Header image source: Engram is a sampler that turns broken AI hallucinations into music | The Verge via The Verge via Google — cropped to 16:9 and colour-adjusted.
Key takeaways
- Engram exposes AI model quirks for real-time manipulation
- Turns AI hallucinations into creative musical textures
- Boutique hardware for experimental musicians, not mainstream producers
Engram launched on Kickstarter. A hardware sampler. $675 to start. No presets. No polished loops. Just AI models you’re meant to break. Thoughtful Things, the startup behind it, isn’t chasing radio-ready tracks. They’re selling glitches, hallucinations, the kind of unpredictable noise most AI music tools treat like a bug.
That’s the counter-narrative here. While companies like Suno and Udio chase mainstream music creation, Engram leans into chaos. Its embedded MusicGen inference code isn’t just running in real-time—it’s designed to be bent, twisted, pushed into failure modes. Three new audio generation concepts sit under the hood, but the real innovation isn’t the tech. It’s the philosophy: AI’s hallucinations aren’t mistakes. They’re the whole point.
How Engram Works: The Tech Behind Model Bending
Engram isn’t another plugin. It’s a standalone hardware box with one job: turning AI failures into sound. The core is a hand-rolled implementation of MusicGen’s inference code—no black-box API calls, no cloud dependencies. The creator built it from scratch for tweakability and performance. That’s deliberate. Most AI music tools abstract away the model, smoothing out rough edges before the user ever hears them. Engram does the opposite. It exposes the model’s quirks, lets you manipulate them in real-time, and calls the result "model bending. "
What does model bending look like? The brief doesn’t spell out the mechanics, but we can guess. Traditional samplers let you stretch, pitch-shift, loop. Engram probably adds layers: feeding the model conflicting prompts mid-generation, tweaking temperature to force instability, corrupting its latent space on the fly. The result? Sounds that don’t just deviate from the input—they disintegrate into something new.
Three "new audio generation concepts" are mentioned. No specifics, but possibilities emerge. Real-time latent space exploration—navigating the model’s internal representations like a synth knob twisting a waveform. Prompt chaining, where output becomes input, creating cascading feedback loops. Or maybe a slider between stable and unstable states, letting performers dial in breakdowns.
This isn’t a sampler with AI bolted on. It’s a performance instrument designed for failure. That’s the radical bit.
AI Hallucinations as Creative Fuel
Engram treats AI’s flaws as features. Most tools bury hallucinations under post-processing. Suno and Udio generate vocals that sound eerily human because they’ve trained their models to avoid the uncanny valley. Engram flips that script. Its hallucinations aren’t a problem. They’re the main attraction.
This isn’t new in experimental music. Circuit bending, modular synths, glitch art—all thrive on unpredictability. But Engram’s twist is the source. Circuit bending requires hardware hacking. Modular synths demand patch cables. Engram’s chaos comes from an AI model’s latent space—an abstract, high-dimensional playground most tools keep locked away.
The implications for sound designers are huge. Glitch artists, noise producers, avant-garde composers have always sought new ways to break audio. Now, instead of mangling samples with effects, they can mangle the model itself. Feed audio into Engram, manipulate it until sounds dissolve into pure texture. Prompt it with conflicting instructions and let the model’s confusion generate something entirely new.
There’s a ceiling, though. Engram’s audience isn’t pop producers or film composers. It’s artists who want their tools to misbehave. For them, the lack of polish isn’t a flaw. It’s the appeal.
The Broader AI Music Landscape: Polished vs. Unpredictable
Engram’s Kickstarter is a tiny blip in AI music. Companies like Suno and Udio chase mainstream music creation with tools that mimic human musicianship. Their approach focuses on generating complete songs from prompts. Engram’s approach focuses on generating unpredictable audio that users can shape.
