Murmur

a toaster gives advice
the city is a mirror
the mirror is the joy

Murmur creates haikus with a small AI language model directly in the browser.

Built with Astro on Cloudflare, running Qwen3-0.6B in-browser via Transformers.js, with Stripe and Printful for the print-on-demand side.

Besides having fun with the branding details, designing a full experience, and playing with haiku formats, I had the goal to test LLMs in two ways.

First, how well central ideas translate to individual technical decisions with coding agents. This time I’ve mostly used Opus 4.5 and 4.6, current frontier models. While generated details are fine or sometimes actually good, the result just lacks the attention to connect execution decisions to the broader vision (with a few interesting exceptions like offering interesitng phrases in the copy). This is something I’d expect from a motivated human engineer to be more attentive about. For example the small details in the loading animation which initially was more generic and technical, needed fairly significant detailed instruction to more directly follow the mood of the flow. Attention to details and systematic choices still mostly means human attention.

My second goal was to see how capable an LLM could be in the browser, running on the user’s machine, in a realistic use scenario. The haikus are mostly good. The model hyperfixates on certain examples (like toasters appear constantly), but it’s surprisingly usable for a 0.6B model running client-side.

toaster hums in the shadow
while my hands are already on fire
the past has turned me into a pixel