Technical work so far
KodAi Research Builds a Foundation
Our work is documented with the belief that transparency in how these systems are built earns its place.
The premise00 / Two engines
Our architecture was built to operate within the present and future of rapidly developing local models.
The macOS build runs on Apple Foundation Models, the on-device model Apple ships with Apple Silicon and A16+. As the model develops, its characteristics will be better understood and reported.
The current energy is spent on the mobile model which utilizes llama.cpp.
01 / The premise
A local model.
An inspectable decision.
KodAi iOS is a local-first AI workspace built inside a glass box. The model runs entirely on the phone through a quantized llama.cpp engine (LFM2.5-1.2B, 4-bit, a 2,048-token window) — no keys, no server, no round trip. And almost every decision the model makes on its way to a word is made inspectable.
Most chat apps show you the answer. KodAi shows you the deciding — by reading the model's raw probabilities before the sampler picks, at every single step.
One source of truth
Four numbers carry the visual language.
- Probability
- of the chosen token color
- Entropy
- the spread of the distribution size / width
- Margin
- the gap between the top two candidates
- Surprise
- −log p of what was actually emitted
Gold is reserved for the exceptional moment where sampling lands on something other than the model's single favorite token.
02 / Three views
One telemetry stream.
Three expressions.
Cyan always means “probable,” gold always means “the sampler overrode the favorite,” width or size always means “uncertain.” Learn the grammar once, read all three.
View 01 / Globe tracing
The Decision Globe
One response, wrapped onto a transparent glass sphere. Each analyzed token is a bead on a pole-to-pole spiral. Color is probability; size is entropy. A tracer ribbon follows generation order. Tiny gold satellites mark where sampling diverged from the favorite.
A scrubber replays generation in order: the past remains a quiet trail, the present glows, the future recedes.
Sampling, made visible.
View 02 / Follow the River
The river
The same response, told as moving water — a scroll-driven, full-screen nocturnal channel. The spine is the chosen path. Channel width is entropy. Rejected candidates fan off as tributaries. When sampling rejects the strongest current, the channel forks and leaves a ghost branch behind.
The playhead stays fixed at center; scrolling follows the current one decision at a time. Brightness and width repeat the signal, keeping it legible without color alone.
Width as uncertainty. Fork as sampling.
View 03 / Thread Atlas
The token atlas
Zoom all the way out. Each traced user-to-assistant exchange becomes a continent. Area is its share of the 2,048-token context window; tint is mean confidence; latitude is chronology. A vine threads the continents in order. A continent containing a sampled-over-favorite choice gets one gold beacon.
Focus a continent and its tokens scatter into a local spiral. Open token replay and the overview gives way to one response, then one decision.
Context as a finite, spatial resource.
The accumulation
Different windows onto one organism.
The engine, telemetry, visualizations, sampler controls, and context engine are not separate features. Add a measure in one place and three views light up.
We're reading a 1.2-billion-parameter mind one decision at a time and finding it has geography, weather, and weather we didn't put there.