This is by far the best AI generated music I've heard yet, and marks the first moment where I started feeling some worry that this stuff is coming for music too.
Up til now, I've been convinced (foolishly, maybe) that there was something special about music that AI just wasn't going to be able to do well. Now, I'm not so sure.
I think modern popular hip hop is probably an easy target - the music itself is relatively simple, and especially for an absolute chart-destroying monster like Drake, there's a huge amount of singing / rapping / voice to use as training data.
I'm somewhat more skeptical on genres that involve complex instrumental playing and more dynamic vocal performances, but I think this track sounding so convincing is evidence that we're making big strides.
On the other hand, there's something about music that I truly don't think AI can touch, and that's this sort of undefined "cool" factor. I'm a rock / metal guy, and there's something just awesome to me about a group of guys up on stage rocking out, but the cool factor is definitely a big part of hip hop too.
Last thought - I'm not sure if this song is as good as I think it is, because I don't really listen to Drake. I recently ran across an AI generated Nirvana song, which was also pretty well done - but it was still just on the edge of the uncanny valley for me. Maybe big Drake fans would feel the same way about this, idk.
“Audio deepfake” would be a better description than “AI generated song”. Someone made the beat himself and wrote and rapped/sang the lyrics, then used AI to replace his voice with Drake’s and The Weeknd’s.
Yeah this is an important distinction that I don’t think is being properly communicated or emphasized. This is a song created by a human, with voice replacement.
That is a _huge_ distinction, which I didn't realize. I was trying out AIVA last week and fed it a couple Lotus songs I liked as source / inspiration - it correctly identified the key and generated a "riff" that was just going up and down the major scale.
I think, at this point, the burden of proof has shifted for skeptics to prove why AI won't be able to produce a certain kind of art.
Still, I think music will be the most resistant to these kinds of trends. Anyone who is familiar with classical music or jazz can tell you that even when the same piece is performed by two very talented musicians, there's often an ineffable and unquantifiable "it" factor that separates the merely incredibly talented from the truly great. Sometimes this is incredibly subtle, or isn't immediately apparent to those who aren't familiar with a genre, but especially because music directly involves the human voice and the movements of the human body, I find this je ne sais quoi much more readily apparent in music than any other artistic medium.
Does that mean AI won't be able to capture it? No, definitely not. But I do think as the marginal cost of producing derivative pop music goes to zero, people will gravitate towards more "organic" forms of expression that are more difficult to emulate through brute force. For example, there's a lot of bad noise music and bad free jazz out there. Most of it probably. I'm sure AI can already produce convincing bad noise music. But since that's already the case, the bar is much higher for AI to produce something that's actually good.
What is art but communication, the best art is communicating the feelings and emotions of the artist, I think an AI can create content that emulates art but without the artist there isn't any communication of emotion. Much modern music is so industrialized and derivative I would say the level of emotion al communication is practically zero anyway though so I'm sure AI will have a great career in the "arts".
The absolute last jobs to be automated by AI will be those that require highly flexible sensorimotor skills, demand real time reactions, and involve significant risks.
Like even right now trains could be 100% automated easily. But the risk involved of transporting so many humans that could die if something goes wrong in a freak scenario, and the material damage of derailing a train makes a human supervisor/driver quite viable.
A lot of light rail and subway systems are entirely automated today. As far as I know, most of the systems that haven't automated conductor work are either grandfathered into union negotiations that preserve those positions or haven't been able to justify the cost yet.
I've also been surprised that AI had hit art first, but in hindsight it makes sense. The AI we have now is imprecise, which is neutral to good for art, bad for sciences.
Humans still have more perspective when creating music or other art. They can create and maintain themes on more levels, for now, but AIs that are multi-modal / multi-paradigm / whatever-you-want-to-call-it will close that gap, too.
A human can look at an art rendering of a humanoid character and immediately spot symmetry errors, missing fingers, strange limbs or joints, and the AI didn't know any better because it was imagining an artistic rendering from a prompt the same as any other art piece. It doesn't "think" of the physical accuracy of what it's drawing. You might be able to ask GPT-4 to rate the physical accuracy of a bad midjourney drawing and get an honest response, in which case AI already has the ability to do better, it's just not trained that way... yet.
