Introducing Quantum Frequency Modulator: Next-Token Prediction Without Attention

The Quantum Frequency Modulator (QFM) is a web app that generates text through a novel method called frequency-domain next-token prediction. Unlike traditional large language models which utilize transformer attention mechanisms, QFM eschews attention matrices altogether. Instead, it converts words into sound frequencies, collides those frequencies within a lattice structure, and then interprets the resulting spectral signature to determine the next predicted word.

This groundbreaking approach eliminates the need for server-side model weights, attention heads, or recurrent state beyond a running frequency signature, with the entire process occurring in-browser.

The QFM system is relatively simple yet intriguing. It consists of a visual interface with four live oscillators, a master waveform, a sequence graph, and a pipeline panel. Users can witness the prediction process in real-time as audio and visual elements are generated simultaneously. The four oscillators work in concert: one sets the initial frequency for the content word, another generates the next token's frequency (often amplified for clarity), a third calculates the interference between frequencies known as lattice collision, and the fourth aggregates these collisions to produce a final frequency signature that drives the final draft.

The pipeline begins with pre-tokenization, where the system extracts the content word from textual input, determines its intent (explanatory, questioning, coding, creative, or statement), and performs OSINT fan research using free online resources like dictionaries and encyclopedias. This stage also includes optional assistance from free LLM endpoints to provide context for the pre-read.

Once pre-processed, each word is transformed into a 64-dimensional vector using character n-gram features, part-of-speech biases, and related-word semantics. This vector is then projected onto a fixed random direction, resulting in a frequency within a specific audible range (around 80 Hz to 1600 Hz).

The four oscillators play distinct roles. Oscillator A sets the anchor frequency for the first word, while Oscillator B generates subsequent token frequencies, often slightly boosted to ensure distinct audio output. Oscillator C performs the lattice collision analysis between the anchor frequency and the current prediction frequency, producing peaks that serve as both visual cues and a measure of interference strength.

Oscillator D accumulates these interference signatures and ultimately emerges as the sole oscillator for the final frequency signature that drives the text generation. All processing occurs in real-time, allowing users to see and hear the prediction unfold as it happens.

One of the key departures from traditional transformer-based models in QFM is the absence of recurrence and attention matrices. The system replaces attention with frequency collision strength and spectral continuity, allowing for a more direct relationship between the generated frequencies and the final text output. The prediction loop is bounded to maintain predictable latency, ensuring the generation process remains responsive and accessible.

The final step involves performing a full discrete Fourier transform (DFT) on the accumulated signature to identify the dominant frequency, which then drives the last oscillator for a brief period before fading out. The sequence of predicted tokens completes the draft text, offering an alternative to the dominant transformer attention paradigm.