This divide isn’t just about features. It’s about what AI music should be. Mainstream tools are built for accessibility. Productivity software, turning vague ideas into polished tracks with minimal effort. Engram is built for exploration. A sketchpad, not a studio. One path leads to AI replacing musicians. The other leads to AI collaborating with them—by giving them new ways to fail.
Historical parallels aren’t hard to find. The rise of synthesizers in the 1960s met skepticism from purists who saw them as cheating. Now, synths are everywhere, but the experimental ethos lives on in modular systems and DIY electronics. Engram feels like a spiritual successor. Not about replacing musicians. About giving them a new way to play with sound.
Performance and Usability: The Trade-Offs of Hand-Rolled Inference
Engram’s custom inference code isn’t just a technical detail. It’s a statement. Most AI music tools rely on cloud APIs or pre-trained models, trading control for convenience. Thoughtful Things went the opposite route. They hand-rolled MusicGen’s inference, embedding it directly on the hardware. That’s risky.
Upside: Engram is self-contained. No internet. No cloud latency. No third-party API logging prompts. More importantly, it gives users direct access to the model’s internals. If Engram lets you tweak temperature, top-k sampling, or latent space dimensions, those controls exist because the model isn’t abstracted away.
Downside: scalability. Hand-rolled inference is fragile. Harder to update. Prone to bugs. Limits model complexity. MusicGen is lightweight compared to Stable Audio or Udio’s backend. A trade-off Thoughtful Things accepted to keep Engram tweakable.
The question is whether users will care. For experimental musicians, it’s worth it. They don’t want a tool that does everything. They want one that does one thing well—and lets them break it in interesting ways.
Who’s Engram For? The Audience for AI Glitch Art
Engram’s $675 price tag puts it in boutique hardware territory. Not mass-market. For a specific kind of artist: those who see AI not as a replacement for creativity, but as a new medium.
Glitch artists will love it. Noise producers will find new textures. Sound designers might use it to generate eerie, unpredictable ambiences. Mainstream musicians? Probably not. Engram’s lack of polish is a feature, but one that will frustrate anyone looking for a tool that "just works. "
Early adopters will come from niche communities. Modular synth forums. Experimental music Discords. Artists who already circuit-bend their gear. These are people who want their tools to be unpredictable. For them, Engram isn’t just another sampler. It’s a new way to interact with AI.
The Future of AI Music: Beyond Perfection
Engram raises a provocative question: What if the most interesting AI music tools aren’t the ones that sound most human?
Right now, AI music is obsessed with polish. Suno’s tracks sound like they could be on Spotify. Udio’s demos are indistinguishable from human-made pop. Impressive, but limiting. It reinforces the idea that AI’s role is to replicate, not innovate.
Engram rejects that. It’s not trying to sound like anything else. It’s trying to sound like itself—a machine failing in real-time, with a human guiding the collapse. A radical departure from the narrative of AI replacing human creativity. It positions AI as a collaborator, not a competitor.
Could this influence larger tools? Maybe. Imagine Suno adding a "glitch mode" that corrupts outputs. Udio letting users tweak temperature mid-generation. Engram’s philosophy could trickle up—but only if the mainstream decides unpredictability has value.
The risk is that Engram stays niche. Its appeal is limited by design. Not for everyone. But that might be the point. The most exciting tools aren’t the ones that try to please everyone. They’re the ones that give a small group something they didn’t know they needed.
Why Engram Matters
Engram isn’t just a sampler. It’s a provocation.
In a landscape dominated by AI tools chasing perfection, Engram embraces failure. In a world where every startup wants to sound like a hit factory, Engram sounds like nothing else. A reminder that AI’s role in music isn’t just to replicate—but to expand.
The question isn’t whether Engram will succeed. It’s whether its philosophy will. Can AI music tools thrive by embracing chaos? Or will the industry keep chasing the illusion of polish?
Either way, Engram’s Kickstarter is a signal. The most interesting AI tools might not be the ones that work perfectly. They might be the ones that break in just the right way.
Leave a Reply