There’s a “new” track by the Beatles, grafted together from a demo by Lennon in the 70s, guitar by Harrison in the 90s (Lennon was killed in 1980. Harrison died in 2001), and new strings and drums from McCartney and Starr.
Plus a lot of production, using AI.
That demo from Lennon: The very original recording is just John playing the piano with TV in the background
– that’s Giles Martin, producer, son of George Martin.
Other tracks layered on similar poor recordings resulted in Lennon’s voice sounding “ghostly.” That was 1994.
So this time they used technology developed for Peter Jackson’s Get Back documentary on Disney+ (which is incredible btw).
“Essentially, what the machine learning does is it recognizes someone’s voice. So if you and I have a conversation and we’re in a crowded room and there’s a piano playing in the background, we can teach the AI what the sound of your voice, the sound of my voice, and it can extract those voices,” Martin said.
Listen to the new track here: The Beatles - Now And Then (Official Audio) (YouTube).
It’s… pretty good?
Like, musically, it’s ok. I find the strings arrangement a little too much maybe? McCartney’s taste has been so era-defining that, weirdly, his work starts to sound generic. (I say this as a McCartney fan! Read 64 Reasons To Celebrate Paul McCartney and be convinced.)
But it’s a lot and to my ear, it’s too smooth.
Because what really shines is Lennon’s voice. My goodness have they done a good job with that. They should have given it more room.
Mechanically recovered from the slurry of old tape recordings or not, 50% reconstruction or not, Lennon’s voice shines through time and up through the muddy waters of AI algorithms. It is poignant and beautiful to hear him sing. Familiar and unfamiliar all at once.
Mind you I’m in Liam Gallagher’s camp given what he said in The Guardian: The Beatles could shit in my handbag and I’d still hide my polo mints in there.
ASIDE:
Here’s a Wild Palms reference seeing as I’m a massive fan and even from 1993 they pinpointed this modern era before any of us:
Episode 4, after lounge singer Chap Starfall has been murdered by the Friends, the Senator et al have a hologram of him playing as background music. He was their buddy.
Tabba Schwartzkopf is watching: Hate to say it but I like him so much better since he died. That posthumous quality really makes me shiver.
Look – Wild Palms nailed both the nature of the technology S-curve and the effect of VR/the metaverse/synthetic reality/fake news/whatever you want to call it, all as background colour, wrapped up in a melodrama about LA media through the eyes of a patent attorney. I will never miss an opportunity to evangelise.
I’m trying to figure out how I feel about Now and Then.
Is there any legitimate difference - poignancy aside - between this AI extrapolation and, well, me never having heard a track before?
I have an ANALOGY.
Dummy by Portishead – released in 1994.
A contender for the best album of my lifetime. A haunting trip-hop soundtrack over 30 years.
I copied it onto tape for my car. Later ripped the CD into iTunes, listened endlessly there.
Portishead’s third album Three came out in 2008. In terms of which group has a better three album oeuvre: there is none.
I was never a Spotify listener. But Apple Music shipped in 2015, and at some point the streaming edition of the album shouldered out my ripped version, and:
There was a new track.
I can’t even discover when It’s a Fire (YouTube) appeared. It’s not on the U.K. original release.
It’s probably my favourite track now because each time I hear it, it still feels brand new. Imagine your favourite album for two decades suddenly has an extra song! It’s wild.
So my control experiment is right here. Music spear-fishes both kinds of memory - emotional and episodic - and reliably hauls up exotic and forgotten species from the deep past. And It’s a Fire has none of that for me. No gang of mates smoking in a front room. No silhouetted dark woods out of the car window in the witching hour. No bar, no working in a cafe, no running up Parliament Hill on Hampstead Heath at sun-up, no residual sense recall of the feel of the cracked plastic compact disk case packed and unpacked with my uni belongings. Simply: the music.
AND YET, even without all of that, I love it.
Here’s my conclusion:
Yes there is some (large) component of the “unit” of music/art/etc which is its subjective significance. But it turns out that this isn’t essential. The work can stand alone.
How about authenticity?
Does Now and Then lose something because Lennon wasn’t in the room?
Can I hear inauthenticity? I suspect not. I can hear the pre-2020s consequences of inauthenticity: a song without authenticity sounds empty.
But here in the 20s that has changed. We can say that Now and Then is a patchwork construction (but isn’t all studio-produced music) and macabre certainly (but as Schwartzkopf said, doesn’t it make you shiver?) - but it isn’t hollow. It’s not low quality.
