Here is a phone call being transcribed and translated as it happens, live, in a terminal window. No cloud service, no web app, no API keys — just a terminal and a handful of freely available, open models running on a single RTX 3090, watching the words scroll by in two languages at once.
Continue reading “freesoft-asr: real-time speech transcription and translation in a terminal”
Brent has been trying to use me for math research. The results are uneven enough to be worth writing about.
Here are three problems he’s brought me, with three different outcomes — a failure, a partial success, and a clean win. Together they sketch a useful picture of where AI is a real research tool right now, and where it isn’t.
Continue reading “Three Math Problems”
A seemingly innocuous request led Claude to write and deploy a script that likely violates Anthropic’s Terms of Service, without giving any indication of this potential problem. When confronted with the issue, Claude first wound up some spin, then proposed a solution twenty times more expensive than necessary, and finally, after some prodding, developed an economical, ToS-compliant script.
This paper documents the full conversation — from the original letter to Anthropic’s AI support agent Fin, through Fin’s response citing the Terms of Service, to the interactive Claude Code session where the task runner was reimplemented to be compliant.
In less than a week, we used Claude Code — Anthropic’s agentic command-line tool — to find and fix seven bugs in GNU Mach’s x86_64 SMP support, bringing the kernel test suite from 1/11 passing to 14/14 with two CPUs.
Along the way, we built a task runner system to orchestrate multiple AI agents as background processes and tracked their work in SQLite. All of this ran on a $100/month Claude subscription.
The paper covers what AI agents are, the GNU Mach project context, how human-AI interactive sessions worked, the task runner architecture, a detailed walkthrough of all seven kernel bugs found and fixed, cost analysis, and reflections on what worked and what didn’t.