softmax

Sacha Krstulovic breaks down the fundamental realities of AI in music that cut through the hype and fear dominating current discourse. As someone who’s spent decades building AI systems for audio—from speech recognition to environmental sound detection—he offers a grounded framework for understanding what AI actually does versus what we imagine it does.The conversation explores why pressing a button to generate a complete song misses the point of creative tools, how data ownership remains the unresolved ethical crisis of the AI era, and why human agency at both input and output stages determines whether we’re witnessing automation or artistry. Sacha shares insights from his work building AI research teams at Music Tribe, where they discovered the real use cases musicians want: evading the blank page, compensating for missing skills, and gaining time—not replacement.Particularly compelling is his framework for thinking about AI as automation, complexity, data-driven programming, and always a function with inputs and outputs. This perspective helps practitioners navigate the difference between assistive mixing tools that teach you about conventions while giving you power to break them, versus generative systems that claim to “make music” while obscuring the human curation required at every step.For anyone building tools for creative-tech professionals or working at the intersection of machine learning and music, this conversation offers rare perspective from someone who’s seen the evolution from unit selection speech synthesis to transformer-based generation—and maintains healthy skepticism about what actually serves human creativity.Episode Chapters[(2:25) Sacha’s Journey: From Speech Recognition to Audio AI Leadership(13:21) Demystifying AI: Four Core Principles(23:32) Beyond Generation: The Full Landscape of Audio AI(28:30) Real Use Cases: What Musicians Actually Want from AI(33:38) The “Press Button, Get Song” Problem(40:44) Breaking the Machine: Creative Exploration with AI(51:04) Data Ethics and the Copyright Crisis(57:49) Digital Hangover and the Return to Real Life Experience(1:03:30) Closing: Finding Sacha and Understanding AIPractical TakeawaysFramework for Understanding AI:* AI is automation with extreme complexity (billions of parameters)* It’s data-driven programming, not hand-coded rules* Always has inputs and outputs—it’s a function, not an entity* Mimics patterns without consciousness or independent agencyDesign Principles for Music AI Tools:* Present outputs as editable parameters, not black boxes* Let AI act as teacher showing conventions you can consciously break* Focus on use cases: blank page stimulation, missing skills, time efficiency* Preserve human agency at input (what to explore) and output (what’s good enough)Data Ethics Standards:* Traditional ML practice: own or license all training data* Current lawsuits challenge the “scrape everything” approach* Ed Newton-Rex’s Fairly Trained advocacy as alternative modelThe Live Music Economy:* Musicians increasingly earn through concerts, not recordings* Local, human interaction offers what algorithms can’t deliver* Fandom culture, tangible experiences, and vinyl collecting as counterweights to digitalResources & LinksConnect with Sacha:* LinkedIn: https://www.linkedin.com/in/sacha-krstulovic-3505544/* Personal website: sacha.today (includes essay on creativity and entrepreneurship)* Consultancy: understand-ai.todayMentioned:* AES International Conference on AI and Machine Learning for Audio* Ed Newton-Rex - Fairly Trained advocacy* Dadabots* Max Cooper - Electronic music artist* Audio Analytic - Environmental sound recognition (acquired by Meta)* Music Tribe - Audio equipment manufacturer (Behringer, Midas, TC Electronics)* Giada Pistilli - Should we be afraid of becoming attached to machines?* Documentary - Re-learning to listen to musicGuest BioSacha Krstulovic is an AI researcher who spent two decades at the intersection of machine learning and audio, from early speech recognition work accounting for vocal tract physics to building the first large-scale environmental sound recognition system. As Director of AI Research at Music Tribe, he led a team of 15 exploring applications for audio equipment manufacturers. His career spans academia (PhD in speech recognition), industry giants (Toshiba, Nuance), successful startups (Audio Analytic, acquired by Meta), and now independent consultancy helping companies structure practical AI applications. He brings rare perspective on the evolution from “machine learning” to “AI” terminology—and maintains focus on what actually serves human creativity versus what captures attention.Related This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit redito.substack.com

