Almost every AI music product is a web service with a text box. You describe a sound, something happens on a server, and audio comes back. The interaction model is a search engine, and the thing you get is a result rather than an instrument.
The Engram, from Thoughtful Things — developer Evan King — is the other shape of that idea. Small neural audio models built directly into the hardware. No internet connection required.
What it is
A sampler groovebox, currently a prototype crowdfunding on Kickstarter:
- 8 knobs — tempo, gain, pan and so on
- 16 programmable beat buttons
- 8 transport/mode buttons
- Small LCD
- MIDI in/out, CV sync/gate, stereo line input, headphone out, onboard microphone
- Voice control — no text prompts
- No internet connection required
$770, early-bird tier sold out, expected delivery May 2027.
”Audio model bending”
King’s term for the working method is audio model bending: pushing neural models beyond their typical limits.
That is a deliberate and rather good reference. Circuit bending is the practice of short-circuiting a cheap electronic instrument to make it behave in ways its designers never intended — a Speak & Spell rewired until it screams. The sounds are not the device working; they are the device failing in a repeatable, playable way.
Applying that framing to neural audio models is more than a marketing analogy, because generative models have a great deal of interesting behaviour outside their intended operating range. Push a sample rate, overdrive a conditioning signal, feed a model’s output back into its input, run it on material unlike anything it was trained on, and you get artefacts that are consistent, characterful, and nothing like the model’s advertised purpose.
That is a legitimate and under-explored instrument design space. Commercial AI audio tools work hard to suppress exactly these behaviours, because for them an artefact is a defect. An instrument builder can treat them as the sound.
The rest of the workflow follows the same lineage: you sample external audio or resample AI-generated sounds, then manipulate them with slice and flip techniques King explicitly connects to musique concrète — Schaeffer and Henry cutting and reversing tape in the late 1940s, which remains the foundation of every sampler that followed.
The two design choices that matter most
On-device, no internet. This is the single most consequential decision in the spec, and it is worth being clear about why.
A cloud instrument is not an instrument. It has latency you cannot control, it stops working on a bad hotel Wi-Fi connection, it depends on a company continuing to pay for inference, and it means your material leaves the room. For anything you intend to perform with, or own in ten years, local inference is the difference between a tool and a subscription. The models are small because they have to be, and the smallness is a feature.
Voice control rather than text prompts. Interesting and risky. The upside is obvious: there is no keyboard on a groovebox, and typing is the wrong interaction for a device you play with both hands. The risk is equally obvious — voice recognition on a stage, in a studio with monitors up, or in any room with other people in it, is a famously poor interaction. Whether this is clever or a demo-driven mistake depends entirely on how narrow the vocabulary is and whether there is a non-voice path to everything.
The licensing claim, and why it is worth noting
King states the models were trained only on commercially licensed audio, with no copyright infringement.
This is a claim rather than an audit, and it should be read as one. But it is a claim most of this industry declines to make at all, and the people shipping the largest generative audio models have been conspicuously vague about training data. A small builder stating a provenance position plainly is a meaningful signal about who they expect their customers to be — people who intend to release the music they make, and who would rather not find out later that the instrument was the liability.
The part to be sceptical about
It is a prototype on Kickstarter with delivery expected in May 2027 — nineteen months out. That is a long runway for a first hardware product with custom silicon-adjacent requirements, and hardware crowdfunding has an unkind history, particularly for instruments with ambitious software. The feature that worries me most is pluggable models created by users or the community, promised as a future update. A model format, a conversion toolchain, and a community around it is a platform commitment, and platform commitments are where small hardware companies go to die.
None of which means don’t back it. It means back it at a price you are willing to treat as a bet, which is the correct posture for all crowdfunded hardware and especially for the interesting kind.
The idea, though, is right. An instrument with a neural model inside it that you can abuse, running locally, playable with your hands, is a much better answer to “AI music” than another website.