I’ve been an avid user of xbox cloud streaming, the service was initially flawless on my Aussie HFC connection, but I noticed that in every game session the stream would freeze and pause for up to 20s before suddenly snapping back to streaming fine.

Not having extensive network experience, I thought this would be a fun excuse to throw an agent at it. This was all done with open weights models, mostly glm-5.3-flash and deepseek-v4-flash via opencode go.

This post explores how to build an agent using an on-device model such as Phi-4-mini or Gemini Nano (available in Edge or Chrome browsers respectively), all behind a simple flag. This approach lowers the barrier to entry since it requires no external API or service.

You can read the original, How to Build an Agent.

I urge you to follow along. The barrier is so low, it’s less than 200 lines of code, just pop open your browser dev tools console and you can follow right along as you read this!

On June 12th, 2025 I gave a talk at Web Directions Code conference about the rise of on-device AI and how rapidly the capabilities are moving, and how capable they are today.

I recently gave a talk at a local meetup, MelbJS.

Talk was about my explorations with the Google Chrome Built-in AI early preview.

I’ve been very interested in the potential of view transitions since they were called navigation transitions. Recently, they’ve landed in Chrome stable v126 with a more stable API. Initially it was a SPA-only API but has now been opened up to MPA, first with an experimental meta tag and now via a CSS at-rule.

A few weeks ago Charlie Holtz (@charliebholtz) showed a pretty cool demo using GPT4’s multimodal abilities to caption and narrate a screenshot of you which was then passed to ElevenLabs to speak out in the voice of David Attenborough with their cloning service.

This demo required you to run the code on your machine using python. I wanted to see if it could be done in the browser since all the generative AI capabilities were done through APIs.

All that was needed was the ability to capture a stream from a users webcam and take a snapshot at required intervals. This is fairly trivial these days in the browser.

We have a large hole, 65mm, cut into our gutter rainbox to allow it to overflow in the case of a large downpour. Rather than have a hole there I worked with my daughter on choosing something fun to cover it while still allowing it to overflow.

After deliberating over a unicorn vs trex vs shark we settled on the low-poly shark head.

Check out sharkbox.

Another year another timelapse of earth as seen from 5 satellites!

I’ve been running this little twitter bot for over 5 years now and it never gets old watching this beautiful pale blue dot.

I’ve previously wrote about how I do the yearly timelapse.

Earth in 2022. A year in review 🧵

— Earth (@earthin24) January 16, 2023

If twitter isn’t your thing it’s also on mastodon! https://botsin.space/@earthin24/109696834017584339

I recently wrote something up on the work engineering publication on a production use of the Proxy & Reflect objects to hook into the createElementweb api and add a CSP nonce to any style tags created that way.

📝 Patching builtin Web APIs using Proxy and Reflect.

The Home app has some frustrating limitations around creating automations or triggering events that don’t require auth from the user.

Apps like Eve allow you to hook into more HomeKit APIs and create automations that can react to more sensors.

What’s missing though is awareness of the automation suite as a whole e.g. overrides, priorities, and general awareness of how a bunch of automations could be modelled as a whole system.