// AI NEWS · AI HARDWARE

Engram Is an Offline AI Sampler That Turns Model "Hallucinations" Into Music

On September 27, 2026, Los Angeles startup Thoughtful Things launched a Kickstarter for Engram — a pad-based sampler and groovebox that runs small generative audio models directly on the hardware, with no internet or subscription. Its trick is "model bending": deliberately glitching the model to sample the weird, uncanny sounds that fall out. Early-backer units start at $675.

2026-09-27 · by Moe Ameen

What happened

On September 27, 2026, Thoughtful Things — a Los Angeles startup founded by Evan King, an applied scientist who works on audio and inference at Cerebras — launched a Kickstarter campaign for its first instrument, Engram. It is a pad-based sampler and groovebox in the vein of a Teenage Engineering or Chompi box, but with one distinguishing idea: it embeds small generative audio models directly in the hardware and runs them fully offline, with no internet connection and no subscription. The company frames it as the first musical hardware to put generative audio models in the box itself rather than calling out to a cloud service.

Instead of pursuing clean, "good" AI music the way a text-to-song service does, Engram leans into the opposite. Its headline feature is what King calls "model bending" — described as circuit bending, but for neural networks. Each parameter of the on-board audio model is mapped to a knob, and turning them the wrong way deliberately breaks the model's audio decoder, producing glitchy, degraded, uncanny textures you can then sample, slice, sequence, and play like any other sound. King has described the device as "a field recorder for latent space": a way to go digging for sounds that don't exist anywhere until the model malfunctions into them.

Around that core, Engram is a fairly conventional sampler: it can record and warp your own audio through an on-board microphone or a stereo line input, generate and glitch short bursts of new audio, and arrange the results with a step sequencer, sample slicing, and on-board FX. Connectivity includes a headphone output, MIDI in/out, and CV sync and gate for hooking it into a modular or hardware rig. Thoughtful Things has also talked about pluggable, shareable audio models — closer to how vintage ROMplers swapped sound ROMs than to a fixed feature set.

Pricing and specifics are still early. Kickstarter units start at $675, described as roughly a 30 percent early-backer discount; the company has not set a final retail price, though The Verge estimated it would likely land somewhere around $850–$900. Engram is still described as a research prototype whose design may change, so treat any single spec as provisional and verify against the campaign before committing money.

Why it matters for creators

  • On-device generative audio is arriving. Engram running small audio models offline, with no cloud round-trip or subscription, is a signal that generative sound is moving from web apps to instruments you own outright — the same shift already happening with local image and language models.
  • The output is original and royalty-free by construction. Samples you bend and record on Engram are sounds you generated, not licensed library music — which sidesteps the copyright-claim and content-ID headaches that come with using commercial tracks under short-form video.
  • "Breakable" is the point, not a bug. Most AI music tools sand off the weirdness; Engram sells it. For creators, a distinctive, slightly-broken sonic texture is exactly the kind of thing that makes a short recognizable in a feed of stock-audio sameness.
  • It is a hardware instrument, not a content pipeline. Engram makes sound; it does not caption, edit video, or publish anything. The value for a creator only shows up once that audio becomes the bed under video that actually ships across platforms.
  • Niche and pre-release, so temper expectations. It is a Kickstarter research prototype aimed at experimental musicians and sound designers, not a mass-market beat maker — useful to watch, but not something to build a workflow on until it ships.

How to act on this with Kompozy

The practical move here isn't "buy a $675 sampler." It's to notice what Engram is actually good at — producing original, one-of-a-kind audio you own — and to have somewhere for that audio to go. A bent loop or a glitchy one-shot is only worth something to a creator once it's the sound under a video that reaches an audience, and that distribution step is where most people stall. That's the job [Kompozy](/) does. It's a content generation and publishing engine: from one source it produces [Persona Shorts](/glossary/persona-shorts) talking-head video, [Clipped Shorts](/glossary/clipped-short) cut from long-form footage, Listicle Video, carousels, images, blog posts, and newsletters, then [Autopilot](/glossary/autopilot) schedules and fans the batch across the eight social platforms plus your blog and a Mailchimp newsletter.

So the honest pairing is sequential. Bounce a texture or beat off Engram, drop it in as the audio bed for a run of Kompozy-generated shorts, and you get videos with a sound that is genuinely yours — no license, no content-ID strike, and no risk of the same trending track that's under a thousand other clips this week. Because Kompozy generates the video, writes per-platform captions, and publishes on a schedule, the audio you made offline on a box in your studio becomes a recurring, on-brand presence in your feed instead of a folder of samples you never used. If you'd rather stay fully software-side, Kompozy also connects to licensed AI-music tools — see [Sonilo](/ai-tools/sonilo), which scores a soundtrack directly from your video, and consumer song generators like [Suno](/ai-tools/suno) — but the Engram angle is the one that gives you audio nobody else can reproduce.

Quick takeaways

  • Thoughtful Things launched a Kickstarter for Engram, a pad-based sampler and groovebox, on September 27, 2026.
  • It runs small generative audio models directly on the hardware — fully offline, no internet or subscription.
  • "Model bending" deliberately glitches the on-board model, via one knob per parameter, to sample uncanny, degraded sounds.
  • Founder Evan King calls it "a field recorder for latent space"; it also samples your own audio via mic or line-in, with a step sequencer, slicing, FX, MIDI, and CV.
  • Early-backer units start at $675 (~30% off); no final retail price yet, with The Verge estimating roughly $850–$900. It is still a research prototype.

Frequently asked questions

What is Engram and who makes it?

Engram is a pad-based AI sampler and groovebox made by Thoughtful Things, a Los Angeles startup founded by Evan King. It embeds small generative audio models directly in the hardware and runs them fully offline, letting you generate, glitch, sample, slice, and sequence sounds without an internet connection or subscription. It launched on Kickstarter on September 27, 2026.

What does "model bending" mean?

Model bending is Engram's signature feature — described as circuit bending, but for neural networks. Each parameter of the on-board audio model maps to a knob, and pushing them past their intended range deliberately breaks the model's audio decoder, producing glitchy, uncanny textures you can then sample and play. The idea is to treat the model's failures as a source of new, original sounds.

How much does Engram cost?

Kickstarter early-backer units start at $675, described as roughly a 30 percent discount. Thoughtful Things has not announced a final retail price; The Verge estimated it would likely be around $850–$900. Engram is still described as a research prototype, so pricing and specs may change before it ships.

How would a content creator actually use Engram?

Engram makes original audio; it does not make or publish video. The practical workflow is to bounce a bent loop or one-shot off Engram and use it as royalty-free audio you own under your videos. A content engine like Kompozy then generates the shorts, carousels, and posts, adds per-platform captions, and publishes them on a schedule across platforms — so your original sound ends up under content that actually ships.

Related news

← All AI news · Get started →