๐ช๐ต๐ฎ๐ ๐ถ๐ณ ๐๐ฒโ๐๐ฒ ๐ฏ๐ฒ๐ฒ๐ป ๐ฐ๐ฎ๐๐ฒ๐ด๐ผ๐ฟ๐ถ๐๐ถ๐ป๐ด ๐๐ฒ๐ป๐ฒ๐ฟ๐ฎ๐น ๐ฃ๐๐ฟ๐ฝ๐ผ๐๐ฒ ๐๐ (๐๐ฃ๐๐) ๐๐ต๐ฒ ๐๐ฟ๐ผ๐ป๐ด ๐๐ฎ๐?
โข What if it isnโt just another โgeneral-purpose technologyโ like computing, but a bundle of many specific-purpose technologies out-of-the-box?
โข And what if calling AI a โcapital investmentโ is also a category errorโwhen in reality it behaves more like inexpensive labour?
Yesterday at the Cost-Benefit Analysis Forum organised by the Economic Society of Australia (NSW), I offered a small technical perspective on the long-running productivity paradox. Many leading economists have addressed it far more deeply; my contribution was simply to show how GPAIโs unique properties complicate familiar frameworksโfrom project-level costโbenefit analysis to firm productivity and national accounts. Unlike electricity or computing, GPAI:
โข delivers ๐ผ๐๐-๐ผ๐ณ-๐๐ต๐ฒ-๐ฏ๐ผ๐
๐๐ฝ๐ฒ๐ฐ๐ถ๐ณ๐ถ๐ฐ-๐ฝ๐๐ฟ๐ฝ๐ผ๐๐ฒ ๐ฐ๐ฎ๐ฝ๐ฎ๐ฏ๐ถ๐น๐ถ๐๐ and even “๐๐ป๐๐ฒ๐ป๐๐ถ๐ผ๐ป๐ ๐ผ๐ณ ๐บ๐ฒ๐๐ต๐ผ๐ฑ๐ ๐ผ๐ณ ๐ถ๐ป๐๐ฒ๐ป๐๐ถ๐ผ๐ป๐ (๐๐ ๐)”
โข drives ๐๐ต๐ฎ๐ฑ๐ผ๐ ๐ฎ๐ฑ๐ผ๐ฝ๐๐ถ๐ผ๐ป outside formal programs
โข ๐ฐ๐ผ๐น๐น๐ฎ๐ฝ๐๐ฒ๐ ๐ฒ๐ป๐๐ถ๐ฟ๐ฒ ๐๐ผ๐ฟ๐ธ๐ณ๐น๐ผ๐๐ into single and longer steps
โข blurs the line between labour and capital
โข Usage and inference costsโand even seemingly large capital investment in model trainingโmay be overestimated if divided by the many powerful specific-purpose capabilities
These shifts mean productivity gains are often misattributed, undercounted, or show up only as โmysteriousโ residual uplifts, total factor productivity, or multifactor productivity.
I also shared how CSIRO’s Data61 approaches this challenge:
โข Using impact evaluation frameworks designed for high-uncertainty R&D with delayed and diffuse benefits
โข Developing new risk assessment methodologies and a science of measurement tailored for GPAI
โข Applying AI itself to enhance costโbenefit analysis, from scoping and precedent extraction to tracing impacts, collating risk evidence, and synthesising reports
The day closed with a lively panel on whether AI can be trusted within costโbenefit analysis itself. The mix of economic, policy, and technical perspectives reinforced why I value multidisciplinary forumsโthey generate insights no single field can reach alone.
text-rich full slides https://gamma.app/docs/AI-Cost-Benefits-Analysis-CBA-bmch54uelzzzpy8
