By James Pethokoukis

It’s always good to have concrete examples of technological progress, especially since macroeconomic stats aren’t great at capturing the phenomenon. Enter: The Stanford Institute for Human-Centered Artificial Intelligence. It recently released its 2022 AI Index Report, which surveys the current landscape of AI research and industrialization. Among the stats tracked in this data-driven report is the cost to train AI systems like ImageNet, an application for identifying and categorizing visual objects.

The report finds that training an image-classification system cost only $4.60 in 2021, compared to over $1,000 for a similar system in 2017: “In four short years, image classification training costs have decreased by a factor of 223.”

(Note the log scale on the Y-axis in the chart below. The apparent leveling off in cost declines since 2018 is actually steady progress.)

And the decline in costs goes beyond image recognition: “The trend of lower training cost but faster training time appears across other MLPerf task categories such as recommendation, object detection and language processing, and favors the more widespread commercial adoption of AI technologies.”

What might the commercial adoption of this technology look like? In a recent two–part series for a16z, Niko Grupen, a former machine learning engineer at Apple, speculates on such applications as smarter search engines that can “summarize any passage of text and extract key statistics” and personalized education that adapts to student learning needs. Another possibility is an AI writing assistant that enables us to modify the tone and style of a sentence the way we switch between fonts today. And with the cost to train AI systems falling, more and more applications like these are becoming commercially viable.

The post Chart of the day: Declining AI training costs appeared first on American Enterprise Institute – AEI.