene moved to YouTube and Spotify, as playlist curators got in the game, there was this sort of flattening effect that happened. Certain types of music from this subculture were being put onto playlists for studying, and then the types of music that did well on playlists for studying were financially incentivized, so more people started making that type of music.

I mean, that push and pull that you identify is so interesting. On one hand, this doesn't feel great, because now people aren't really hearing the authentic lo-fi beats. They're hearing the cheap version, and they don't even know that they're hearing the cheap version.

But on the other hand, the playlist became so popular that it incentivized the creation of a lot more of this music. People wound up hearing a lot more of this thing that they like. So how do you think about that kind of push and pull? Is the picture of what Spotify is doing to the culture more mixed than just, “Algorithms are flattening everything”?

I think for me it's an important distinction: I don't necessarily just think the issue is that people are listening to less authentic music or that people are listening to fake music. For me, my concerns have more to do with the reality that there are so many independent musicians today who are trying to figure out how to make a living in the streaming era—or maybe not even make a living, but just how to connect meaningfully with listeners in the streaming era. They're all impacted by these practices.

One of the things that Spotify would say as a defense was that they turned to the stock music because they had found a need for content. But there's no shortage of music in genres like lo-fi hip-hop, jazz, classical or ambient that fill out these lean-back playlists, by musicians who could really use the boost.

For me, I'm always thinking more about those musicians who are really impacted by being removed from these playlists, replaced with stock music, or who have never been able to access these sorts of promotional opportunities in the streaming era.

There's a great excerpt of your book in Harper's Magazine that I read and loved and shared on Bluesky, and I had a surprisingly heated back-and-forth with a reader who I think was a musician himself. He said to me, essentially, “Look, this ghost musician stuff that you're talking about, this stock music—musicians have always taken stock music gigs to pay the rent, right? It's always been a tough job and, on some level, a gig is just a gig. So let's not shame musicians for taking these gigs, making stock music.”

And I know that you're not shaming them, but I wonder what you made of that argument: that this isn't as different as we may think.

Well, I would encourage that person to read my book because—

Yeah.

—I’m not shaming the musicians who make this work. In fact, there’s a whole chapter where I talk to a number of musicians who make this work. What I try to explain is that this practice is as deceptive for listeners as it is for them. When I was interviewing musicians who made music for companies that were part of the PFC practice, they—

That’s the perfect-fit content for Spotify.

Yeah. These musicians didn’t know anything about the broader arrangement that they had signed up to be part of. They would tell me things like they make their tracks, submit them, get paid, and don’t know what happens after that.

Their arrangements are dictated by the contracts between them and the production company that’s hiring them. So that’ll look different from contract to contract. Some of them told me that their arrangement was a buyout, where they were getting a flat fee for the master, and then maybe there were some other royalty rights that they were entitled to. Each company has its own arrangement.

I talked to a couple of composers from this songwriter advocacy group based in the U.K. called the Ivors Academy, and they talked about how, from their perspective, they felt that companies like Epidemic Sound, through these arrangements, were trying to buy composers out of their luck.

When you make production music, part of what you’re doing is making lots and lots of music. You never know which tracks might take off and take on a life of their own, then end up being a really sustainable source of income for you for years to come. They felt that, by encouraging these buyouts and this flat-fee model, companies were buying out composers’ luck. This was a change, and composers should hold onto their chances of a song going viral or being used in a commercial and being able to see some success off of it.

I think a lot of listeners to our show will be thinking about AI when it comes to the future of services like Spotify. We’ve talked on the show about services like Suno and Udio, which can basically generate new music along the lines of existing music.

To me, it just seems inevitable that, at least for these ambient, lean-back playlists, Spotify will eventually just start creating music on the fly using AI so that they don’t have to pay any royalty to any human artist or any production company. Is that happening already and we don’t know about it? Do you think that this is the future of this kind of music?

In recent years, Daniel Ek has said in the press that he finds the potential of generative AI music exciting, and that it could be great culturally and help boost engagement on Spotify. To me, that framing, or that optimism about it, signals that it would be unsurprising if that direction were explored eventually.

