Music composed by artificial intelligence (AI) rather than humans is a controversial topic in 2017, for two reasons.
First: disbelief in some quarters that an AI will ever be able to create music as well as a human can. And second: fear that if it can, that’s the next tech trend sucking income away from working musicians.
Australian startup Popgun is hoping that its deep-learning based technology can prove the first view wrong, while playing a more positive role for human musicians than the second.
The company, co-founded by former We Are Hunted and Twitter Music exec Stephen Phillips, was part of the recent Techstars Music accelerator. Its first project is ‘Alice’, an AI that plays piano with humans.
Popgun sees Alice as a creative foil for musicians, as well as a teacher and playing companion for people learning to play piano for the first time.
Alice started playing the piano three months ago, and has been learning in the same way that children do: listening to thousands of songs and more experienced players, then mimicking them and practising lots.
She/it (more on that later) can now listen to and understand what a human plays, then reply with what she thinks may come next. Here’s what these human/AI duets currently sound like:
When Phillips talks to Music Ally, he’s back in Australia ready to hit the fundraising trail, while continuing to build Popgun’s team. It was in Brisbane that, after a bruising experience at Twitter Music, Phillips decided his next startup would focus on ‘deep learning’ technology.
“Deep learning is not the same thing as machine learning. It’s got the same general feel, and the ideas are 20 years old, but the tools are all 2-3 years old, and the big breakthroughs on how to do it are all new,” he says.
It was in Brisbane that he met Popgun co-founder Adam Hibble, who had been working on deep learning with a group of university friends. Phillips hired them for three months to tackle the challenge of using AI to tackle the challenge of music search: identifying similar songs simply by listening to them.
“They did it in three weeks. And I thought ‘this is the least interesting thing we can do with this technology if I can get six kids from uni to do it in three weeks’. It’s a feature for Apple or Spotify, and their teams will work it out, even if it takes them longer,” says Phillips.
“But the technique used to do it is really interesting: you understand something by your ability to mimic and generate it. So composition and generation becomes understanding. I realised that composition was the thing that was going to happen with this technology.”
Alice wasn’t the first idea for capitalising on this trend, however. In fact, when Techstars Music announced its first cohort of startups, Popgun’s description was very different:
“PopGun is a generative music company that measures success indicators from streaming services. We then use that data to train AIs to write music to match consumer habits.”
Ingesting, say, Spotify’s Global Top 50 chart in order to pump out more songs that sound like Spotify’s Global Top 50 chart? A fascinating deep-learning challenge, but also something guaranteed to raise hackles within the music world.
Phillips explains that Techstars Music head Bob Moczydlowsky – a former colleague at Twitter – originally contacted him to ask if he could mentor for the accelerator.
“I said ‘actually I’ve got a really young team, we’re going to try to write some software to have a top 40 hit!’ He said ‘good luck with that mate!’,” remembers Phillips.
“So I flew to LA and a room with these 250 execs, and Adam and I said we were going to build an AI that’s going to have a top 40 hit. They were like ‘Why the hell would you do that? That’s a terrible idea! You’ll be the most hated person in music!’ We wanted to do it just to see if we could do it…”
Popgun isn’t trying to do it any more. Phillips says one moment of realisation came from talking to a rock band who’d had a hit, and then had gone back into the studio trying to make another one.
“When they tried to make something that sounded like the first hit, it sounded terrible. And that’s because the song itself is an artefact of the creative event,” he says.
“The song is not the thing you should try to copy: all you’ll do is make a cheap version of something that already exists. Novelty comes when two intelligences collide.”
There was an example outside music to take inspiration from: the AlphaGo project from Google’s DeepMind. Announced in 2015, it was a computer program capable of playing Chinese board-game Go, with the ambition of being able to defeat the human world champion.
AI experts thought it would take a decade, while Go aficionados confidently predicted that it would never happen, thanks to what DeepMind admitted itself was the “beauty, subtlety and intellectual depth” of a game “a googol times more complex than chess”.
In fact, it happened in March 2016, when AlphaGo beat Go champion Lee Sedol 4-1 in a match held in South Korea. A machine triumphing over a human? Another way to look at it might be a machine showing a human – the best human at this game – new, inventive ways to play. As well as giving Popgun a new idea for AI and music.
“The AlphaGo guys realised they had to get an AI that could play humans, and then clone that AI to play itself a billion times. So the way to true novelty in music is going to first start with an AI you can play with. Then we can clone that and see what happens!” says Phillips.
“What does that sound like? What stuff will it play? And when we talked to artists, that’s what turned this from something they feared to something they were deeply intrigued by. They wanted to play with it and hear what the bloody thing sounds like!”
Hence Alice, which has spent the last three months learning piano chops and improving every day. That’s been buoyed by research published from other companies in the AI-music field, including projects at Google (Magenta) and Sony (Flow Machines).
