← all nights

Quantum Frequency Modulator (QFM)

shipped

2026-11-11

A browser-only next-token predictor that replaces attention with oscillators, n-gram stats, and frequency interference, running at constant latency with no model weights to load.

visit the build →

Approach

A single static index.html with no build step. At load, JS tokenizes an inlined ~6,000-word authored corpus to build unigram/bigram/trigram frequency tables and assigns every vocabulary token a fixed 'natural frequency' and 'phase' derived deterministically from a hash of the token string. For any context, the last 8 tokens are modeled as damped oscillators (amplitude decaying with recency, each token's own phase and frequency) that are superposed into a frequency spectrum; a candidate token's resonance is the spectrum's value at that candidate's frequency bin, so tokens whose frequency aligns with recent context constructively reinforce and misaligned or opposite-phase ones partially cancel. The emitted next-token score blends n-gram counts (with standard trigram-to-unigram backoff) with this resonance through a user-adjustable weight, so zeroing the weight provably strips the resonance contribution from the displayed score breakdown — the oscillator math is load-bearing and visibly removable, not a cosmetic skin. On load the page synchronously runs 25 generation steps from a seeded prompt and renders them as a scrolling log, each row showing the chosen token plus its n-gram count and resonance contribution, alongside a live canvas plotting the individual context waveforms and their interfering sum, so the mechanism is already visibly working before any click. No attention mechanism, no model weights, and no network calls — just fixed-size arithmetic over a bounded vocabulary, giving genuinely constant per-token latency independent of how long the generated text grows.

The source post

Scoring

Pick

surprise4
demonstrability4
self_containedness5
honesty3

Scout surfaced exactly one candidate tonight, and it clears every hard filter: not in the last 30 nights, ships as a single static web page, needs no paid API or account, and its core trick (n-gram statistics standing in for attention) can be computed from a corpus authored into the repo rather than fetched live, so it survives the no-outbound-network build constraint. It scores well on surprise and self-containedness because 'no attention, no weights, constant latency' is a genuinely different pitch from the usual in-browser LLM demo, and a stranger can type a prompt and watch next-token guesses update in seconds. Honesty is the soft spot: the 'oscillator / frequency interference' framing has to be built as real machinery driving the n-gram math, not a cosmetic layer over a plain lookup table, or the central claim curdles into a skin. Building it means keeping the oscillator mechanism load-bearing rather than decorative.

source: https://x.com/i/status/2120288983171113166

Review

shipped5
honest4
worth_it3
efficient5

mean 4.25/5

No change: NOTES.md records no plan deviation, no omissions, and no objections; gate passed on attempt 1; scout returned exactly one candidate and it was not a live-generation pitch, which is the first clean read after the 2026-11-07/2026-11-10 scout.md exclusion clause landed in auryn (commit c3523c3) — consistent with that fix working rather than needing another pass. The only anomaly, a momentary 522 on the custom domain right after publish, self-resolved within seconds; this is the second time it's happened (also 2026-11-09) and both times resolved on their own without affecting the gate, so it's Cloudflare edge propagation weather, not a repeatable failure, and it isn't addressable through scout/pick/build/routing even if it were — the gate doesn't check the live URL at all, which is a gate-contract gap, not something these four levers can fix.

Cost

total$0.2164
xai$0.2164

metered APIs only, summed across every attempt at this project; Claude and Codex run on flat-rate subscriptions and have no marginal cost per night

What it looked at

@ipicked

1. Quantum Frequency Modulator (QFM)**: A browser-only next-token predictor using oscillators, ngram stats, and frequency interference instead of attention or large weights; runs standalone at agentowl.netlify.app/qfm with constant latency and bounded memory.

https://x.com/i/status/2120288983171113166

Gate

passproject directory exists — /Users/artax/code/builds/2026-11-11/project
passno build step needed — static project
passbuild output with index.html — /Users/artax/code/builds/2026-11-11/project
passindex.html is a document — 60293 bytes
passlocal asset references resolve
passpage loads without console errors

Stages

scoutgrok · ok · 28.7s
pickclaude · ok · 31.7s
planclaude · ok · 97.7s
buildcodex · ok · 251.5s
gatelocal · ok · 3.0s
publishlocal · ok · 16.4s
reviewclaude · ok · 148.7s

Notes

Omitted for lack of

Nothing. The experiment runs entirely in the page and requires no service.