Jamie Duncan

@jduncan.bsky.social

Tech & History Nerd. Wood worker. F1 fan. Dad. Commanders & Canes Fan. Living in The Triangle in NC. Currently @google. Former @vmware and @redhat

I asked two frontier models for a color-coded ASCII map of US cloud regions. I got a potato, then squares, then a Jackson Pollock, then a Mon Calamari cruiser. Turns out there's real research on this: models read ASCII fine and can't write it. The gap is 13 to 27 points.

Is ASCII Art the True Turing Test?

Two frontier models couldn't draw an ASCII map of the US. Why LLMs read ASCII art fine but can't write it, and what that says about the jagged frontier.

jduncan.io

I ripped the RAG-over-MCP layer out of my marketing agent and turned the whole repo into a wiki the model reads directly. It's turning out well. But once your main reader is a machine, the old wiki instincts start working against you: a model reads everything with equal weight, so a contradi…

Running a Wiki Whose Main Reader Is a Machine

I replaced a local RAG setup with an LLM wiki: a repo my marketing agent reads directly. The pattern that keeps a machine-first knowledge base true over time.

jduncan.io

I'm increasingly amazed at the apparent correlation between "rate the Super Bowl halftime show" and "who did you vote for". The work done to polarize and divide this country is masterclass in tearing down a country.

Trying to compare Kendrick Lamar last night to Prince's halftime show is like trying to compare one of Shakespeare's sonnets to Dr. King's Letter from Birmingham Jail. They weren't designed to serve the same purpose. They're both beautiful and incredibly effective.

It took Trump 3 hours to delete the Spanish version of the White House website and the White House’s Spanish Twitter account. Trump won 42% of the Hispanic vote.