Right-sized AI and sustainability on the Web
Choose model architecture proportionate to tasks and devices.
Problem first, model second
Not every feature needs a large generative model. Conventional search, explicit rules and specialized models can solve many tasks with lower complexity and cost.
Comparable measures
Compare quality, response time, energy, bandwidth, model size and actual costs. Parameter count alone does not summarize footprint or performance across devices.
Local, hybrid or remote
Local architecture can reduce some data transfers but require substantial downloads and capable hardware. Remote architecture centralizes updates while introducing transmission and privacy trade-offs.
Honest experimentation
Benchmark a defined task on realistic scenarios and multiple devices. Publish samples, versions and failures rather than cherry-picking favorable results.
Choose the lightest sufficient approach
For sorting tasks, deterministic rules or small specialized models may suffice where a large generative model adds complexity and cost. Compare downloads, memory, response time and network use. Local execution may reduce transfers while requiring more capable user devices.
Exercise: compare three architectures
Define a simple document classification task and test a rule, a small local model and a hosted service if authorized. Measure accuracy on the same examples, latency, data transfer and deployment constraints. Report contextual results without inferring full environmental impact from request duration alone.
Core steps
- 01Define the target task
- 02Compare simpler alternatives
- 03Measure across devices
- 04Publish trade-offs
Three questions to ask
Does this task require a generative model?
Are quality and resource use both measured?
Is the fallback accessible?
Read the source documents
Versions, reuse terms and document status can change. Check them on the original sites.