03 · Scale
Built at 20B. Aimed at 1T.
Scale is stated plainly — where the project is now, and where its architecture is designed to go. No inflated claims in between.
current scale
20B params
next milestone
100B params
architectural target
1T params
Why start at 20B
A scale we can actually train, evaluate and learn from honestly. Small enough to iterate the loop daily, large enough to generalise. Every capability ships on the 20B system before scale becomes a distraction.
The 1T question
The target is not 'a bigger single model'. It is the architectural question of how memory, routing and self-modification should behave when the substrate is a trillion parameters across a distributed structure. A research question, not a promise.
Honestly hard
Compute cost, data quality, catastrophic forgetting and evaluation are all unresolved. We report real blockers, not smoothed curves. Scale is a consequence of the loop working, not the other way around.