Though I should say that, in my reporting on PFC and ghost music, it’s not necessarily something that I directly observed. Companies like Epidemic Sound, which work with Spotify in this way, have directly signaled that they’re excited about the potential of their composers working with generative AI tools and are open to it. So it’s not hard to imagine.

I think that, from my perspective, there are certainly a lot of important concerns about generative AI content and its impact on streaming services. There already is so much AI-generated music flooding streaming services every day.

But I also think it’s important to remember the different ways in which different systems that might be called AI—systems of machine learning, automated recommendations, algorithmic recommendations, and personalization—have reshaped the way that people understand music over the past 15 years. Our recommended music and the context within which music is served and presented to listeners are equally worthy of consideration and critique.

Yeah. Well, let me put some of my own cards on the table. I have to say that Spotify often feels like a miracle in my life. I still remember being a high schooler who had to scrimp and save to buy a single CD for $18. I wanted to know so much more about the canon of pop music, and it was just completely inaccessible to me. But now Spotify exists, and I can just inhale it.

Liz, your writing on this subject really unsettles me because it reveals the extent to which Spotify has built systems to manipulate my listening in ways that are completely invisible to me. I do worry that, as the years go on, my taste in music is becoming less and less my own.

So I wanted to ask you how you might reconcile those two things, or how you think I might reconcile them, and how you try to personally cultivate your own taste in music in this age.

Absolutely. Yeah. Something that I think can be complicated or seem like a contradiction in some ways is that I’m actually in favor of universal access to music. I don’t think universal access to music is a bad thing, and I’m someone who came up in the era of Napster and file sharing. Being able to access a lot of music that way was really influential and formative for me, I should say.

So I don’t necessarily think that universal access to music is the problem. For me, I think it’s more the rise and championing of lean-back listening, of a sense of passivity, and of a devaluing of music—not just on a financial level but, in some ways, on a more cultural level—that I think happens when this relationship with music is watered down in this way.

Of course, Spotify and streaming didn’t create these conditions. They didn’t create the idea of the lean-back listener, for example. But I think that this way of engaging with music has been really exacerbated by streaming, by making lean-back listening the most frictionless way to engage with music.

I think that optimization and frictionlessness have been really disastrous for culture beyond even music. I’m a music critic and a cultural critic, and I think that thinking is really important. I think that encouraging people to think is really important.

When I talk to former Spotify staffers, when I look at the ways in which optimization and frictionlessness are seen as these goals of streaming curation in the platform era, there was one interview that I did with a former staffer who talked about the goals of the curation ecosystem as trying to reduce the cognitive work that people have to do when they open the app.

In my book, I trace Spotify’s long-term goal to create a product where the user can open the app and be met with a perfect recommendation at the perfect moment without having to do any deciding, choosing, or thinking.

Well, it seems like the TikTok model just applied to music rather than video. Lots of social media platforms have had the same realization: If we just remove all of the choice from the user and give them an endless scroll of algorithmically selected content, we can keep them hooked for longer because, statistically, most people don’t want to do the work of searching out the things that they want.

But that is a very cynical view, and unfortunately, I think it does appear to be profitable. Spotify just had its first profitable year, so something they’re doing is working. But I’m not sure it’s working for culture at large.

Yeah.

Yeah.

One of the former Spotify engineers that I spoke with referred to the TikTok feed as the ultimate distillation of lean-back listening. You’re not putting in any input. You’re just giving signals based on how long you linger on something.

I guess what I was trying to get at earlier is that, as a critic—as someone who thinks a lot about the way in which music is contextualized as a way of opening up the possibility of new connections with music—to me, this idea of encouraging people to think less about what they’re listening to is troubling.

I think that this process of listening, thinking, deciding, hearing things that you don’t like, deciding why you don’t like them, and being challenged by music that is outside of your comfort zone—this is all important from my perspective.

Yeah.

Yeah. I think there should be a ghost-musician stage at Coachella this year, where everyone who’s made these playlists that we listen to all year long just gets on stage. It’s just little twinkly piano from, I don’t know, 8:00 to 9:00 p.m.