Guest: Jordan Davis, solo founder of TensorPunk with 15 years of music industry experience spanning performance, recording, and sync licensing. Musician-technologist building AI tools from a producer's perspective.Topics: AI music tool development, Mace generative sampler, Anvil model training platform, DAW integration, dataset curation, performance-focused AI instruments, hacker ethos in music techAboutJordan Davis represents the emerging wave of musician-technologists who are building AI tools from the perspective of actual producers rather than data scientists. His company Tensor Punk has created Mace, a generative sampler that runs natively in DAWs, and Anvil, a tool for training custom models. This conversation reveals hard-won insights about product development, model training challenges, and the philosophical tensions shaping AI music tools.Key InsightsBuilding for Real Workflows: Jordan emphasizes that most AI music tools fail because they live in web browsers, disconnected from producers' actual creative environments. Mace runs as a VST plugin directly in DAWs, allowing seamless integration into existing workflows. This seemingly simple decision reflects deeper understanding of how musicians actually work.The Art of Model Training: Jordan shares fascinating technical details about his journey from TensorFlow to PyTorch, discovering that high-frequency content like hi-hats trains faster and produces better quality than low-end samples. He deliberately chose "grittier, grainier" results over pristine fidelity to create genuinely experimental sounds rather than polished imitations.Accessibility Without Compromise: TensorPunk tools run on CPU when GPU isn't available, supporting everything from Windows 7 to modern systems. Jordan developed much of the initial code on an old Windows 7 machine, demonstrating commitment to democratizing access to AI music tools.The Performance Future: Rather than focusing on one-click generation, Jordan envisions AI tools as performance instruments. He describes frustration with traditional DJ equipment and imagines knobs controlling "live neural samplers" that morph genres in real-time, freeing performers to explore "radically different sonic territory."Notable PerspectivesOn AI Music's Identity Crisis: "There's going to be really cool and interesting instruments that haven't even been invented yet... I see the future of there being really, really cool and interesting instruments."On the Hacker Ethos: "The branding behind Tensor Punk is that maybe there can be a company that has that hacker ethos and spirit, and is there supporting the creative and the artists. It's creative first, hacker and punk mentality first."On Controllability: "A lot of producers that might be hesitant with AI tools... want more control of the outcome and less black box results. I think there's a lot right now that just needs to be re-engineered and can have amazing results."Practical Takeaways* Dataset Quality Over Quantity: Focus on curated, labeled datasets using producer vocabulary rather than generic metadata* Platform Integration: Build tools that live inside existing creative workflows rather than isolated web applications* Community-Driven Development: Engage with actual users to discover unexpected use cases and workflows* Performance-First Design: Consider AI tools as instruments for live manipulation rather than just generation enginesFollow Jordan and TensorpunkTensorpunk DiscordTensorpunk IGTensorpunk WebsiteMusicMemory States - ȶʀǟɨռɨռɢ ɖǟȶǟ [12] - mini_epRelated This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit redito.substack.com

Guests: CJ Carr and Zack Zukowski, Dadabots founders, AI music researchers since 2012, known for intense neural metal streams and frontier AI music researchKey Topics: Live AI performance, prompt jockeying techniques, real-time music generation, cultural implications of AI training data, future of music interfacesMust-listen conversation with Dadabots founders CJ Carr and Zack Zukowski, who've evolved from their legendary 24/7 neural metal streams into live "prompt jockeying" - performing with tracks that don't exist until they're generated on stage. These creative technologists break down their journey from 2012 SoundCloud remix bots to performing at the UN, revealing the virtuosity behind real-time AI music generation that pushes far beyond typical button-pushing stereotypes.Key SegmentsThe Art of Live AI PerformanceCJ and Zack unpack "prompt jockeying" - their term for DJing with AI-generated tracks created in real-time. Using Jupyter notebooks as their interface, they generate tracks in seconds while performing, creating everything from "twinkle trap" (children's music fused with trap beats) to "ghost pepper salsa" (Afro-Cuban percussion meets New England hardcore). The process requires riding "the edge of chaos" - mixing tracks they've never heard while reading crowd energy and venue acoustics.Technical Innovation Beyond the HypeThe duo explains how diffusion models now generate full 3-4 minute tracks in seconds (versus the hour-per-minute of early sample RNN), enabling live performance. Their