Lex Fridman: evaluation

Transcript context

…So when AI, when I first started making these AI videos, it started back in 2023, I made my first one, and I would take my phone, come up in the kitchen, I’d play a song, and then my youngest, Layla, and I have three kids, and my oldest, Dylan, as soon as I played it, “Why are you listening to AI?” And it’s like, oh my God, instantly. It’s like, how do you know? Oh, it has this ringing sound in the thing. So it took me probably about four or five days to figure out, “Okay, what are they hearing that I’m not hearing?” So I did it, I separated all the parts, and what they’re hearing was the artifacts that are in the vocal reverb. That sound that were… That made incomplete- It just couldn’t do the ambiances correctly, right? Because it’s trained on… A lot of these AI programs are trained on very low bit-rate MP3s, right? So, they feed all this stuff in there. So, they’re getting really inferior information in the training process, whereas now when they make these deals with the major labels, they’ll get the multitracks, and they’ll get high-quality WAV files to train from, right? And whoever opts in, they get the solo vocal tracks. You know, if Ed Sheeran wants to do it, or Drake, or whoever wants to give their voice to it, let it do its thing, and then get the royalties from it. I’m not saying that any of them are doing it. I’m just giving an example. But every time that I would do it, I could be down the hall, and I would play something in my phone just to see if they’ll… “Why are you listening to AI?” They can instantly tell. Then it eventually started getting better. And then, it’d be like, “Is this AI?” I’d be in the car with Layla coming back from taekwondo practice, and she’s like, “Is this AI? Why? Does it sound like AI? Sounds like it could be AI.” And I’d be like, “Yeah, it’s AI.” She’s like, “Oh, it’s getting better.” And then I did this song for… It was an NPR interview, and I created a song with a fake artist. And the song was called “Neon Ghosts,” and I played it for Layla in the car. She’s like, “Can you separate the tracks?” I said, “Yeah, I have them separated back home.” “Okay, I want to go down to hear it.” So, we go down to the studio, and I play it for her, and she listens to the soloed vocal. She said, “Wow, this is really realistic.” “This is very hard to tell, even with the soloed vocal.” I think the room for creativity right now for humans is lyrics. It seems like the lyrics that are being generated, they lack soul somehow. And that’s- I don’t know the words correctly. I mean, they can be incredibly sophisticated, but there’s something, the edge is not there. Some kind of edge that we want in our lyrics. Some kind of surprise, but not cringe or not cliché. Or something truly novel in the lyrics. But if that’s the case, it’s kind of sad that that’s where the creativity has to come from, but not from the music. Because then if we can create very realistic music that sounds really damn good, where’s the role of the musician there? I think the role of the musician is that in actually… If they use AI to assist them in coming up with ideas, they could as a creation tool. Then the musician… Like, some of the stuff is just not high quality sonically. So, the musician goes in and redoes stuff and changes things and adds parts, and then they actually do music production, and maybe they re-sing the parts, and they change the stuff. And then it’s just basically like an idea generator, and I think that that’s a great use of AI, is for that.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.