Georg Kucsko, Co-Founder & CTO at Suno скачать в хорошем качестве

Georg Kucsko, Co-Founder & CTO at Suno

Welcome Georg Kucsko, Co-Founder & CTO at Suno on this episode of 🔥The Lantern 🔥 Suno originally began with the release of an open source, text to speech model, that was "open ended" called Bark. People started to experiment and even make music with the model. Suno, which is a Punjabi word that means “to listen”, was born! ⚒️ What were some of the technical breakthroughs that Georg has been working on with his team to reach this level of scale? ⚒️ At its heart, Suno is a research company. Audio is an interesting space because it lies in between the two “major” domains - text and image. Image has large diffusion flow type models, text has the even larger sequential GPT style models. Audio is somewhere in between. It’s a space where you borrow from both of those fields figuring out ways to distill flow, scale context window sizes, and have a good attention pattern competing with the largest research labs in the world. Preference data and preference tuning is a huge part of where they spend their time - collecting data from the most amazing creatives in the music industry - to figure out what model outputs are what they’re looking for to better deliver the type of creative output they want. 🔍 Where is Music AI headed? What’s Georg excited about? 🔎 Suno’s approach has been to build one single foundation frontier model that powers the entire platform. 🎶 First? Make decent music, piece by piece. 🎮 Next? Lean into control. Figuring out what musicians are really trying to achieve is crucial and then helping them iterate on it without needing to start from scratch. 🤔 What should machine learning and technical builders know about Suno? 🤔 Suno is mostly about the people. For music, like the word Suno, it’s about listening - to music, understanding their users, teammates, etc. It helps to work on a product that you use & love yourself. Everyone at Suno loves music, playing instruments, and using Suno to combine those efforts. Music AI is a hard space and takes a bit of everything you have to build large scale frontier models that can understand, create, and scale with the users. Taking a large preference dataset and then translating it into a product that serves consumers at scale is one of the biggest challenges.