approach includes multi-step generation processes, where tracks are fed back through models for layered complexity, and context-aware prompt chaining that creates coherent musical narratives throughout sets.Research Meets PerformanceDadabots exemplifies the creative technologist archetype - building the very AI tools they perform with while maintaining roots in metal and punk virtuosity. Their 10+ page process documents for the AI Song Contest reveal the technical depth behind seemingly simple outputs, from custom RVC voice models to neural synthesis parameters that achieve timbres "halfway between guitars and synths."Cultural Implications of AI MusicFascinating exploration of how AI music reflects "the collective consciousness through a hall of mirrors" - training data becomes warped reflections of human culture. Their UN performance, taking genre requests from international diplomats, demonstrates how AI models enable unprecedented cultural fusion while raising questions about attribution and musical authenticity.Practical Takeaways* Live AI Performance Setup: Combine DJ software (DJ Pro with Neural Mix) with Jupyter notebooks running diffusion models locally* Prompt Engineering: Use BPM constraints for seamless mixing; develop signature techniques like multi-step generation* Venue Adaptation: AI generation enables real-time genre switching based on space, sound system, and audience energy* Technical Approach: Prioritize local models for performance reliability; build custom tools rather than relying on consumer appsNotable Insights"We're music hackers in a band slash hackathon team slash research lab... Code is this other frontier of music.""It's like DJing, but harder because the tracks don't exist. And so we're going back to back on beats, but we don't know where the beat goes.""The history of music is influences upon influences upon influences... it is a hall of mirrors."The Bigger PictureThis conversation captures a pivotal moment in creative technology - the shift from "using AI tools" to "performing with AI" as a creative partner. Dadabots represents the vanguard of human-AI collaboration, where technical skill becomes performative art and code itself becomes an instrument. Their approach signals what's coming for all creative fields: real-time, context-aware AI systems that enable entirely new forms of artistic expression.For creative technologists across disciplines, this offers a template for authentic AI integration - not replacing human creativity but amplifying it through deep technical understanding and performative mastery. The implications extend far beyond music: imagine live coding visual art, real-time AI-assisted writing, or dynamic design generation that responds to environmental context.Most significantly, Dadabots demonstrates how the next generation of creative professionals will need to be fluent in both artistic tradition and AI systems architecture - true creative technologists who can build their own tools rather than simply consume them.MentionedMusic and Artificial Intelligence: Artistic TrendsAlgoravehttps://tidalcycles.org/Reductive, Exclusionary, Normalising: The Limits of Generative AI MusicMusic for ComputersPrompt MeAI Song ContestLil Internet Mix (NTS)https://everynoise.com/Magenta RTPrompt Jockey DocumentaryFollow Dadabotshttps://dadabots.com/https://www.youtube.com/@dadabots_https://x.com/dadabotsPrompt Jockey DocumentaryMusicR. Tyler - raise ValueError()Related This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit redito.substack.com

Guest: Ezra, Founder of AudioCipher TechnologiesTopics: AI Music Workflows, Creative Tech Stacks, Film Scoring, Industry EthicsAboutMusicians wrestling with AI's place in their creative process will find practical frameworks here. Ezra breaks down how serious producers are building "tech stacks" with Suno, Udio, and ElevenLabs—moving beyond one-click generation toward intentional creative workflows. This isn't about replacing musicians; it's about understanding how AI tools fit into existing production pipelines.Key Insights & Practical TakeawaysThe Creative Tech Stack Approach* Suno excels at style transfer: Upload your rough ideas, transform them across genres with sophisticated prompt engineering* Udio dominates extension work: Superior at maintaining musical timbre and creating film score-style dynamic compositions* ElevenLabs offers licensed alternative: Text-to-music with commercial licensing, though less control than competitorsWorkflow Innovation in PracticeSophisticated production techniques: generating ambient loops in Udio, then conditioning Suno with unrelated prompts to create unexpected hybrid outputs. This "cross-pollination" approach demonstrates how creative professionals are developing personal AI methodologies beyond basic prompting.The Licensing Landscape Reality CheckThe ethical debate is evolving faster than anticipated. While UMG's lawsuits against Suno/Udio created