Thoughtful Things just launched a Kickstarter for Engram, a sampler and groovebox built to glitch audio into something new, not spit out a finished song.
Engram takes in live or sampled audio and runs it through a custom, in-house-trained AI model that distorts, warps, and occasionally hallucinates sounds that were never in the original signal. Unlike cloud tools such as Suno, it has no internet connection at all. The model runs entirely on-device, described by the company as "tiny AI," and the Kickstarter listing frames the whole project around pushing audio models past their limits rather than smoothing them into something radio-ready.
That's the interesting bet here. Most AI music products chase the push-button-song promise, competing on how convincingly they can fake a finished track. Engram instead treats the AI's failure mode, the hallucination, as the instrument itself, closer in spirit to circuit-bent gear and glitch art than to a hit-generating machine. Running locally also sidesteps the usual data and latency questions that come with cloud-based music AI.
It's still a Kickstarter, which means the only thing shipped so far is a pitch. Custom-trained local models are notoriously hard to make reliable outside a demo reel, and "embrace the weirdness" is also a convenient way to describe an AI that doesn't always do what you ask.