Hello,Soya.tech

Gemma 4 runs on 4GB, fits on a MacBook, and writes production code. Open weights, the exact setup, the benchmarks, and the honest limits — after two weeks of use.

15 hours per week back, same output quality. The exact delegation matrix between me and the AI — which 70% to hand off, and which 30% you must never give up.

What happens when you run research, drafting, editing, and publishing through one model for 90 days? The time savings, the failure modes, the full Claude workflow.

Anthropicが公開を見送った未公開モデル「Claude Mythos」と、背景にあるProject Glasswing。日本語で読める最長級の解説と、AI業界の構造分析。

Anthropic built Claude Mythos, decided not to release it, and told the world why. Inside Project Glasswing — and what withholding a frontier model actually signals.

DeepSeek, Qwen, GLM, Kimi, MiniMax — the five Chinese AI models Western media keeps quiet about. Real daily use, real benchmarks, real $200/month savings. Reviewed.

90 days of real production work across Claude Code, Cursor, and Copilot — same tasks, same days. The final scorecard, the winner for each type of coding, the cost.

The exact $20/month AI stack I use to run a 700K-view content business — every tool, every cost, every cut. No VC money, no paid Claude Plus. Full breakdown.