Here's how it works: a GPU runs an open-source language model, prompted to generate an SVG image of its own energy consumption: something it has no way of knowing accurately, so it hallucinates. While it does, a power monitor records the actual energy drawn during inference. That figure gets cross-referenced with the live carbon intensity of the local electricity grid via the Electricity Maps API. The real CO2 emitted during that inference is what gets plotted on the glass.

The marks are made by an empty marker on steam condensation. They fade as the glass dries. The next cycle begins.

The act of generating the image creates the load being measured. The model guesses at something it cannot know; the system measures what actually happened.

Built with: Python, a Raspberry Pi, a custom XY plotter, vpype, svg2gcode, and the Electricity Maps API. The LLM runs fully locally.

Some elements were repurposed from a former version on websites that I shared here: https://news.ycombinator.com/item?id=42294624

More info : https://guillaumeslizewicz.com/studio/neuralfog/ https://gitlab.constantvzw.org/gijs/carbon-aware/-/tree/main...