YY's Random Walks: Preparedness paradox, research advice, and coding agents with open source models

Three things this time: why good prevention ends up looking wasteful, a compilation of research advice, and a recipe for running a coding agent on UVA’s hosted LLM, which may be useful for those who want to set up open source AI agents.

Preparedness paradox#

https://yyahn.com/wiki/Preparedness-paradox/

In July 2025, following budget cuts, the CDC made FoodNet reporting optional for Cyclospora and five other pathogens; a year later, the 2026 cyclosporiasis outbreak is unfolding. This is a familiar pattern: a “cycle of panic and neglect”. I put together a quick note on the “preparedness paradox”: effective prevention makes a danger look less serious, because we never experience the crisis that prevention averted. In other words, the better prevention works, the more wasteful it may look.

This happened again and again. A crisis makes the value of preparedness obvious, so governments build teams and fund surveillance. Once the crisis fades, successful prevention leaves little visible evidence that the investment worked, preparedness starts to look wasteful, and capacity erodes until the next crisis restarts the cycle.

Research advice#

https://yyahn.com/wiki/Research-advice/

I’ve been maintaining a compilation of research advice—“looks-useful-to-me” pieces I encountered over time, organized by topic and career stage, from applying to grad school all the way to tenure. It has classics like Hamming’s You and Your Research along with collections, books, and job-market resources. You may find something useful in there (there also may be some dead links and outdated advice).

Coding agents with hosted open source LLMs#

https://yyahn.com/wiki/Coding-agents-with-UVA-RC-GenAI/

UVA Research Computing runs RC GenAI, a university-hosted LLM service with an OpenAI-compatible API (currently serving Kimi K2.5). I think many universities are setting up similar services and you may be setting up your own locally as well, due to privacy concerns or other issues with big AI companies. If that’s the case, I thought that this tutorial may be useful.

The API is gated to HPC-internal traffic, so you can’t reach it directly even on the VPN. Here’s a recipe for using it with a coding agent from your laptop: tunnel through an HPC login node with sshuttle, then point Pi at the endpoint. Pi is one of the simplest and most customizable agent harnesses, and a single shell function (pi-uva) brings up the tunnel and launches the agent.

Coding-agent CLIs differ in which API they speak, and that matters more than it used to. RC GenAI exposes only OpenAI Chat Completions—not the newer OpenAI Responses API (which Codex CLI now requires) and not Anthropic’s Messages API (which Claude Code speaks). So the choice of agent CLI is constrained by the endpoint, not just by preference. There are ways to have a proxy that translates different APIs but that may be too much.

The recipe is UVA-specific, but the ingredients probably translate to other settings: a hosted OpenAI-compatible endpoint, a network gate you can tunnel through with sshuttle, and an agent CLI that matches the API the endpoint exposes.

That’s it for now!

yy