Artificial intelligence may be better at Go and Poker than humans, and it may be gunning for some humans’ jobs, but we can at least be content that DeepMind won’t be replacing the rich and famous rappers appearing on MTV’s Cribs anytime soon.
While humans may still get the exclusive right to flaunt their rap wealth on national TV, a teenager from Virginia has taught an AI to mimic the artform using 6,000 lines from Kanye West. Its flow may not be the best, equipped as it is with a stop-start synthesised robotic voice, but you can’t really argue with the rhymes.
It’s far from perfect (by my count, 11 of the 18 lines in the fourth verse end with the word “shit”, rhymed with “lit”, “fit” and “Brad Pitt” – twice), and you can still figure out which Kanye West lines the AI is parroting if you examine them close enough, but it flows pretty well all the same. It’s more impressive when you realise that the first build of the AI was created in just a week to prove a point to Barat’s programming club. “Originally it just rearranged existing rap lyrics, but now it can actually write word by word,” Barat told Quartz.
Perhaps more importantly, this project – on a micro scale – shows the difficulty of the black-box-learning method, where everything about how the AI is learning is invisible to its trainers. At first, Barat’s data wasn’t correctly formatted, meaning that the algorithm was creating nonsense, giving him this epiphany: “We’ve got to figure out the way to make more transparent machine-learning models. If we’re able to make progress into seeing into the thought process of a neural network, I think it’s going to teach us a lot about Go, or language, or the structures that we aren’t able to pick up on yet.”
Yep, that’s one of the handful of issues we really do need to address with AI as soon as possible. Still, the very fact that the open-source AI community is so far evolved that a neural network can be trained to rap by a newcomer in just a week should give us some hope, at the very least.
Image: U2soul, used under Creative Commons
Disclaimer: Some pages on this site may include an affiliate link. This does not effect our editorial in any way.
