Research approach

Artificial Intelligence Research by Hatsunama

Hatsunama's artificial intelligence research examines AI models and the larger systems required to turn models into agents capable of useful work.

A model may produce a correct answer but still fail a real task. The surrounding system must provide context, choose a tool, construct arguments, execute the action, observe the result, recover from errors, and verify the outcome.

Model → inference → harness → tools → result

These layers can change task performance even when the underlying model stays the same. Evaluating an agent therefore requires looking at the full path from a request to a verified result, not only the model's final sentence.

Google visibility experiment: closeout — October 27, 2026. Findings and limitations.

Research notes

• Hatsunama Small-Model Training: What We Know and What Still Needs Testing · October 20, 2026. Verified findings, resource-limit lessons and the next questions for more reliable local AI.
• Hatsunama Small-Model Research: Balanced vs Repair-Weighted Training — October 18, 2026. A paired development experiment comparing balanced and repair-weighted training.
• A Translation Review That Can Say “Unresolved” — a bilingual review method and reusable synthetic template.
  Does the Harness Speak the Model's Training Format? — a practical interface debugging method.
  Give a Small Local Model a Prompt That Fits — a practical runtime budget worksheet.
  What My Local Translation Candidate Tests Actually Tell Me — documented checks and open questions.
  A Tool Timeout Is an Unknown Outcome — inspect state before repeating consequential actions.
  Correcting AI Captions Without Changing the Speaker's Meaning — a practical review workflow.
  Did a Local AI Change Help? A Reproducible Comparison — fixed cases, explicit expectations and a downloadable JSONL template.
• Tool-Call Output Is Not Task Completion — what model tests measure and what agents must still verify.
• AI Model vs AI Agent vs Agent Harness — the distinct roles of models, agent loops, and the software around them.
• AI Tool Calling vs Tool Use — how a valid call differs from execution and a verified result.

Translation and captions

Testing Translation: Meaning, Speaker Intent and Instruction Boundaries — a reproducible method for evaluating semantic preservation and application behavior.

Announcements

• Hatsunama’s Free App Projects: Small-Model AI Research in Practice — October 13, 2026. Current app development, evidence limits and the practical questions behind the work.
• Independent AI research and open-source systems work — September 30, 2026.

Public artifacts include model experiments and software repositories. See the artifact index for direct links and the identity page for background.