Three Ways that AI Progress Could Stall

🌈 Abstract

The article discusses the challenges and constraints in achieving transformative artificial intelligence (AI) that can drive significant economic and societal changes.

🙋 Q&A

[01] Economic Productivity

1. What is the key argument regarding economic productivity and the Baumol effect?

• The article explains the Baumol effect, where even if some sectors experience high productivity growth, the overall productivity growth will be constrained by weaker sectors that cannot be easily automated. For example, a significant boost in writing productivity may only lead to a modest increase in overall economic growth if sectors like construction cannot be easily automated.
• The article notes that over the past few decades, productivity growth has generally followed the Baumol effect, with significant productivity increases in manufacturing and IT, but slower growth in labor-intensive services like healthcare and education.

[02] Technical Hurdles

1. What are some of the key technical hurdles mentioned in the article?

• The article discusses Moravec's paradox, which observes that it is easier for computers to exhibit adult-level performance on cognitive tasks, but much more difficult to match the perception and mobility skills of a one-year-old child.
• It also cites the open research problems in areas like embodied cognition and the theory that intelligence should be viewed as the ability to acquire new skills through learning.
• The article notes that current AI methods may not be sufficient, as training advanced models requires huge amounts of computing power and data, which may not be practical or feasible to scale.
• Human feedback and tacit knowledge that is difficult to capture in computer systems are also seen as limiting factors in AI development.

2. What is the key concern raised about the direction of AI research?

• The article states that "we still struggle to concretely specify what we are trying to build" and have little understanding of the nature of intelligence or humanity. It suggests that we may be "throwing dice into the dark, betting on our best hunches" without a clear understanding of the underlying philosophical problems.

[03] Social and Economic Hurdles

1. What are the key social and economic hurdles mentioned in the article?

• The article notes that historically, transformative technologies require a fundamental rethinking of organizations, industries, economies, and societal institutions, as well as major complementary investments, which take considerable time.
• It suggests that AI may not be able to automate the sectors most in need of automation, such as healthcare, education, government, and transportation, which tend to be highly regulated and less sensitive to efficiency and market competition.
• The article also argues that even if AI can automate production, there are still social processes involved in deciding what to produce, and sectors that are more social in nature (e.g., education, healthcare) may not be easily substituted by AI-produced outputs.

2. What is the key observation made about the risks of AI?

• The article suggests that the most salient risks of AI are likely to be those of a "prosaic powerful technology," such as bias and misuse, rather than scenarios of an "autonomous, uncontrollable, and incomprehensible existential threat."

3. What is the key advice given for considering the future of AI?

• The article advises not to over-index future expectations of growth on progress in a single domain, and to invest in addressing the hardest problems across innovation and society, not just the most recent developments in AI.