LLM hallucinates its carbon footprint, we measure and plot it on steam [video]

by G_S·Mar 8, 2026·3 points·0 comments

Visit Project View on HN

AI Analysis

○Pass

Poetic installation, but 49 seconds of steam disappearing—no reproducible tool or insight.

Strengths

•Genuinely novel concept linking ephemeral medium (condensation) to invisible infrastructure cost
•Recursive conceptual depth (model hallucinating its own carbon footprint)

Weaknesses

•No tool, no methodology, no reproducibility—pure art performance without documentation or open-source output
•No measurable contribution to carbon accounting; plotter traces prove nothing about LLM efficiency

Post Description

Carbon Technostructure traces the material weight of immaterial systems. A modified plotter, glass panel, and steamer work in concert to render visible what typically remains unseen: the environmental cost of computational processes.

Steam condenses on glass, creating an ephemeral drawing surface. An empty marker moves across this temporary canvas, inscribing patterns derived from real-time energy data. These marks exist only as disruptions in condensation, fading as quickly as our attention to digital infrastructure’s environmental impact.

For this iteration, we run a GPU hosting an open-source language model. We prompt it to hallucinate representations of its own energy consumption—a recursive gesture that produces both poetic output and measurable load. The installation monitors this consumption, cross-referencing it with the carbon intensity of the local electrical grid and subsequently plots it.

Each vanishing drawing offers a glimpse of the carbon technostructure we’ve built, one calculation at a time.

Produced during Ohme/BrIAS residency in Brussels, Avec le soutien de la Fédération Wallonie-Bruxelles

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