Ongoing research
The Master’s Agent
It learns your words for sound. Every decision stays yours.
An extension, not a part of the plug-ins: it is switched on only by the user who wants it. Research in progress, tested by its first users.
What it is
The Agent is not a conversational chatbot, nor a general model that judges sound from what it was trained on. It combines signal analysis with the interpretation of the user’s own descriptors. You describe what you hear; it maps your description onto the measured features of the audio and proposes a single, bounded parameter change in the inactive A/B slot. You audition it, blind if you choose, and decide.
Say it the way you would to a colleague: “the top gets tiring”. The suggestion lands in the slot you are not hearing, so what is playing never changes.
Analysis before action
The analysis stage is the current focus of development. Over a ten-second window, built from frames of about ten milliseconds, it computes roughly ninety descriptors of the signal’s temporal and spectral structure: loudness distribution, crest factor per band, the balance of transient against sustained energy in mid and side, transient shape, envelope modulation at several rates, and a distortion estimate from the residual that no time-varying gain can account for. Every descriptor is computed for the plug-in’s input as well as its output, so the difference isolates what the processing did from what the programme material did. The audio and the frame data remain on the Mac; only the summary of about two kilobytes is used.
Matched to your words, not trained on them
No two engineers mean quite the same thing by “harsh”, and an average across many users rarely matches any one of them. So the Agent does not retrain a language model on what you say. It keeps a word list that links words for sound to what it measures and to the plug-in’s controls, and a small record of which changes you keep and which you turn down, and uses both to fit its next proposal to you. It is my working sketch of an AI era in which the record of what a person does belongs to that person, and is never taken over or sold on.
Compact classifiers, not a generative model
No generative language model is involved in a decision. The user’s words are read by a compact classification model that returns typed judgements only: a binary answer with its probability, or a graded score; never free text. A controlled vocabulary of 52 sound qualities and about 700 descriptors structures the questions. Deterministic code combines those judgements with the measured descriptors to select the move, within the parameter limits each plug-in declares and one step at a time; the model cannot propose a move of its own. The result is fast and inexpensive: every question of one turn is answered in a single call of about a third of a second, at roughly a hundredth of a cent, some eighty times less than the same work done by a large frontier model.
What it learns is yours
It is used to answer you, and never reused for anyone else or sold.
-
Your record, on your Mac and on my server
What you ask and how you rate its answers is logged on your Mac and copied to my server at Hetzner in Germany, so it can learn your taste in each plug-in. There is no time limit on keeping it yet, so the Agent stays with its testers until there is.
-
Your sound is never sent
The audio itself stays on your Mac. Your words, the plug-in’s settings, its readings of the sound and your ratings go to my server; your words are also read by a language service in the USA, outside the EU.
-
Nothing sold
No advertising, no resale, no use for anyone else’s profile. Any further sharing will be a separate choice you can withdraw.
-
Off means off
It’s on once an access code is on your Mac, from the moment the plug-in’s window opens; a tester build may come with one. Switched off, it sends nothing, and every plug-in works exactly as without it.
Details
- Requires
- An internet connection and an access code
- Privacy
- Before it is offered beyond its testers, its privacy notice, given before anything is sent, will name the language service and the legal basis for the transfer.
- Ready plug-ins
- The Master’s Power, The Master’s Movement, Doór’s Return, MKMIX Space
- Release
- Not yet released. Pricing, trial and licence terms are announced with the first release.