A lab that shows its work
We rebuilt Cubby from a documented failure, one validated component at a time. The negative results counted as much as the positive ones.
Measure before you build
Our rule is simple: don't add a component until you can measure what it does. It's slower than throwing everything in at once and hoping the benchmark goes up, and it's the only way we've found to actually know why something works — or doesn't. That includes what it costs to run: compute and energy are things we measure, not things we wave away.
The collapse, and the rebuild
An earlier version of Cubby fell apart. Rather than paper over it, we documented the failure and rebuilt from it, adding one validated piece per rung and scoring each on generation quality before the next went in. A result that says "this doesn't help" is signal, and we keep it. Most of what we've learned came from the things that didn't work.
The ecosystem
Cubby isn't one repository. It's a language cortex in training, a symbolic reasoning machine, a live brain layer, and a world-model agent — separate pieces that each earn their place, developed and tested on their own before they're wired together.
Grilly — our own compute layer, open to everyone
Most AI tooling quietly assumes NVIDIA GPUs and CUDA. We started on an AMD card we couldn't replace, so we built past that assumption. Grilly is a neural-network framework we wrote from scratch with a PyTorch-like API, running on Vulkan compute shaders instead of CUDA. That means it runs on almost any GPU — AMD, NVIDIA, Intel, even Apple silicon — with a graceful fall back to the CPU when there's no GPU at all. It's the same instinct as everything else here: when the standard path is expensive and locked to one vendor, build the one that isn't. Cubby can run on commodity cards plenty of people already own, instead of the scarce, costly accelerators everyone is fighting over.
And it isn't ours to hoard. Grilly is open source under the MIT licence, free for anyone to use: github.com/Grillcheese-AI/grilly. We develop it in parallel with our own models, not as an afterthought — so anyone has the same power we do to build, run, and test this kind of work for themselves.
What we're not showing yet
Some of our reasoning research is currently under submission, so we're keeping the figures off the site until that lifts. When the numbers are ours to share, we'll share them here — and not before.