GregRietta AI
FIELD NOTE / 004Back / Lab ↖

EXPERIMENT 004 / SYSTEM ARCHITECTURE

THESYSTEM.

G4-v6 now has a completed pretrained checkpoint, and instruction tuning is teaching it the assistant role. This experiment asks what still has to exist around that model before GregRietta feels like a useful AI instead of a checkpoint sitting on disk.

SYSTEM PROFILE / 01

A model can talk. A system has to remember, act, and connect.

ScopeAI system / evolving
Model layerG4-v6 base complete / SFT underway
RuntimePython / in development
WebsiteSeparate / not connected

THE IDEA / 02

Build only what the model actually needs around it.

G4-v6 finished pretraining across 1.63B tokens, and its best checkpoint is now in supervised instruction tuning. That teaches the model how to answer a user, but it still does not automatically know what to remember, when to retrieve information, which tool to call, or how to fit into a useful product.

Those surrounding pieces will be added one at a time: inference, memory, retrieval, tools, orchestration, and eventually an interface. Gregrietta.com is still separate from the AI today, and that stays explicit until the system is ready.

WHAT I AM TRYING TO LEARN / 03

The hard part starts after training.

01

Is the model the product?

No. The model is the part that predicts the next token. The useful system is everything around it that gives those predictions context, memory, tools, and a place to go.

02

What should get built first?

Only the pieces that solve a real limitation. Memory, retrieval, tools, and orchestration should earn their place by making the system more useful, not by making the architecture diagram larger.

03

When does Gregrietta.com become the interface?

When there is a stable system worth exposing. The site documents the work today; it should not pretend to be connected to capabilities that are still being built.

WHERE IT IS NOW / 04

Do not build the whole future at once.

The model and system layers still stay deliberately separate. G4 now has a local role-aware chat harness for testing the instruction-tuned checkpoint, while this track keeps asking the product question: after the model can answer reliably, what capability should be added next?

NOWInstruction-tune G4 while keeping the model and system layers separate
NEXTEvaluate the tuned checkpoint, then let real limitations pick the next capability

RELATED NODE / EXPERIMENT 001

See the model-generation track.

Open GregRietta G3