I love that.

Let’s give these people some attention.

And here’s my feature request. I want a toggle switch on Spotify where, before I go to sleep and put on my sleep playlist, I can just say, “Only use human musicians,” because those people are performing a valuable service for me.

I love the idea of some obscure classical pianist just waking up to a giant royalty check from Spotify because millions of people have been using their music to fall asleep.

Spotify, of course, is famous for its giant royalty checks.

I write in my book that there’s no shortage of really inspiring ambient music to be discovered these days, made by actual musicians. So if there were going to be an ambient stage at a major music festival, I would hope that it featured those artists.

And some of the musicians making the music for these ghost-artist playlists have their own creative practices, so I would also hope that they'd be able to share that music.

No, I'm very glad that we had this conversation, Liz, because I think I am realizing that I am the problem. It's me.

Which was my hope with this conversation, so I'm excited too.

And I do think that the sort of loss of agency and loss of taste, essentially, that you're describing in your book applies to me. I think that I used to be a person who sought out specific musicians and artists, and I think I have just gotten lazy about that. So I do actually feel challenged by what you have told us today, and I am going to start being more intentional about the music that I choose.

Well, maybe on our way out, Liz, as you mentioned, you are a critic. Do you have any ambient, chill, lo-fi artists that you might suggest to Kevin so that when he's in more of a lean-back mode, when he wants to hear the genuine article and not the dollar-store version, he might be able to enjoy it?

Okay. I will say that probably my favorite ambient music of the past few years has been by Emily A. Sprague, who is the singer of this band called Floras but also makes ambient music that is really beautiful. I would recommend checking that out.

Perfect. Human beings recommending music to each other, just like in the old days.

How do you spell that? Emily A. ...

S-P-R-A-G-U-E.

Yeah, and just for a change, Kevin, try listening to Emily while you're awake. I just want you—It might actually improve your appreciation of music, would be my guess.

That's a good tip.

Liz, fascinating conversation. Congratulations on the book. Thank you for joining us on Hard Fork.

Thanks so much for having me.

When we come back, we're going to tell you what AI tools we're using in our new segment called Tool Time.

Well, Casey, it is time for a new segment that we're calling Tool Time.

It's Tool Time.

Now, if you are a '90s kid, you might remember Home Improvement.

Only '90s kids will remember Home Improvement.

Tim “The Tool Man” Taylor had a TV show called Tool Time, but this is different.

That's right, because whereas Tim Taylor was often working on his car in his garage, we are working on laptops in our home offices.

Yes, this is more of a knowledge-worker Tool Time.

Yeah.

But we do get a lot of questions from listeners about the tools that we're using, whether it's AI to help us be more productive at work or maybe just in our personal lives. People want to know what is going on out there and what the latest and greatest tools on the market are.

Yeah. They hear us doing Hard Fork, and they think, “They clearly are not doing this without computer assistance. There is something they're using to aid themselves.” And today we're actually going to tell you what those things are.

And Casey, this is a segment about AI and AI tools, so we should make our AI disclosures.

Well, here's one for you, Kevin: Casey's boyfriend works at Anthropic.

Kevin works for The New York Times, which is currently suing OpenAI and Microsoft over alleged copyright violations related to the training of large language models.

Is that the first time we've ever referred to ourselves in the third person on this show?

No, but I kind of like it.

All right.

9. Deep Research Delivers Reports

So the first tool that we want to talk about on Tool Time today is Deep Research. This is a new feature from OpenAI. It's available to subscribers to the $200-a-month ChatGPT Pro plan, although they have said that they plan to make it more widely available. And Casey, just explain Deep Research for people who haven't heard about it or tried it.

Sure. So Deep Research is a way to get a lengthy, extensive, detailed report on a subject that you are interested in. You access it through the normal ChatGPT interface, but when you type in your query, you click a button that says “Deep Research,” and then Deep Research will read your query. It'll ask you a few follow-up questions so it can really hone in on what you want, and then it will use this as-yet-unreleased model called o3. What sort of reports have you been asking Deep Research to create?

