The Jevons Paradox is when an increase in the efficiency of a resource leads to an increase in overall consumption. This happens when the elasticity of demand is sufficiently high. The classic example is as we learned how to convert coal into energy more efficiently, we consumed more coal overall. Consumers with fuel-efficient cars tend to travel more and therefore consume more fuel.
In software: video compression algorithms improved, which increased the demand for streaming so that overall more bits were transferred.
- Page load times improved, and people spent more time on websites.
- Hardware became more efficient, but our software consumes more resources than ever (e.g., Google Chrome).
- Cloud computing became more efficient: serverless, scale-to-zero, and function-as-a-service. These abstractions unlocked more use cases for cloud, and developer spend and utilize more cloud resources than ever.
Jevons Paradox is coming for LLMs. We already have parameter efficient training, but we’ll continue to make LLMs more efficient — optimizations in design, hardware, and software. Small improvements can lead to dramatic increases in usage.
What if an LLM ran locally on your phone? We might have one (or more) on every iPhone.
What if latency was lower? We could add LLMs in critical paths other than chat.
What if the cost were lower? We could use larger models with higher context lengths.
What if the developer experience was better? More developers could integrate LLMs into their applications.