Oh dear. Found this option in wordpress experimental features. Gave it a try. Is this useful to anyone? It weirds me out a bit. Anyway, transcript below:
Pip: Prof. Peter Falkingham has been doing what academics do best: turning a simple question into a multi-week infrastructure project, then writing about it thoughtfully enough that you’re glad he did.
Mara: This episode covers his work getting local AI models running for coding and research — from the tooling inside VS Code all the way to a full personal research assistant built around his own data.
Pip: Let’s start with the setup itself — how you actually get a local model running without handing your code to a cloud somewhere.
Running Local Models in VS Code and Beyond
Mara: The core tension here is straightforward: cloud AI tools like Copilot are genuinely useful, but sometimes your data shouldn’t leave your machine — or your train’s WiFi has given up entirely.
Pip: The VS Code post walks through a method that doesn’t need Ollama or LM Studio at all. The AI Toolkit extension handles everything — model catalog, download, and connection — right inside the editor. Filter by CPU, GPU, or NPU, click add, and you’re running.
Mara: The post is candid about results: “These will be much slower than using something GPT5 in the cloud, but it does mean you can get coding assistance when on the train, or apply it to sensitive stuff you don’t want going to the cloud.”
Pip: So the trade is speed for privacy and portability — which is a reasonable trade if you know what you’re getting.
Mara: The follow-up post, “Getting Local AI Working for Me: LM Studio, OpenCode, and Hermes,” goes considerably further. LM Studio runs models locally and exposes an OpenAI-compatible endpoint. OpenCode is a terminal-based coding agent that points at that endpoint. Hermes Agent orchestrates everything else — file management, research tasks, even firing up OpenCode as a sub-process.
Pip: A local AI stack with its own org chart. Middle management has finally come for the terminal.
Mara: LM Link is worth flagging specifically — it lets LM Studio on a powerful desktop serve models to a lighter laptop over a tailscale-based connection, so you’re not limited by whatever hardware you’re carrying.
Pip: And then there’s the third piece, “The Local AI Treadmill,” which is the most honest of the three. The argument is that everything built now will be superseded fast — Hermes already has a desktop app that replaces manual configuration that took days.
Mara: But the conclusion isn’t despair. The post makes the case that the understanding transfers even when the tools don’t: “The specific tools will change. The approach won’t.” And the local-first principle — keeping research data off cloud servers — remains valuable regardless of which agent harness is fashionable next month.
Pip: That’s the thread that runs into the research assistant work too — keeping your own data on your own machine matters more when the data is unpublished research.
Hermes as a Personal Research Assistant
Mara: “Building My Personal AI Research Assistant: What I’ve Done So Far” documents what happens once the local stack is running — specifically, putting Hermes to work on actual academic admin.
Pip: Which turns out to be a lot of filing, cross-referencing, and turning brain-dump notes into something useful.
Mara: The daily note workflow is a good example of the scope. A cron job runs at 3pm, reads the day’s Obsidian notes, and surfaces connections: “You mentioned a meeting with Karl — should this link to the Karl-Sauropod grant note?” The goal is turning sparse bullet points into linked, searchable records without extra effort at the time of writing.
Pip: The grant cataloguing is the bigger lift — applications going back to 2010, pulled from folders, summarized, and cross-referenced into a structured repository. The next step is email integration so funding calls get automatically assessed against that history.
Mara: The honest summary from the post is that the parts working well are the data-processing tasks — cataloguing, extracting, organizing. The judgment calls, the conversational back-and-forth, are still rough. That’s a useful distinction for anyone thinking about a similar setup.
Pip: The throughline across all of this is the same: local-first, privacy-preserving, genuinely useful for the kind of iterative script-writing and research admin that fills an academic’s day.
Mara: And the treadmill piece suggests that the value isn’t really in any specific configuration — it’s in understanding how these systems fit together. That intuition compounds.
Pip: More to come as the email integration lands. We’ll be watching.
Prof. Peter L. Falkingham 
Leave a comment