In writing music, humans can conjure and enforce large-scale patterns in music. An AI would currently get lost writing a symphony. An AI might construct an interesting harmony, but it wouldn't dwell on and play with the harmony the way a human composer would. Yet.
I think there's enough demand for non-trivial music and non-trivial artistic representations of humans that these problems will be figured out soon. Whether that involves different training, different neural net architecture or a conglomerate of different sub-neural-nets, or multiple independent neural nets with some filtering others' output.
Kind of hard to say this, but unless some miraculous slowdown happens, it's coming for the entire human experience by the look of it. As Max Tegmark put it recently, we're basically Neanderthals building homo sapiens. I don't care what people think about Lex as an interviewer, it's worth a watch just to be more prepared for where things are likely headed: https://www.youtube.com/watch?v=VcVfceTsD0A
Wait, what’s wrong with Lex Friedman? I’m not familiar with the guy, but after briefly listening to that video & checking his background…he seems fine?
Thanks for the link! To summarize the Nassim Taleb guy's criticisms as I understand them, he:
1. Is ignorant of the Lex guy's affiliation with MIT, and speculates about how tenuous the particulars could be.
2. Thinks Lex is misusing said MIT connection in unassociated emails.
3. Points to Lex's book-per-week pacing as indigestible given the material, concluding that Lex's persona is fraudulent, and his advocates are all idiots.
4. Uses Lex's courteous response to criticism, and subsequent choice to block said critic, as evidence that Lex is a "fake." There's allegedly supporting evidence of this regarding Kanye West/Vladimir Putin? (I'm ignorant of the reference.)
Full disclosure -- I don't use Twitter and gradually lost the mental wherewithal to slog through the "discourse" within that thread. Let me know if I missed/misunderstood something important.
I'm frankly more confused now. A popular podcaster apparently likes revisiting mostly high school reading material, is polite online, and leans on their institutional affiliation. Okay? Did he also like...shoot someone's grandma or something? I feel like a restaurant patron eavesdropping on a dysfunctional couple having a meltdown over ordering an appetizer that isn't reaaaally about ordering an appetizer. What the hell did this guy actually do?
I love Lex Friedman interviews personally, he tend to just kinda let the person he interviews talk a lot and his follow up questions tend to lean towards exploring things the people he interviews are passionate about. I think the criticism is that he generally doesn't push back against those he interviews but I don't think that should be the default style for all interviewers necessarily.
I don't know why, people on here always seem to accuse him as being a bad interviewer. Maybe I shouldn't have said that. I think he is fine / good. At least he discusses important things.
I agree and I think this is a banger. But your last couple lines are important here: we're both not huge Drake consumers so we can't tell if this is uncanny valley material. Can you link the Nirvana one?
To my ear, it's surprisingly good, but not quite good enough that I'm fooled by it. It might be good enough that if it was slipped into a Nirvana playlist I might not notice - I like Nirvana OK but have never been like, intimately familiar with their catalog.
Up til now, I've been convinced (foolishly, maybe) that there was something special about music that AI just wasn't going to be able to do well. Now, I'm not so sure.
I think modern popular hip hop is probably an easy target - the music itself is relatively simple, and especially for an absolute chart-destroying monster like Drake, there's a huge amount of singing / rapping / voice to use as training data.
I'm somewhat more skeptical on genres that involve complex instrumental playing and more dynamic vocal performances, but I think this track sounding so convincing is evidence that we're making big strides.
On the other hand, there's something about music that I truly don't think AI can touch, and that's this sort of undefined "cool" factor. I'm a rock / metal guy, and there's something just awesome to me about a group of guys up on stage rocking out, but the cool factor is definitely a big part of hip hop too.
Last thought - I'm not sure if this song is as good as I think it is, because I don't really listen to Drake. I recently ran across an AI generated Nirvana song, which was also pretty well done - but it was still just on the edge of the uncanny valley for me. Maybe big Drake fans would feel the same way about this, idk.