Ternary computing · Wave architectures · Efficient AI
The creators of TAARDIS
Efficient Ternary LLMs and Experimental AGI Architectures
Big questions. Hands-on experiments.
Building what comes next, one idea at a time.
Selected work
Open the source. Inspect the approach.
Help shape what comes next.
01 / Ternary computing
Give model weights
a smaller footprint.
Base-3 weight packing, selectable weight and scale formats, GPU kernel experiments, and a Model Atlas for inspecting per-layer choices.
Tap a tile to change its value and watch the encoding update.
Each tile has three states: −1, 0, and +1.
From the published specification
TK k1 can describe a full group of 20 weights using 32 code bits and a 12-bit scale.
Four bytes hold the 20 one-trit codes. A packed 12-bit scale adds 12 bits: (32 + 12) ÷ 20 = 2.2 bits per weight. This nominal full-group comparison excludes padding and file metadata. It does not measure inference speed or model quality.
Calculated representation size · full groups · packed scales
Source & assumptions ↗
02 / Agent workflow
Keep an AI-assisted project moving. Save progress, capture what changed and what was tested, then continue when you choose.
Explore GRIP
03 / Full-ternary model
Explore the lab’s full-ternary model project, its published files, and the runtime work behind it. Usage and limitations live in the model card.
Explore the modelOn the workbench
Research in progress.
Room to experiment.
A visual workspace for AI components, model flows, and training pipelines, built around the GNU Radio Companion interface.
Exploring geometric models using in-phase and quadrature (IQ) waves.
06 / Geometric architectures
From geometry
to the emblem.
A visual study of the lab’s emblem. The folding shapes in this sequence are planned directions for future models.
Explore current model shapesThree outer triangles fold around a fourth face.
Current model geometries
Explore the individual geometric components, then the complete IQ wave model built from connected fractal crosses and dodecahedron chambers.
A cross branching into smaller crosses in three dimensions.
Twelve pentagonal faces enclosing a connected volume.
A ring-shaped surface looping around a central opening.
Geometry working together
Connected fractal crosses link large dodecahedron chambers, with smaller repeating cross forms along each branch.
This combined geometry is the wave model currently being trained, exploring geometric models through in-phase and quadrature (IQ) waves.
Color = phase · brightness + size = amplitudeIllustrative wave activity
Three component studies and the complete wave model currently being trained. The emblem sequence above shows additional shapes planned for future models.
These projects are not released yet. Follow the lab for public updates.
HUEYTOWN, ALABAMAThe person behind the experiments
INTERNECTUAL (adj.)
Someone who has cultivated a substantial body of knowledge via the web.
I’m Cody Dixon, a hands-on researcher and builder in Alabama. I connect ideas with published research and use AI to help turn those ideas into working tools.
“Today’s minds are always on the internet, but it’s what you do with it that sets you apart.”— Cody DixonMore about my work ↗
Support independent research
More GPU time means more room to evaluate models, reproduce results, and keep practical tools moving forward.
Build with me
Reproduce a result, report a bug, improve a tool, or bring a different perspective.