So authenticity doesn’t matter either.
We’ve seen AI-extrapolated art before (as previously discussed) and it just hasn’t been very good. Like, it’s novel but it lacks something, and my preconception has been that there was no way that AI art can be good because it doesn’t have the original human touch, or the provenance (societal or personal).
But what I believe now is that it’s do with the effectiveness of the machine.
AI art will be good. We will one day have new Van Goghs, new Beatles tracks, new episodes of Firefly, and so long as they’re good - which they will be - my prediction is that we won’t mind where they come from. Quality will overcome it all. Caring about the origins will be nerdy like specifying which show runner you preferred on The West Wing.
So we should lean into extrapolated art maybe?
And that would be an easy thing to do?
Simon Willison recently did a deep dive into OpenAI’s new image synthesis AI: Now add a walrus: Prompt engineering in DALL-E 3.
The random number seed caught my attention. Simon
- generated an image from a prompt
- asked ChatGPT for the seed for that image:
1379049893
- tweaked the prompt to add a bow tie, maintaining the same seed
- and got an almost identical image, only with the asked-for bow tie.
Simon: I’m pretty stunned by this.
It’s a surprising result! My mental model of AI image synthesis has been that (a) yes images look great, but (b) they’re pretty chaotic. i.e. a change in the prompt, even an extra bit of punctuation, will send the AI spiralling off in a different direction and the resultant image will look very different.
What this deep dive shows that is that variance in the results is down to a random number generator. That’s what the “seed” is: it’s a number to feed into the image synthesis. The seed is randomly generated – but now I need to update my mental model to state that the image synthesis itself is actually deterministic.
(I’ve tried to reproduce Simon’s experiment but ChatGPT will not tell me the seed for DALL-E-generated images. So I think that hole has been closed.)
The consequences of there being a seed:
It makes AI synthesis way more reproducible. That means you can sit there tweaking and tuning a prompt and narrowing in on something intentional – this makes AI much less of a novelty, and much more like the creative process. It will reward work and skill.
It also reminds me that the words of the prompt become numbers too: vector embeddings, which can be mathematically combined.
As AI imagery, so AI music.
Now John Lennon probably isn’t a 10 digit integer. The seed isn’t deterministic like that. But passable Lennon probably is a 1,024 dimensional vector in embedding space.
And given that, there’s not just Now and Then but an infinity of possible extrapolated Beatles tracks.
We’d have to figure out the ethics and the IP (both fascinating rabbit holes in their own right) but perhaps we should lean into that, instead of pre-emptively blocking the possibility in the UI, which is what happens now, because who knows what would happen next.
There’s a “new” track by the Beatles, grafted together from a demo by Lennon in the 70s, guitar by Harrison in the 90s (Lennon was killed in 1980. Harrison died in 2001), and new strings and drums from McCartney and Starr.
Plus a lot of production, using AI.
That demo from Lennon: – that’s Giles Martin, producer, son of George Martin.
Other tracks layered on similar poor recordings resulted in Lennon’s voice sounding “ghostly.” That was 1994.
So this time they used technology developed for Peter Jackson’s Get Back documentary on Disney+ (which is incredible btw).
Listen to the new track here: The Beatles - Now And Then (Official Audio) (YouTube).
It’s… pretty good?
Like, musically, it’s ok. I find the strings arrangement a little too much maybe? McCartney’s taste has been so era-defining that, weirdly, his work starts to sound generic. (I say this as a McCartney fan! Read 64 Reasons To Celebrate Paul McCartney and be convinced.)
But it’s a lot and to my ear, it’s too smooth.
Because what really shines is Lennon’s voice. My goodness have they done a good job with that. They should have given it more room.
Mechanically recovered from the slurry of old tape recordings or not, 50% reconstruction or not, Lennon’s voice shines through time and up through the muddy waters of AI algorithms. It is poignant and beautiful to hear him sing. Familiar and unfamiliar all at once.
Mind you I’m in Liam Gallagher’s camp given what he said in The Guardian:
ASIDE:
Here’s a Wild Palms reference seeing as I’m a massive fan and even from 1993 they pinpointed this modern era before any of us:
Episode 4, after lounge singer Chap Starfall has been murdered by the Friends, the Senator et al have a hologram of him playing as background music. He was their buddy.
Tabba Schwartzkopf is watching:
Look – Wild Palms nailed both the nature of the technology S-curve and the effect of VR/the metaverse/synthetic reality/fake news/whatever you want to call it, all as background colour, wrapped up in a melodrama about LA media through the eyes of a patent attorney. I will never miss an opportunity to evangelise.