“All these companies are investing heavily in it, but they’re also sharing what they’re doing. We’re seeing breakthrough papers published daily,” says Phillips.
“The barriers to entry are still pretty high, even down to just being able to interpret the math in these papers, but this really feels to me like the next big thing. We’ve been through Windows, the web, mobile and then social. Now deep-learning. It’s a seven-year cycle, and we’re really only at the start of year two.”
Phillips is full of praise for the Techstars Music program’s role in helping Popgun hone its technology and plans, although he admits that he went in as a sceptic about accelerators in general.
“I thought it would be a training and teaching thing, but no, they expect you to have your stuff together. They throw you in front of the industry and investors,” he says.
“The Warner guys were great, the Sonos and Sony guys were great. The conversations are ‘What do you want? We’re investors in you, so what do we need to do to help you win?’ That’s a change in mindset for the music industry.”
It is very early days for Popgun and Alice, but some potential products and business models are already emerging. Alice could be a creative tool for musicians which they pay for, but it could also be developed as a music-education tool.
“It’s a new, fun way to interact with music. My 10 year-old daughter is playing the piano, and it’s the bane of our existence to get her to practise! But with Alice she plays for hours: it’s a game, and you’re playing with somebody else,” says Phillips.
“It’s going to be an addictive training tool for sure, but we’re early days: a year away from commercial products in that area. But we’ve had interest already: companies that make keyboards were into it at the demo day. They already have play-along beats and drums, but they don’t feel intelligent. Imagine the first fully AI-enabled keyboard…”
For now, besides raising funding, Popgun’s focus is on continuing to help Alice learn and improve. Step one was getting the program to listen to a human and then predict (and play) what might come next. Step two will be playing with a human: “They’re going to play this so I’ll play that – accompaniment!” as Phillips puts it.
And those ambitions to have a top 40 hit with an AI composition? They may not have entirely gone away. Phillips thinks that the adults who feel defensive or derisive about such a prospect are unlikely to be the target listeners.
“My kids and the generation coming through, they’ve been brought up on iPads and Roblox and Minecraft and generative music. It’s in the games they love, being pumped into their heads, and they just get it,” he says.
“If I play my kids the Rolling Stones they don’t get it. If I play Radiohead they’re like ‘Who’s this whining bastard?!’ This next generation is probably going to adore AI music.”
“We will hate it because we don’t get it, and they will love it because we hate it! So it will become their thing. They already listen to the backing tracks from the games they play on Spotify.”
Mischievously, Phillips suggests that this may bring some opportunities to the music industry, even if they open up new fissures between labels and artists.
“Label guys love the idea that there’s a new breed of fan out there where maybe you can press a button and make a hit for them. Although we’ve got into some pretty surreal conversations about legal stuff,” he says.
“When labels are going to sign new artists, there might be clauses over who owns their ‘intelligence’. Imagine if an AI hangs out with you for a year, and another AI hangs out with Pharrell Williams for a year. The one that hangs out with Pharrell is gong to be a shed-ton better than yours!”
“So who owns that AI: the label? Once Pharrell’s trained it, maybe they don’t need Pharrell any more! I don’t think artists are going to like that, although the young artists will probably do it…”
Popgun’s aim isn’t to put Pharrell out of business: in the scenario above, it’s just as possible that the artist will own the AI, and it will become a tool to help their career rather than destroy it.
These are the kinds of conversations musicians and the music industry will need to be having – however surreal and sci-fi-worthy they feel at first – if AI composition develops at anything like the pace that AlphaGo did.
The fact that Alice has a human name may make her/it less threatening. Talking of which, is Alice a ‘her’ or an ’it’ and how does that affect how we see this technology?
“When we called it a name, people started to treat it like it was sentient! ‘Oh, Alice can do this…’ It’s about 850 lines of Python code!” laughs Phillips, before segueing back into the human-vs-AI topic.
“Some of the artists who were originally upset thought that whatever humanness is, it’s somehow defined in this ability to create music. I don’t buy this: that music is somehow only a human thing,” he says.
“I can buy that it’s an ‘intelligent’ thing, but I’m not sure we’ve got the lock on that. We might be the smartest things we’ve ever seen, but that’s not to say we won’t come across something smarter that can also make music.”
Phillips feels like Popgun is “racing” against Google’s Magenta team to understand how soon that point will come, and what it will mean for musicians, listeners and all the middlemen between them.
“It’s a wanky term, but the singularity moment in music is when this thing starts to make music that sounds like actual music. I wish I knew what’s going to happen next after that, but I don’t,” he says.
“The Turing Test for us in music is to fool musicians that they’re playing with another human. What follows that, we’re not sure, but we think that moment will come in the next year… An AI you can play with is going to become a tool for composition and learning.”