AI
Machine Intelligence in the Loop
On 16 September 2026 I gave the PHAST Department Colloquium, Machine Intelligence (MI) in the Loop: A Physicist’s Research and Teaching Relationship with MI Agents. The talk was about methodology for research and teaching. Machine intelligence that works on its own, as an agent, is the most sudden new research methodology I have seen. It has changed how my computational astrophysics group works and what it attempts, and I expect it to change all knowledge work at a university. The talk shows how the group works with such agents, from our own measurements and the agents’ own words, set against outside reports. It then asks what a university should do with this new power and argues for taking it up on our own terms. For teaching, it proposes practical steps through the platform Vela.
The slides as delivered and a written version of the talk are in a shared Nextcloud folder, open to everyone with no login. The written version is a draft. Comments are welcome, and page 2 of the PDF says how to send them.
The AI-for-science repository
Materials on working with AI in research and teaching are collected in the repository AI-for-science on UVic’s GitLab. Unfortunately, for now it can only be opened with a UVic GitLab login, on campus or through the UVic VPN.
The repository holds:
- AI Sessions, a book on working with AI sessions, from a single session to agents. It is an early version and under constant revision.
- Machine Intelligence (MI) in the Loop, the slides of my colloquium.
- Session dashboard, one local web page for every Claude Code session running in tmux.
- Marina, a keyboard-first file manager for macOS with a terminal in each tab. It is also public on GitHub.
- Vela and the token spigot, described below.
- A draft paper on 1D models of massive main-sequence stars, the research example in the colloquium.
Vela
Vela is a workspace in the browser where a person and an AI agent do real work together on the same files, for a course or a research project. What the agent may do is limited on the server, and every turn is recorded, so the work arrives with its history. Each person’s work is closed to other users, and the work stays on the Vela server.
- UVic pilot. The UVic pilot is open for testing by invitation.
- Token spigot. A self-hosted open-weight language model on UVic’s Arbutus cloud that your own tools can use. See tokens.velella.ca.
- Explainer and slides. A three-page explainer (PDF) says why I built Vela, what a vela is, where it stands in November 2026 and how to get involved. The slides (PDF) go with it.