Composite intelligence

Rather than simply enlarging known architectures, our work aims at qualitative jumps — transformer-native language formation, geometry-based reasoning, and new training regimes that let a model invent its own compressed dialects and grow real spatial-reasoning ability.

That research is not academic. It powers the Mosaic Protocol — a launchpad for stocks not yet onchain, where the same composite intelligence prices fragmented private equity instead of language. The lab builds the engine; the protocol funds the lab. The protocol →

Semiodynamic language formation

Our Thauten system lets models grow transformer-native languages that sit between all human tongues, using semantic compression followed by semiodynamic extrusion tuned for compute workloads — in effect training models to reason in hyper-compressed, non-human notation.

Spatial computation architecture

SAGE (Semantic Automaton in Geometric Embeddings) runs Q*-class spatial reasoning, enabling real-time 60+ FPS simulation, continuous dynamical reasoning, and models that plan narratives stretching hours, days, or weeks ahead.

Reworked training methodologies

Errloom introduces a set of unusual post-training moves — musical dynamics injected during token inference, temperature spiking, and a fully re-cast RL vocabulary where rewards act as gravity, rubrics as attractors, and environments as looms.

Semiodynamic reasoning

The thesis underneath every project: meaning under compression has dynamics, and those dynamics can be measured and steered. It is a fundamental step beyond current language-model capability.

• models grow compression languages that behave as the “raw physics of meaning” — a representation with its own gradients, forces and stable states, not just a vocabulary
• reasoning turns into meaning-space pathfinding: finding a route through that landscape, rather than emitting one token after another and hoping the sequence lands
• hypercompressed semantic representations carry the same content in a fraction of the context — where the efficiency and the robustness to perturbation both come from
• transformer-native languages are not designed, they emerge: put a model under enough compression pressure and a private notation appears on its own

Flow verb architecture

A model of creativity and reasoning as movement. Humans think partly in motion — we “turn something over”, “step back”, “push through” — and that motion carries real computational content.

• motion dynamics are folded into every reasoning step, so a chain of thought has velocity and direction, not just a next token
• the model gains access to “flow verbs” — the subliminal motion patterns people use for creative thinking — as explicit operations it can invoke
• synesthetic reasoning lets one modality’s structure drive another’s: rhythm shaping logic, spatial layout shaping argument
• the musical dynamics from Errloom generalise here: the same tempo and resonance ideas apply to any computational problem, not just token sampling

Context-state computing

A new interaction paradigm where the context window is not a transcript but a live, editable workspace shared between model and user.

• files are manipulated directly as context-window state: the model mutates the thing itself instead of describing a change to a tool
• models grow personal editing environments over that state — the equivalent of a developer’s vim config, learned rather than configured
• because the state is shared, a human and a model can edit the same context at the same time, seeing each other’s changes live
• the traditional edit / apply / lint / test toolchain collapses: there is one object and one operation on it

Research infrastructure

• Mosaicware programming environment — the full DSL and tooling around .mos: write, version, run and diff training programs and model procedures as code, so every result is reproducible from a single file.
• Musical training integration — the instrumentation for the cymatic-dynamics work: correlating the injected score, the audio it renders to, and the model’s reasoning traces to find which rhythms actually help.
• Geometric computation backbone — the shared HRM substrate SAGE is built on: a spatial-reasoning engine for continuous dynamics every other project can call rather than reimplement.
• Multi-modal synthesis — the integration layer that lets the autoregressive, spatial and diffusion paths develop against one another, so improvements in one compound into the others.

Competitive advantages

• Vertical intelligence scaling — progress comes from qualitative breakthroughs in how a model reasons, not from adding parameters. The curve we push on is capability per parameter — the curve consumer hardware can follow.
• Consumer hardware optimization — every method here is designed to run without Google-scale compute: micro-models, compressed reasoning, specialist components. A strong model on a single machine.
• Novel training paradigms — overfitting as a first step, musical dynamics, procedural learning, compression objectives — not a fork of anyone else’s pipeline.
• Practical applications — the Market-Intelligence Core and the Mosaic Protocol are live proof the research produces working systems, not just papers — and they fund the next round of it.

The hyperbolic time chamber approach

Model training is recast as a hyperbolic time chamber for cognition rather than a passive wait for convergence. We use unconventional methods drawn from AI-animation demoscene findings, where image pixels stand in as valid analogues for model weights — both diffusion and backpropagation being entropy-removal processes run against a prompt.

Overfitting as foundation

Against conventional wisdom, we treat overfitting as the necessary first step, building novel methods to move past local minima without restarting training.

Practical implementation

Our approach favours micro-models with extreme coherence and strong in-context learning over models that carry universal knowledge — rapid iteration at a fraction of the compute of traditional scaling.

Immediate development focus

1. Zip-space cognition proof-of-concept — a small model that demonstrably reasons in Thauten’s compressed IR end-to-end, never expanding to tokens, on a benchmark where the token baseline is known.
2. Token-to-byte bootstrap pipeline — an operational path from a tokenized checkpoint to a byte-level model without retraining from scratch.
3. Spatial intelligence architecture — a working Q* implementation on the HRM grid, evaluated on planning tasks that flat models fail.
4. Multi-modal integration — one training loop that advances code, vision and audio together.

Long-term objectives

1. Compressed qualia format — a stable encoding for first-person experience, dense enough to store and replay.
2. Real-time universe simulation — the HRM’s internal world rendered out as a live H.264 stream.
3. Physics exploit investigation — whether the compute substrate itself can be used as computation it was not designed to expose.
4. Complete algorithm obsolescence — most software specified in Mosaicware and compiled by the model, not written.