Decoder with Nilay Patel
00:43:23 · S7, S9
Ad[00:43:23] S7: Support for this show comes from Outshift by Cisco, which is Cisco's incubation engine for frontier technology. Imagine a company where every department can talk, but none can work together. Decisions stall, people spend all day chasing updates, mistakes slip through the cracks. That's the challenge facing multi agent AI today. The fact is when AI agents work in isolation, more AI doesn't mean better AI. And if agents can't work together, you're just scaling the confusion. Outshift by Cisco is here to help. Outshift by Cisco knows communication doesn't equal collaboration. And in the age of AI, that difference changes everything. Outshift by Cisco is building the Internet of Cognition, an open source foundation for building multi agent systems. Agents share context, memory, and guardrails to get excellent results from design all the way to production. So instead of simply passing messages, agents can build on each other's work and move towards shared goals. Read the paper, experience the demo, and grab the code from outshift.com. That's outshift.com. Ad[00:44:30] S9: All new holiday drinks are now at McDonald's, like the frosted cranberry splash with Sprite, the Red Bull frosted cranberry energizer, and cookie butter coffees. It is the season for seasonal drinks at McDonald's.
00:44:51 · Nilay Patel
[00:44:51] S3: The initial debates about copyright law and music were about player pianos. Like, we've built huge institutions to manage these rights. I said I was gonna talk about Google a lot and the tension inside of Google when it comes to this is the next sort of nightmare law school hypothetical for you. Google has two conceptions of itself. YouTube describes itself as a licensing regime, right? There's a lot of content on YouTube, the creators make it, Google takes a license to it whether or not any creator ever gets to negotiate that license as a secondary, of secondary concern. There's a lot of music on YouTube and the industry had to fight its way into licenses on YouTube because music is existential for YouTube and the industry knows it has that power. There's a real back and forth. Maybe the only industry that has true negotiating power with YouTube is the music industry. And you can see this play out over and over and over again. On the other side of the house, Google search, it's a huge business is they will describe it to you as a fair use regime. Right, we're gonna ingest all the photos in the world and let you search them and that's fair use and we're gonna litigate our way through that. We're gonna adjust all the books in the world for Google Books and that's fair use and we're gonna litigate that and they have done that over and over and over again. You can go work with the people at YouTube, they're very reasonable. Nilay Patel a smart guy, he understands that he needs music on that platform to survive and he will build you the tools that you're asking for and he has. And then the people running Google are like, we're just going to eat everything and training an AI model is fair use. And just in that one company, they are totally at odds with each other. They're going to kill each other somehow. How do you make sense of that?
00:46:27 · Mitch Glazier
[00:46:27] S4: What's so fascinating to me about YouTube is that it is the first real lesson in the arc of sort of the DMCA where we real... Where where I think a technology based company realized that even if you win the fight on intellectual property rights, if you don't have partnership through a license in the actual market where fans are, you're not going to have the service that fans really want. And so it was the first time that you know we we took everybody realized that, you know, you can win the case but lose the business. Mhmm. And music is such a driver of YouTube, and I think this applies to sort of AI too, that having a pristine... Having pristine copies, right, that fans really want, and then having marketing partnerships and partnerships with people who invest with and partner with creators who have a better sense of how to create the bond between that particular artist that they've been living with and sort of building a brand around and understand versus, like, sitting in a cubicle in a tech company and, like, trying to decide how to connect the fan with the artist in the best way. Like that former piece is irreplaceable. And I think YouTube realized that and developed Content ID even though they technically won in court. Right? And didn't have to so to speak. And now it's so ingrained, the partnership is so ingrained that there's a there's a dependency or a codependency and and that's how it's supposed to work. It also works because of this concept in copyright law called substitution. Right? Like if what you're creating substitutes for something else, it's probably not fair use. That's what we're arguing in our case against Suno. You're basically creating music that is competing on the same platform right next to human made music and you can click on either one and you're paying the same subscription fee and so you're affecting the economic pool and you're effectively substituting in the marketplace.
00:48:39 · Nilay Patel, Mitch Glazier
[00:48:39] S3: This is the fourth factor of the fair use. [00:48:41] S4: This is [00:48:41] S3: the fourth factor of [00:48:42] S4: fair use which, yeah, which court say is the most important which is are you affecting the potential market for the underlying work. And you know, if you are that sort of problem number one. Google in their case, unlike the YouTube case under the DMCA, Google in their case that they won in the Supreme Court was about creating an index and the idea is if you click on it, it could lead you to the work that you might then buy. So instead of substituting for the work, the search result doesn't substitute for the work, but it could lead you to the work, which at the end of the day may not have the same economic consequence. It could even have a positive economic consequence. At least they would argue that. So there's this weird thing like Google won its case and YouTube won its case, but YouTube formed a partnership despite it because it needed to do so because that's the best fan experience. And Google didn't because it wasn't as related to the fan experience. Fan experience was potentially pointed to by their activity. So that's my weird, like, [00:49:46] S3: lost world take on the two running Google, if you're if you're in charge of Google Gemini, you need all of the data in the world. [00:49:55] S4: Yes.
00:49:55 · Mitch Glazier, Nilay Patel
[00:49:55] S4: Yes. [00:49:55] S3: Right, like that is the that is just how it works. To make the models better and to win against OpenAI you need more data and all of those companies just took it. I do not think Google came to you and said can we license all the music in your database Even though we already have a license on YouTube. In fact, they just scraped YouTube, right? I don't think creators have like woken up to the fact that they've given a lot of economic value to Google and its models. The other AI companies also scraped YouTube and Google can't yell at them because then they will have to reveal that they scraped YouTube and this is about as tense a fight in AI industry as there is. [00:50:30] S4: Yes. [00:50:31] S3: You're gonna come to Google at some point and you're gonna say no you got to pay us for that training. Right? I know you're suing Suno but that feels like a little test balloon for the people with the real money. At some point they're gonna say, no, this training was fair use, right? We're watching this dynamic play out in dozens of these cases across the industry. How do you make sense of that? Where you're gonna say, you're gonna kill us. If the AI models flood YouTube with fake music, you're actually gonna kill the supply of the data that you need.
00:50:31 · Nilay Patel, Mitch Glazier, Nilay Patel
[00:50:31] S3: You're gonna come to Google at some point and you're gonna say no you got to pay us for that training. Right? I know you're suing Suno but that feels like a little test balloon for the people with the real money. At some point they're gonna say, no, this training was fair use, right? We're watching this dynamic play out in dozens of these cases across the industry. How do you make sense of that? Where you're gonna say, you're gonna kill us. If the AI models flood YouTube with fake music, you're actually gonna kill the supply of the data that you need. [00:51:03] S4: So there's the there's the legal side and the business side. On the on the legal side, obviously, we're bringing this case as a precedent setting case to show that it is not fair use to train on music. Whatever exemption you wanna think of, it's hard for us to imagine how it applies. Right? I I don't I don't think national security is at risk because you have to copy Drake. I I just I just don't I don't think [00:51:26] S5: that there's [00:51:27] S3: What do we play in, like, the war in Panama? We, like, played Metallica at them at full volume. [00:51:31] S4: Right. Right. Now... Right. Maybe maybe it's... Right. Maybe it'll be used for, like, you know, for soft power or, you know, propaganda blasting on the screen. [00:51:39] S3: The Clash was very upset when Rock the Casbah was the soundtrack. I'm I'm just letting you know.