I’m trying to figure out how I feel about Now and Then.
Is there any legitimate difference - poignancy aside - between this AI extrapolation and, well, me never having heard a track before?
I have an ANALOGY.
Dummy by Portishead – released in 1994.
A contender for the best album of my lifetime. A haunting trip-hop soundtrack over 30 years.
I copied it onto tape for my car. Later ripped the CD into iTunes, listened endlessly there.
Portishead’s third album Three came out in 2008. In terms of which group has a better three album oeuvre: there is none.
I was never a Spotify listener. But Apple Music shipped in 2015, and at some point the streaming edition of the album shouldered out my ripped version, and:
There was a new track.
I can’t even discover when It’s a Fire (YouTube) appeared. It’s not on the U.K. original release.
It’s probably my favourite track now because each time I hear it, it still feels brand new. Imagine your favourite album for two decades suddenly has an extra song! It’s wild.
So my control experiment is right here. Music spear-fishes both kinds of memory - emotional and episodic - and reliably hauls up exotic and forgotten species from the deep past. And It’s a Fire has none of that for me. No gang of mates smoking in a front room. No silhouetted dark woods out of the car window in the witching hour. No bar, no working in a cafe, no running up Parliament Hill on Hampstead Heath at sun-up, no residual sense recall of the feel of the cracked plastic compact disk case packed and unpacked with my uni belongings. Simply: the music.
AND YET, even without all of that, I love it.
Here’s my conclusion:
Yes there is some (large) component of the “unit” of music/art/etc which is its subjective significance. But it turns out that this isn’t essential. The work can stand alone.
How about authenticity?
Does Now and Then lose something because Lennon wasn’t in the room?
Can I hear inauthenticity? I suspect not. I can hear the pre-2020s consequences of inauthenticity: a song without authenticity sounds empty.
But here in the 20s that has changed. We can say that Now and Then is a patchwork construction (but isn’t all studio-produced music) and macabre certainly (but as Schwartzkopf said, doesn’t it make you shiver?) - but it isn’t hollow. It’s not low quality.
So authenticity doesn’t matter either.
We’ve seen AI-extrapolated art before (as previously discussed) and it just hasn’t been very good. Like, it’s novel but it lacks something, and my preconception has been that there was no way that AI art can be good because it doesn’t have the original human touch, or the provenance (societal or personal).
But what I believe now is that it’s do with the effectiveness of the machine.
AI art will be good. We will one day have new Van Goghs, new Beatles tracks, new episodes of Firefly, and so long as they’re good - which they will be - my prediction is that we won’t mind where they come from. Quality will overcome it all. Caring about the origins will be nerdy like specifying which show runner you preferred on The West Wing.
So we should lean into extrapolated art maybe?
And that would be an easy thing to do?
Simon Willison recently did a deep dive into OpenAI’s new image synthesis AI: Now add a walrus: Prompt engineering in DALL-E 3.
The random number seed caught my attention. Simon
Simon:
It’s a surprising result! My mental model of AI image synthesis has been that (a) yes images look great, but (b) they’re pretty chaotic. i.e. a change in the prompt, even an extra bit of punctuation, will send the AI spiralling off in a different direction and the resultant image will look very different.
What this deep dive shows that is that variance in the results is down to a random number generator. That’s what the “seed” is: it’s a number to feed into the image synthesis. The seed is randomly generated – but now I need to update my mental model to state that the image synthesis itself is actually deterministic.
(I’ve tried to reproduce Simon’s experiment but ChatGPT will not tell me the seed for DALL-E-generated images. So I think that hole has been closed.)
The consequences of there being a seed:
It makes AI synthesis way more reproducible. That means you can sit there tweaking and tuning a prompt and narrowing in on something intentional – this makes AI much less of a novelty, and much more like the creative process. It will reward work and skill.
It also reminds me that the words of the prompt become numbers too: vector embeddings, which can be mathematically combined.
As AI imagery, so AI music.
Now John Lennon probably isn’t a 10 digit integer. The seed isn’t deterministic like that. But passable Lennon probably is a 1,024 dimensional vector in embedding space.
And given that, there’s not just Now and Then but an infinity of possible extrapolated Beatles tracks.
We’d have to figure out the ethics and the IP (both fascinating rabbit holes in their own right) but perhaps we should lean into that, instead of pre-emptively blocking the possibility in the UI, which is what happens now, because who knows what would happen next.