| |
|
| |
Join the Snorkel AI Reading Group, a recurring forum to explore the latest frontier developments in AI while building meaningful connections within the community.
As AI workloads grow, data center infrastructure is strained by rising energy & capacity demands. Yet many queries are everyday tasks (drafting, summarizing, looking things up) that may not need frontier-scale models running in the cloud. Knowing how to measure efficiency brings us closer to building AI that is within everyone's reach without outgrowing the power we have to run it.
In this session, Stanford researchers Jon Saad-Falcon & Avanika Narayan will present their work on Intelligence per Watt (IPW): Measuring Intelligence Efficiency of Local AI, where IPW serves as a unified metric to measure intelligence efficiency by capturing capabilities delivered against power required.
Among other things, you'll learn:
How measuring intelligence per watt gives researchers, builders, & policymakers a common yardstick for deciding where AI should run as demand outpaces power supply
How efficiency measurements across models & hardware can show which queries belong where
Why weighting capability against real-world economic tasks reveals where AI creates value, & where gaps remain
Across 20+ models & 8 accelerators, tested on 1M real-world queries, how local AI efficiency has changed over time
Agenda:
4 pm - doors open
4:30 pm - talk begins
5:30 pm - research discussion & networking
Boba tea & other refreshments will be provided !
This work has been featured in the Financial Times, Reuters, & Foreign Affairs, & was recently accepted to NeurIPS 2026. arXiv preprint available here.
The broader research has also been featured through interviews, talks, & industry coverage from Foundation Capital, AMD's Advancing AI conference, PyTorch Conference's Measuring Intelligence Summit, IBM Research, & SambaNova.
Jon Saad-Falcon & Avanika Narayan are PhD students in Computer Science at Stanford University & members of Hazy Research.
|
|
|
|
|
|
|
|
