THELAB.

Training a compact transformer from scratch instead of treating the model as a black box.

A 150M-parameter TinyStories run used to study tokenization, training behavior, inference, and what it actually takes to make a model become useful.

A small, modular compute system built around compact x86 nodes and physical iteration.

A 3D-printed rack for t740-class nodes designed around airflow, networking, cable routing, serviceability, and the ability to expand one machine at a time.

Finding the boundary between what needs the cloud and what can live on hardware you own.

Experiments in local inference, orchestration, latency, privacy, and distributed workloads—less about benchmarks, more about useful systems.

The public surface of the experiment, built to change as the underlying work changes.

Part identity, part interface, part proving ground. The site is intentionally small now so future tools, models, and experiments have somewhere coherent to emerge.