initial friction, ElevenLabs' partnerships with Cobalt and Merlin are establishing new industry standards. The question shifts from "is this ethical?" to "how much are artists actually earning from these licensing deals?"Film Scoring: The Creative FrontierAI music's future lies in multimedia storytelling, not pop music generation. Film scoring offers creative freedom across multiple styles within single projects—exactly what AI tools enable. Ezra demonstrates how composers can leverage AI's stylistic range while maintaining the deep musical literacy that drives meaningful narrative scoring.Notable PerspectivesOn AI Music in Public Spaces: "I've been in dozens of cafes running 12-hour Suno playlists. It's functional music with no care given to sonic branding—this is where AI music risks poisoning public spaces."On Creative Authenticity: "There's a silent minority of serious producers actually experimenting with these tools, not just clicking 'generate' but investing time to understand where AI fits their artistic mission."On The Matrix Score Easter Egg: The film's composer embedded 12-tone serial music matrices within the soundtrack—the music literally reflects the movie's mathematical reality concepts, demonstrating the depth possible in multimedia storytelling.The Bigger PictureThis conversation cuts through typical AI music discourse by focusing on practical application rather than existential concerns. For creative technologists, the value lies in understanding these tools as part of broader production ecosystems—not replacement technologies, but specialized instruments requiring the same musical literacy and taste-making skills that have always defined great music production.Essential listening for anyone building creative workflows with AI tools, particularly those interested in multimedia storytelling and film scoring applications.Mentioned* Futureproof episode with Ezra* AI Song Contest * Music for Computers* Udio Prompt Guide* Tools: Stable Audio Open, Suno, Udio, Eleven Music, Veo3, Midjourney* the AI-Generated Ripoff Songs That Got Udio and Suno Sued* N. Korea’s AI Nukes* Twelve Tone Matrix* AudioCipher’s Ad* Feltzine Instagram* Music and Artificial Intelligence: Artistic Trends (Paper)Follow Ezra Sandzer-BellLinkedIn: https://www.linkedin.com/in/ezra-sandzer-bell/AudioCipher: https://www.audiocipher.com/MusicR. Tyler - raise ValueError() This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit redito.substack.com

In this episode of softmax, I invited my good friends for a roundtable discussion. Joining me are Shamanic, a multidisciplinary creative technologist; Levi, a musician and founder of on-chain co-op Amerge; and Alex, a community organizer, spiritual explorer, and founder of Mothership, a creative accelerator. We dive into some fascinating topics that have been on our minds lately, starting with Shamanic's insights into the burgeoning world of VR raves and virtual music culture. It's amazing to see how these digital spaces are creating new opportunities for artists and fostering a sense of global community.We then explore the intersection of AI and creativity, sparked by Levi's mention of Brian Eno's generative documentary. This led us to discuss the potential and challenges of AI in music, including the ethical principles proposed by AI4music.info. Alex shared an intriguing concept for a new type of social network powered by personal AI models, which got us all thinking about data sovereignty and the future of online interactions.Throughout our conversation, we touched on themes of human creativity, technological advancement, and the importance of maintaining our humanity in this rapidly evolving digital landscape. As always, I'm grateful for these opportunities to chat with such insightful friends and explore these complex topics together. This episode is a great example of the kind of Discord conversations we have - deep, wide-ranging, and full of parallel tangents. I hope you enjoy listening to our discussion as much as we enjoyed having it!Links:VR Chat: https://hello.vrchat.com/VRC Party hub: https://twitter.com/vrchatpartyhubVRC Rave Schedules: https://vrc.tl/Brian Eno Generative Documentary: https://www.theverge.com/24197153/eno-documentary-generative-film-gary-hustwit-interviewPrinciples for Music Creation in AI: https://aiformusic.info/GuestsShamanic:https://twitter.com/shamanic_artshttps://objkt.com/@shamanicartsLevi Downey:https://www.instagram.com/levi.downey/https://x.com/levidowney?lang=enhttps://www.amerge.xyz/Alex Park:https://www.instagram.com/alex.park/?hl=enhttps://www.mothership.wtf/softmax:Join our discord! https://discord.gg/3zTywUX9BYMark Redito:https://x.com/markreditohttps://www.instagram.com/markredito/MusicR.Tyler - raise ValueErrorPurchase on Bandcamp: Podcast ArtworkShamanic: https://objkt.com/tokens/KT1MoBFWosexpEkoCuXimxzrcffGANr7Twi7/1 This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit redito.substack.com