Explained: The conspiracy to make AI seem harder than it is! by Gustav Söderström

We’re back! And today we dig into Spotify’s AI DJ. Since its inception, radio provided a music format that was convenient and simple, but far from personal. Spotify’s solution? AI DJ, a bespoke experience that provides listeners with the content they want through generative AI technology. After years of experimentation, we created a product that captures the fun and usefulness of a radio DJ, but builds on it to make it a personalized experience meant just for you. From tackling the gap between “mind free” listening and “mind busy” listening, to relentless weekly user testing, the path to the AI DJ took many twists and turns before it landed in the lap of listeners. When the concept of the product was ready, we just had one more question: Who would become the first-ever “Voice of Spotify?” Follow along and hear the full story as we dive into AI DJ’s journey, from conception to launch. This episode features interviews with Ziad Sultan (Spotify’s VP of Personalization), Xavier “X” Jernigan (Spotify’s head of cultural partnerships and the voice of AI DJ), Emily Galloway (Spotify’s Head of Personalization Design).

2023 may be a year that people still speak about 100 years from now, the year computers passed the Turing test! You know what these things can do, but do you actually understand how they can do it? How is it that we have services like Chat GPT that can write entire novels, and services like Stable Diffusion and Midjourney that can create amazing images or even music from just a text description or even white noise? Straight from the halls of Spotify, this is an educational talk from an internal executive offsite that we’re sharing with the world. The premise of this talk is that AI is made to seem harder to understand than it actually is, and that after this presentation, you will feel like you understand how all of what’s now happening is possible - even if you don't work in tech and you don’t know a lot of math. 00:00:00-Intro 00:04:01-What is an LLM? 00:20:09-What about Creativity? 00:24:00-How do you steer it? 00:34:26-Why did no one see it coming? 00:39:00-Everything is a vector! 00:57:44-What is a neural network? 1:05:53-Intelligence is compression! 1:15:12-Diffusion Models - Generating Images, video and music 1:21:10-Conditioning on text Sources used to build the talk: ⁠www.mdpi.com/2076-3417/11/21/10267⁠ ⁠openai.com/blog/chatgpt?ref=assemblyai.com⁠ blog.acolyer.org/2016/04/21/the-amazing-power-of-word-vectors/ https://aclanthology.org/N13-1090.pdf ⁠www.researchgate.net/figure/Perceptron-neuron-with-three-input-variables-with-a-single-output-0-or-1-The-inputs-are_fig1_338989845⁠ www.researchgate.net/figure/Schema-of-Autoencoder-architecture_fig1_33899555 www.this-person-does-not-exist.com/en ⁠developer.nvidia.com/blog/improving-diffusion-models-as-an-alternative-to-gans-part-1/⁠ There are great resources available, for anyone interested to dig deeper

What’s Spotify up to next? Tune in to the season finale to hear about Spotify’s plans for the future of audio and the brand new formats that we’re launching today and in the near future. As always, when building something new, we don’t know how it’s all going to play out yet, or exactly what the product strategy lessons will be. But we do know which bets we’re making, and why we’re making them. This episode features interviews with Spotify insiders Michael Mignano and Nir Zicherman (Anchor founders), Howard Akumiah (Locker Room founder), and Maya Prohovnik (Head of Product for Spotify’s Creation Platform team). You can find the transcript here

This is the story of Backstage -- a “platform of platforms”, built by Spotify and open to anyone as open source software. Tune in to hear insights about when to build vs buy - and when to open source - or not! Interviews with Jeremiah Lowin, CEO of Prefect.io, a company that runs on what is called an open core model, as well as from Spotify insiders like Pia Nilsson, Director of Engineering and Tyson Singer, Head of Technology and Platforms.

How do you go “from zero to one”? In this episode, you’ll learn how Spotify managed to enter a market late: a market that was already dominated by a single player - our biggest competitor in music. This is the story of how we entered podcasts. Daniel Ek, Dawn Ostroff, and Sten Garmark shed light on the insights that led Spotify into the podcast arena. Tune in as we explore Spotify’s bet on podcasts, the shift to audio-first and four hard-won lessons in product strategy. The transcript is available here. Thanks to Dan Willis for letting us use a clip from his YouTube video  

In this episode you’ll learn about Spotify’s journey into the painful complicated world of hardware. You’ll hear how we worked our way down to the very bottom of the stack in our quest to create the “perfect listening session” in the home, how we had to build a completely new media streaming protocol in the process and how we ultimately found a way for users to pass something called: “The Beer Test.” Things you’ll learn in this episode: - How do you get a company to sustain a risky multi-year bet with high uncertainty? - How do you partner well (and not so well!) - How do you go about introducing a new standard in an industry? The episode features interviews with Thomas Cullen, Sonos co-founder, and Spotify insiders, Sten Garmark, VP User Experience, and Nicole Burrow, Senior Director of Design. Check the episode transcript here and Spotify Connect site here

This episode is especially geared towards CEOs/CTOs that have to "kill their darlings" and make hard choices. Listen to Google employee number 8 - and SVP of infrastructure at Google - Urs Hölzle, as well as Spotify insiders Tyson Singer, Nicole Bouchard and Emil Fredriksson as we journey from "brilliant" proprietary tech on a few servers in a cupboard, in an apartment, through buying thousands of servers in anticipation of our own growth, to eventually shutting down our own server infrastructure team completely, instead partnering with Google to be able to move up the stack and compete where our resources can make the most difference. Transcript here 

Will AI truly replace us humans when it comes to music? Listen to machine learning (ML) legend Andrew Ng and Spotify insiders on what it really means to develop products in an AI/ML first world, how we invented the term “Algotorial” and how reinforcement learning (RL) applies to music. Features interviews with Andrew Ng, Oskar Stål, Ajay Kalia, Meg Tarquinio, and Tony Jebara. Transcript here 

This episode tells the story of how Spotify went into negative growth and nearly didn’t survive to it’s 5th birthday as the world around us suddenly changed…. You’ll learn how we had to reinvent our entire business model and find a new free tier that worked in a mobile first world.  You’ll hear from people like world-class technology investor and analyst Mary Meeker, as well as early Spotify employees like Sten Garmark, Charlie Hellman and Babar Zafar. But most importantly, you’ll finally get to know why it’s better to be lower on a taller mountain than higher on a smaller one. Transcript here https://bit.ly/3mIV3M6

In this episode we tackle how Spotify went from being a popular free desktop app - but that nobody paid for - to being a mobile app that people actually paid for - even though users already had all their pirated music on their phones for free…  It includes interviews with: Kenneth Parks, Mattias Arrelid, Petra Hansson. Bonus - If you want to see what the magic trick actually looked like back in 2009, check out our first public demo here! https://www.youtube.com/watch?v=Z5whfaLH1-E Script here https://bit.ly/2Q24TN5