TL;DR: A musician used AI coding tools to help an older Zoom effects pedal run small neural models of guitar amplifiers. The interesting part is fitting a useful sound into hardware that was never designed for this job. It builds on years of human reverse engineering, and other owners have begun testing the results.

The maker’s overview and sound demonstration:

YouTube

The problem: a working pedal with a fixed menu

A guitar effects pedal takes an electrical signal from an instrument and changes it before it reaches an amplifier or recording interface. Delay creates echoes; distortion changes the waveform; reverb suggests a room. A digital pedal does this with software running on a specialised processor, or DSP.

That means the hardware may be capable of sounds its manufacturer never supplied. Getting new code onto it, understanding how the processor receives audio, and finishing the calculation in time are separate problems.

Roman Rusinov, who releases music as Farlow, took on the original Zoom MS-70CDR. In an October 19 interview with The Verge, he described using AI coding tools to help develop custom effects and support for neural amp captures. His project builds on Jannik Assfalg’s existing MultiStomp toolchain and other developers’ hardware research. AI assistance accelerated a particular contribution to a much longer community effort.

What is an amp capture?

An amplifier changes a guitar’s sound in ways that depend on the incoming signal, including how hard a note is played. Neural Amp Modeler, created by Steven Atkinson, uses machine learning to model that response. Its training software exports a model as a .nam file; compatible software or hardware then runs that model on new audio.

Think of a capture as a learned description of a sound-producing setup. It does not automatically reproduce every knob setting or every amplifier ever made. The capture and its supported model architecture matter.

A cabinet impulse response, usually shortened to IR, describes another part of the chain: the filtering associated with a speaker cabinet and its recording setup. Amp capture and cabinet processing can therefore be separate building blocks.

The solution has to fit inside an audio deadline

The project’s browser loader packages supported A2 Lite captures into .ZDL effects the older pedal can load. It also fits cabinet IRs into a compact filter. New effects can run on the stock firmware; replacing the pedal’s entire operating system is a different operation.

The small model is a practical constraint. Larger NAM architectures do not fit this implementation, and several effects compete for the same processing budget. An effect that finishes eventually is useless if the next piece of audio needed it earlier.

The published investigation is unusually instructive. The developer compared the network’s output with a reference implementation, tested actual target instructions in an emulator, and made hardware diagnostics. Passing the arithmetic comparison did not establish that the physical pedal could meet its deadlines: real memory behaviour and scheduling still mattered. Reduced-rate processing created another trade-off, gaining headroom while changing the sound.

Other owners make the story stronger

In the public test thread, JPM100P3 confirmed NAM and cabinet effects on an MS-50G running firmware 3.10, then reported working knob changes, saving, backup and restore. They also found naming and control problems that led to fixes. This is a specific owner’s report, rather than proof of compatibility across the whole Zoom range.

The Plus and Four families use a different effect format and are not supported by this ZDL project. The maker still labels parts of the workflow experimental and asks users to back up their patches. I reviewed the documentation and reports; I have not installed it or tested a pedal.

For musicians, the wider possibility is appealing: start with a sound or workflow you understand, use AI to help bridge the implementation gap, and let actual measurements and listening decide whether the result works.

Project, loader and compatibility:
github.com/themanro/ZoomMultistompZDL

Neural Amp Modeler:
www.neuralampmodeler.com/

Hardware investigation:
github.com/themanro/ZoomMultistompZDL/blob/main/docs/NAM-RUNTIME-INVESTIGATION.md

Owner’s test report and follow-up:
github.com/themanro/ZoomMultistompZDL/issues/5

Maker interview:
www.theverge.com/gadgets/1013493/vibe-coding-ai-music-gear-hacking-custom-firmware