Thoughtful Things launches Engram, an offline AI sampler built around hallucinated sound

Engram turns AI misfires into instruments

According to The Verge, music startup Thoughtful Things has launched a Kickstarter campaign for Engram, its first hardware instrument. Engram is a sampler and groovebox that uses an on-board AI model to reshape incoming audio — and, at times, to hallucinate entirely new sounds.

The device is deliberately not a song generator. As The Verge notes, it is not aimed at producing polished, radio-ready tracks from a single command. Instead, Thoughtful Things positions it as a tool for experimental sound design, one that produces uncanny, glitchy textures by pushing a small audio model past its intended limits.

Tiny AI, no internet

Engram does not connect to the internet. It runs what the company calls a "tiny AI" model locally on the device. According to the Kickstarter listing, the model was designed and trained in-house, and the company makes an unusually explicit claim about its data: the models are trained only on open datasets containing audio licensed for commercial use, such as CC-BY material, and Thoughtful Things says it will never train on non-commercial, pirated or otherwise stolen audio.

The company also plans to open up the device's firmware, so users can tweak its behaviour or load their own custom models. That places Engram closer to the hackable end of music hardware than to sealed consumer electronics.

A field recorder for latent space

In the campaign video, founder Evan King demonstrates the concept by asking the device, using his voice, for "piano." Engram responds with something glitchy and only vaguely piano-like — which appears to be the point. King describes the instrument as a "field recorder for latent space," framing it as a way of exploring the interior of an AI model the way a field recorder explores the physical world.

The design also borrows from circuit bending, the long-standing practice of modifying electronic instruments so they fail in interesting ways. Engram applies that ethos to software: users are invited to tweak and even break its small models in pursuit of compelling errors.

Pricing and availability

The Kickstarter offers a limited initial run starting at $675. Thoughtful Things has not announced a retail price, but The Verge points out that the $675 pledge tier is listed as a 30 percent discount, which puts the likely street price somewhere between $850 and $900.

Why it matters

Engram represents a contrarian take on AI music tools. Where mainstream generative audio services chase clean, usable output, Thoughtful Things is betting that a model's failures are more musically interesting than its successes — and that musicians would rather hunt for strange sounds than accept a finished track.

Three details make the device worth watching beyond the niche synth crowd. The model runs entirely offline on dedicated hardware, avoiding dependence on a cloud service. Its training data is, by the company's account, exclusively permissively licensed audio — a direct answer to the data-provenance concerns that hang over much of the generative AI industry. And the promised open firmware would let owners replace the model entirely, treating the AI as a swappable component rather than a locked black box.

None of this guarantees Engram will find an audience. Experimental instruments occupy a small market, and crowdfunding hardware carries familiar risks. But as one of the first standalone instruments built around generative audio models, it sketches a template — local, transparent, hackable, and cleanly licensed — that larger AI audio efforts have largely avoided. Its reception may indicate whether "AI instrument" comes to mean a song generator in a box, or something musicians actually want to play.