Generative AI impact on sustainability

Which analog?

Thermodynamic analog 1: Compression algorithm

JPEG compression

Compression algorithms already work as classifiers

Thermodynamic analog 2: A cup of coffee

Implications of AI as a hot compression algorithm

Compression smooths over outliers

Before training

After training

Heat as generative power

Adding heat lets compression algorithms generate new content

⚔️ Novel content can be a bug as well as a feature

Pre-AI archival trajectory

Post-AI archival trajectory

"AI is an archive in reverse"—Eryk Salvaggio

Archive in reverse

If we continue to build systems that consume provenance and output plausibility we are engineering the decay of knowledge—Jessica Talisman

Image generation in place of search?

Leonardo.ai Realtime Generation

Replacing the entire Internet?

Websim.ai

AI-generated artist websites

(Dis)connected knowledge?

The sycophantic web

Unlike sci-fi's cybernetic automata, transformer architectures are neither neutral nor cold. (I Do Not Think It Means) What You Think It Means

Implications for sustainability

For a compression algorithm to generate new variations on existing content incurs a thermodynamic cost that is both statistical and environmental.

GPU's get crazy hot

Inference at scale keeps data centers humming, but inescapable inefficiencies in those GPU racks radiate heat, which demands more water and power

Conclusion

If we don't confront AI's lack of sustainability on both fronts, we risk a web that may be easier to query but will be harder to trust, and far more costly to our culture and planet

Back to title screen