Rootcomputer is an independent AI research lab focused on understanding, building, and evaluating modern machine learning systems.
Our work centers on small and mid-scale language models, efficient transformer architectures, training dynamics, alignment, model behavior, and the internal representations that emerge during training.
We are especially interested in how architectural choices, data design, and optimization strategy affect reliability, reasoning, grounding, and failure modes in neural networks.
Rootcomputer conducts applied and exploratory research across several areas:
Rootcomputer is built around the belief that meaningful AI research does not require scale alone.
Small and mid-scale models provide a controllable and experimentally rich environment for studying how models learn, fail, generalize, and align. By constraining size, we hypothesize that we will gain some clarity into the mechanics of training and behavior that are often harder to isolate in very large systems.
Haiku Mini is a compact conversational research model focused on grounding, and behavioral consistency in small language models.
It is designed as a controlled platform for studying turn-taking, instruction following, hallucination, and conversational stability under tight parameter constraints.
Tanka extends the Haiku research line into simple reasoning, more consistant factual recall, utilizing about twice as many total parameters as Haiku Mini.
It is intended for internal research, evaluation, and architectural experimentation. Limited public access due to the nature of an experimental project.
Rootcomputer treats training data as a first-class research object.
Our datasets and corpora are designed to study:
We also build custom training software and infrastructure for controlled experimentation, including transformer training pipelines, reproducible evaluation, and limited-hardware friendly solutions.
Safety and alignment are central to Rootcomputer’s research direction.
We study alignment as a property shaped by architecture, data, training dynamics, evaluation pressure, and deployment constraints. Our work emphasizes predictable behavior, controlled interaction, failure-mode discovery, and transparent model limitations.
Rootcomputer AI Development Branch