So I have been experimenting with this, and I've been really impressed so far. I have done it with a couple of different topics. One of them, I was just curious about the history of the term AGI—artificial general intelligence—and so I asked Deep Research to make a report for me about the intellectual history of this term and the idea behind it: a computer capable of doing everything the human brain can.

First it asked me some questions to clarify. It said, “Are you looking for an academic-style research document with citations or a more general historical overview? What time frame should I focus on? Do you want to include science fiction or just general references to scholars and other people talking about AGI?” I answered those questions, and then it went away for 10 minutes. It consulted 36 sources, and it returned a 7- or 8-page report about the intellectual history of the term AGI.

Right. And as you read through this report, what did you notice? Is this a subject where you actually had a lot of familiarity with it, so you were able to follow it, or was this something where you really didn't know a lot of the information that it was telling you?

So I have done this kind of research project before. For my last book, I did a document very similar to this. It was really good. It went all the way back to 1956, to the Dartmouth workshop, where the term artificial intelligence was coined. It went back even further than that, into the 17th century, when Thomas Hobbes talked about how reasoning was akin to computation.

So it just traced the entire intellectual history of this term, and I didn't see anything obviously wrong in it. When I started checking some of the citations, it all looked pretty good.

Well, I have been doing my own explorations with Deep Research, and I have to say, this feels like the first good AI agent. There's been a lot of talk over the past 6 months in particular about how this next era of AI is going to be these advanced tools that can do multistep projects on your behalf in the background while you're not paying attention. I haven't used any so far that felt like they were meeting that bar until this one.

I'm somebody who writes a column 3 times a week. That column is usually rooted in some set of historical events that I need to refresh my memory about. Because it involves subjects I've written about before, I'm a little bit more confident as I use it. It does make mistakes, and I have to say, I have never done a Deep Research report where I have not found at least 1 mistake. Other stuff is actually true, and more importantly, it helps to structure my thinking a bit, right?

It can create a timeline of events for me. It can break out different ideas into different buckets and offer citations. I have to say, if I had an editorial assistant and I said, “Hey, in the next hour, put together a report for me about this sort of thing,” I would be surprised if they could do something that comprehensive in that short amount of time.

Yeah, so I think it's also useful for just more personal things. One of the things that I had Deep Research do: I have this stack of parenting books that I have been meaning to read ever since before my kid was born.

Mm-hmm.

And I just never got around to it. A lot of the advice in parenting books overlaps with itself or overlaps with other books, so it's not a very efficient way of understanding what you're supposed to do when a toddler is having a temper tantrum or something.

And so I basically just said, “Go off, read all of the things you can from this set of parenting literature, and give me the CliffsNotes.” It went out, and it did that pretty well.

Wow. And so now, for the first time, you'll know what to do when your toddler throws a tantrum. Which is what, by the way?

Well, I haven't made my way through the—

Okay.

10,000-word report yet, but I'll get there, and I'll let you know.

So we should say Deep Research from OpenAI. This is not the only deep-research tool on the market. Google also has a product called Gemini Deep Research. How would you say this stacks up to other similar tools that you have tried?

So, in short, Google's version is not as good. That is to be expected. Google is using a regular large language model, whereas OpenAI is using what they call a reasoning model, which is just built better to do this sort of thing. I think the fact that OpenAI asks questions before it gets to work is really useful because it does help you hone in on what you want.

You can also watch the chain of thought as it goes. We talked about this recently with DeepSeek. It does something similar, where you can try to understand what this model is doing, and if it's doing something you don't like, you can ask a follow-up later to maybe guide it better. And then finally, you just get much longer output.

So when I ran similar queries in Google's version and OpenAI's version, OpenAI's version was generally at least twice as long. That's a mixed blessing, of course, because now you have twice as much stuff to read, but in general, I found it much more comprehensive. The final thing that I would say is that there is just more stuff in the OpenAI Deep Research results that feels like thinking.

I know that will drive some people crazy, and they will scream at us and say we're anthropomorphizing these things, but I'm telling you: While I do not think that the AI is sentient, I do think it can create very good human reasoning that can verge on the insightful. And that's really powerful, and it is something that Google's version cannot yet do. What do you think?

