Posts
What I've posted on LinkedIn about my work and research.
Every centre I looked at grew after the new metro opened in August 2024.
Chatswood and Macquarie Park rose 9.2% against centres off the line. CBD and North Sydney, where the new stations are, showed no measurable change. Then I faked the opening date 26 times to try to break it.
I gave five Claude models the same prompt: build an animated night scene of Sydney Harbour in a single HTML file.
One shot each, no edits, no retries. Two of the five didn't run at all. A headless browser caught both crashes before I'd looked at a single pixel, then each working one was checked against every requirement in the prompt.
Do you know which file your coding agent is editing right now?
Easy with one. Not when you've got three running in the same repo. So I built them a cursor: a free VS Code extension that shows where each agent is working, live, with edits that glow as they land.
How far can you trust an AI to mark other AIs?
120 hand-labelled pairs, three Claude judges, 4,200 judge calls. A rubric beat a plain prompt for every judge, 1-10 scoring was the worst format every time, and Opus favours its own answers.
I wrote a coding agent in about 300 lines of TypeScript, then put an off switch on every part of it.
Edit tool, test feedback, context trimming, notes, step budget: each off in turn, then all at once, across 1,120 runs for about A$14. None of them moved the pass rate. All 16 failures were the agent trusting its own check.
I wrote backpropagation from scratch, about 120 lines of TypeScript with no libraries, and put it on my website so you can watch it work.
A small neural network starts out guessing, then learns to separate two interleaved spirals right in your browser. It is one of the first projects on dylanlewis.au, alongside Commute Pulse and Aussie Data MCP.
Every so often you get to watch something you had a hand in actually ship.
I did a stretch of the development on Obsidia earlier this year, so it is good to see it out in the world. Huge credit to Eddie Lewis, who led the build.
2,229 Reddit posts. 388 subreddits. 12 months. Here's what's actually happening in AI hiring.
I pulled the data to see whether the noise on LinkedIn matched what people are actually saying when they are not performing for an audience. It does.
Most AI verifications today don't actually test anything.
The real risk to a business isn't whether someone can use Claude. It's whether they can tell when Claude is wrong.
Everyone is carrying an AI "driver's licence" they printed themselves.
AI skill is mostly self-reported. The real question for anyone hiring: can this person actually use AI well in a business context?