EXPERIMENTS / IN PROGRESS
THELAB.
The engineering notebook behind GregRietta: what I am building, why I tried it, what happened, and what the next experiment changes.
G4 / TRAINING
ACTIVEWORKING SET / 04
What I'm working on.
Each field note starts with the question, then gets into the build. The details are still here, but you should not need a research paper to understand why they matter.
GregRietta G4
ACTIVEWhat changes when a pretrained 69M model is taught how to answer people?
G4-v6 completed its 1.631B-token pretraining run. The best checkpoint, selected at step 793,750 with 2.5328 validation loss, is now being instruction-tuned on 27,024 assistant examples.
MicroBlade
BUILDINGCan a small compute cluster grow cleanly one machine at a time?
A modular 3D-printed rack where mounting, airflow, wiring, networking, and expansion are treated as one engineering problem.
Local intelligence
EXPLORINGWhat actually gets better when AI runs on hardware you own?
An exploration of latency, privacy, reliability, and when coordinating several local machines is worth the added complexity.
System architecture
EVOLVINGWhat has to exist around a model before it feels like an AI system?
The longer-term work around inference, memory, retrieval, tools, orchestration, and eventually a public interface.
OPERATING PRINCIPLE / 01
Explain the problem before the implementation.
Show the result before the jargon.
This is a build log, not a white paper. The goal is to make the engineering interesting to follow and useful to learn from.