Yeah, I think Deep Research is really useful, and I think it's potentially a very big deal. A lot of white-collar knowledge work is about research. That is one of the fundamental—

Mm-hmm.

—tasks involved in jobs like consulting or even finance or journalism. Knowing the capsule history of the thing that you are writing or thinking or preparing a presentation about is often quite useful and pretty time-consuming. So I think the implications of tools like Deep Research on the white-collar labor market are potentially very steep.

But just as a tool, I think this is very useful already for people who want to quickly get up to speed on new topics. It's a very good learning tool. I've been using it to teach myself things. So right now, you only get 100 queries per month, even if you pay $200 a month to OpenAI for the Pro subscription.

It is very compute-intensive, and you are somewhat limited in what you can do. But I think we should keep tabs on this, and I'm personally going to keep my Pro subscription just so that I can have access to this.

I feel the same way. I subscribed to ChatGPT Pro within the past couple of weeks because I wanted access to this Operator agent, which it released a week ago, which we talked about in our most recent episode. I used it, and I wrote about it, and I thought, “I don't want to use this anymore. It's not that good.”

So I was truly getting ready to cancel, and then OpenAI said, “Well, if you subscribe to Pro, we'll also throw in these 100 Deep Research queries a month,” and I thought, “That actually might be worth $200 a month to me.”

Totally. All right. Next tool on our list. This is a tool that I've been using for the past week or two called Granola AI. Casey, are you a Granola user?

I am, and this one tickled me because I had started using Granola in November, I think, and I just hadn't mentioned it to you yet. And so when you told me you were into it, I was like, “That's cool, because I am, too.”

So the way Granola AI works is that it's an app. You download it. You install it on your computer, and then any time you open a new video meeting, like a Zoom or a Google Meet or—

Or a Cisco Webex.

Or a Cisco Webex if you're still at one of the 3 companies in America that still uses that. You can have Granola take notes on your meeting, and what it does is interesting. It's not recording the meeting.

I'm sure you've also seen these meeting note-taking tools where the robot joins the video meeting as a hidden participant and records and transcribes. Granola works slightly differently. It doesn't record the meeting. It basically just takes the sound that's coming out of your computer and transcribes it in real time, and then presents you with a pretty detailed summary of what happened in the meeting.

So if you are a person who likes to take notes on meetings, this is a good replacement for that. It's also got some cool features where you can chat with the meeting transcript afterwards, and you can say, “What did Bob say? What are some action items that might come out of this?”

I used it to say, “What was Kevin's worst idea this week at our editorial planning meeting?”

I have found this very useful. What about you?

Yeah, I have as well. And look, I'm sure that at this point, people have seen these AI note-takers. They might be wondering, “What's so special about this one?” To me, what has made it stand apart is these summaries that it gives you after the meeting.

It's really good at identifying, “Here were the most important things that came out.” “Did you talk about a milestone in the meeting?” We're going to put that at the top. “What were the things you wanted to work on?” That's going to be at the top. And so it's just really smart, and they have different templates depending on what you're doing.

So Granola, I think, some of their first big users were venture capitalists, and many of the meetings that VCs are taking are people who are pitching them for their startups. So Granola has a template for that. The AI essentially knows what information to look for that is going to be useful to a VC afterwards.

Similarly, if you have a one-on-one meeting with the same person every week, Granola has a template for that. So it's really bringing in a lot of structure to the kinds of regular standing meetings that people have, and just making those notes super useful.

Yeah, I like the feature where if you have a bullet point of something that happened in the meeting, you can click on it in the Granola transcript, and it will enhance that by giving you a direct quote from the part of the conversation where you were talking about that thing. It hones in on the sentence or the two sentences that most directly talk about the thing in the bullet point.

Yeah. Now, I will say Granola is not perfect for me as a journalist, because I want to be able to use this for my interviews as well, but because it is not keeping a recording, I have to use something else in addition, right? Because sometimes I actually do need a direct quote, and I need to double-check it to make sure I'm quoting the person absolutely accurately, right?

