AI: From Disappointing Past Performance to the Coming Productivity Boom

Summary:

• Tech’s impact on productivity & growth

  • Quality of life has risen dramatically since 1700s

  • Impact most driven by General Purpose Technologies (e.g. steam engine, AI)

  • Power of AI has grown dramatically in doing tasks cheaper and/or better than humans

  • However, productivity growth is dropping

    • In 1990’s it was 2.8%/year

    • Since 2005 it has been 1.3%/year

    • Productivity = GDP/hours worked

  • Explanations:

    • Recent tech boom is a false hope

    • Mismeasurement of productivity

      • Many digital goods have $0 nominal price even if they are very valuable

      • GDP measures creation of physical goods

        • E.g. Wikipedia much more valuable than Encyclopedia Britannica but is valued as less because it is mostly free

      • Erik Brynjolfson currently building a new measure of well-being: GDP-B

        • Survey: how much would you pay to avoid losing a digital good/service?

        • Alternatives:

          • GDP: very precise but measures wrong thing

          • Happiness/wellbeing surveys: very noisy, varies strongly across societies with different understanding of good living

    • Benefits are real but concentrated among the powerful 

    • Benefits are lagging due to restructuring

      • Has happened before:

        • It took 30-40 years for introduction of electric motors to propagate to improvements in productivity

        • Steam engine-driven factories were physically centralized around single large engine

        • Electric motors were integrated into individual devices, enabling more physically distributed and flexible factories

      • AI/Software Systems

        • Effective use requires changes to business processes

        • Usually 90% of investment is transitioning the business to new processes

        • Correlating the value of physical goods owned by companies relative to their market values shows that there is a large “extra value” multiplier associated with IT capital, and a smaller one for physical capital.

        • Investment in technology varies a lot across companies

          • Few “superstar” companies investing vastly more than rest of economy

      • More accurate to model impact of tech as 

        • An initial drop in GDP while tech investments are being made

        • Followed by a rise once tech has been incorporated into the business

• Tech’s impact on society

  • Tech affects wages via

    • Substitution

    • Complementarities

    • Demand, Income & Supply Elasticity

    • New Tasks via invention & transformation

  • They categorized O*Net tasks by more or less automatable

    • Plotted automatability by income, geography & companies

  • Job2vec: ML embedding of job postings to understand how they’re related

  • Can use techniques to measure jobs based on their “remotability”