So I actually just wrote in to the hello@granola email address and was like, “Hi, I'm a journalist. I would really like it if I could do that.” And they wound up putting me on a meeting with the CEO, and I got to make my case in real time.

What I learned was that they've been nervous to do this because they like the fact that they don't keep recordings. It feels much more private and secure, and I respect that. I think it's good to build technologies that preserve privacy, but I'm like, “Man, if Granola did this, then I could get rid of my other thing that does the—

Yeah.

—that does that part for me.”

Yeah, but we should talk about this, because this is an interesting question that's come up a few times in my usage of this, which is that because it is not joining your meetings, it can be running in the background without the other person that you're talking to, or the people that you're talking to, knowing. So do you have any privacy concerns about using a note-taking AI like this without informing the other person?

As a podcaster, I assume I'm being recorded at all times. And when I'm not, I get upset because I think, “We could've used that for the podcast.”

But certainly when you are recording somebody in an interview setting, you always want to tell them that you're doing that upfront. Obviously, there are circumstances where you don't want to be recorded, or the other person doesn't want you to record them. You kind of have to work that out.

But I think for the most part, if you're in meetings, it's because you're generating some sort of information that you want to use afterward, and having a tech tool that helps you do that makes all the sense in the world to me.

Yeah. Granola has a page up on their website saying, “You should definitely get people's consent before you do this.” I did not get the consent of our editorial team before I started using this in meetings, although I did notify them afterwards, so I apologize.

Yeah.

I'm sorry.

Well, let—

That was bad of me.

—me just say, you're in a lot of legal hot water, my friend. So lawyer up.

All right. Casey, the third tool that I want to talk about today is not really a tool.

Yeah.

It is a request for a tool.

Yeah.

For months now, I have been wishing and hoping for a tool that would essentially allow me to automate my email.

Mm-hmm.

Email overload is a huge problem for me.

Mm-hmm.

I get way more email than I can deal with. I spend hours a day trying to slog through my inbox. It is a huge time expenditure. And so one of the exciting things for me when large language models came on the scene was maybe I can have an AI take a first pass at responding to my emails, or at least populate a draft for me that I can just go through and click send on or edit to my own liking. But that tool has not arrived yet, at least in a form that I have used.

Casey, when it comes to AI and email, what have you tried? What are you using? What's your level of automation of your own email inbox?

It is much lower than I want it to be, Kevin. I have all of the same frustrations that you do. When I look at email, I see a data extraction problem, right? There are only 9 or 10 different kinds of emails that I get. Some of them are pitches for me to write about. Some of them are people who are inviting me somewhere. Some are people who want me to go on the radio.

It should be almost trivial at this point for some kind of AI to just notice that, bucket it out, draft responses, and let me just click a couple of buttons and be done with it. But nothing I have tried gets close.

I have tried 2 different AI-enhanced email apps so far. One is in a very early beta: Notion, which makes the popular collaboration software. They have a Notion Mail client. I don't want to give a full review of that one because it truly is in beta. They're making a lot of changes over there, but I would just say that so far it has not been able to do what I wanted it to do.

The other one I tried is called Shortwave, which I paid a subscription for and which promised to do what I just described in terms of extracting all of that data out of my email. I found it couldn't do that. I remember running the query, "Of the emails in my inbox, which ones have action items that I need to do?" It completely failed to do that.

So I canceled. The CEO emailed me and was like, "You know, we're changing it, we're making it better." I'm sure they have improved it since the last time I used it, but I have felt burned by my experiences with AI email, and so I'm no longer using them. What about you?

I've been interested in a few different solutions here, but one of the things that worries me about these third-party apps is that I don't want to send— I have 20 years of Gmail—

Yeah.

—sitting in my account, and I don't want to send all of those emails to a company that I don't necessarily trust to keep that information private. All these apps that are popping up want to learn how to write like you, which involves ingesting a ton of your previous emails, and that's just a privacy concern for me.

I don't want to hand over that many years of my email to OpenAI or Anthropic or another company without knowing if they're training on that or retaining that in some way. So what I've been trying to do is build my own homespun email autopilot app.

Hmm.

A couple of days ago, I went into Claude, and I just said, "Here's my problem. I want it to all run locally on my machine so that it's not sending my emails anywhere else, and can you help me build it?"

And can it?

Well, TBD, because so far I've only been working on it for a few days. What I have is a bad prototype now. Claude helped me install local LLMs on my machine. We've been going back and forth about how this app should work, how it's going to learn from an archive of my old emails, and how to write like me, but it is still pretty buggy.

It did start responding to spam emails for me, so I—

It said, "This sounds amazing. I'll take all that Viagra that you got."

So I still have to do some more fine-tuning, but I think I am rapidly approaching the limits of my own very limited programming skill. And so if there are any Hard Fork listeners out there who are programmers—and I know there are—you would earn my undying devotion and gratitude if you helped me build an app that would do essentially the following: 3 or 4 times a day, scan my inbox—

Mm-hmm.

—pick out anything important, and draft a reply to it.

Mm-hmm.

Populate a little box, and give me 1 button that I can hit to send it or another button that I can hit to edit it.

Mm-hmm.

Also, give me a digest every day of the most important things that happened in my email inbox—

Mm-hmm.

—and any action items. If you're feeling fancy, connect to my calendar, but you don't even have to really do that. I would just settle for the email drafting tool that I just described.

Yeah. I think that's beautiful, and I would like to make a prediction, Kevin.

What's that?

We are going to publish this podcast, and you are going to get several emails from people who make email apps, and they're going to tell you, "We can actually do this already." Then you're going to go through the trouble of setting it up, and you're going to find it cannot actually do that. I don't know why this happens, but this happens.

Okay, I need to send another message.

What's that?

This is to the people who listen to this podcast who work at the Google Corporation.

Mm-hmm. The total badasses.

The total badasses. I need you to do this yesterday. You have my email. You know everything about me. You have my browsing history. You have my photos. You know everyone that I've ever contacted in my life and everywhere that I've ever been and everything that I've ever searched for.

The fact that there is not a tool built into Gmail that allows you to put your inbox on autopilot is a failure of imagination, and I want it fixed.

This sounds like a great job for 2.0 flash thinking experimental apps, Kevin. That, of course, is a new model that Google released this week.

Okay, so Casey, let's end this Tool Time segment with a question from a listener.

Love a listener question.

This came in just today. It's from a listener named Ray Keen.

And we're keen to answer it.

He asks, "If you were to choose just 1 paid subscription"—I guess he means AI subscription—"which would it be?" So Casey, what is the answer for you?

You know what's crazy about this question, Kevin? I feel like my answer to it probably changed within the past couple of days.

Hmm.

Because I think the truth is that if I could only pay for 1 AI subscription, if I'm on a desert island with only 1 AI, it would be ChatGPT Pro.

Hmm.

And the reason really is Deep Research. We've talked before about how the AI labs are mostly at parity when it comes to the basic questions that people ask. Some LLMs seem to have a better personality. Maybe they're a little bit better at mentoring, tutoring, coaching, whatever.

But Deep Research felt useful to me in a way that made me feel like I am going to use this most days now, and I don't think I would want to be without it. So I don't know, maybe a week or 2 will go by, and the bloom will come off the rose, and I'll say, "Oh, yeah, this Deep Research thing—it turns out I don't actually want to read 3 10,000-word reports a day."

But right now, I feel like it. Obviously, at $200 a month, it is extremely expensive for a software subscription, but if I can only pick 1, I think it'd be that. How about you?

Yeah, I think it's good for people who have an interest in this stuff to at least try the latest and greatest coming from OpenAI, but for me, the answer to this question is Claude.

Hmm.

I pay for Claude, the Pro version. It's $20 a month, and Claude has some limitations. It can't browse the web. It's not good at everything. It doesn't have some of the same multimodal capabilities that some of the OpenAI models do, but it is just a very good daily-driver, all-around AI model for the things that I use it for.

Yeah, makes sense. Claude is really, really great. But I do wish it could browse the web, and I do wish it had some research feature or even just a reasoning model.

Yeah.

Yeah.