Interview
#31 - Prof Dafoe on defusing the political & economic risks posed by existing AI capabilities
- The Governance of AI program aims to build a career pipeline and community to direct talent to addressable problems in the near term, with expectations that the AI strategy research community will become substantially larger within approximately two years.
- Growth in interest in AI strategy is predicted to follow concerns about transformative AI, lagging the initial rise in AI safety interest by two years, while the median of expert survey opinions suggests a 10% probability of human-level machine intelligence arriving in 10 years and a 30% probability in 25 years.
- An arms race dynamic driven by the perception of an existing race among other sides is expected to increase the probability of competition occurring, leading to corner-cutting and reduced safety on a global scale.
- AI services arising from service digitization could trigger a reversion to economic nationalism, protectionism, and mercantilism, potentially ending the current liberal economic world order.
- Ubiquitous sensors and algorithms could be deployed to monitor the public, track behavior, and profile individuals to support political objectives, while imagery intelligence and social network analysis could reveal most submarine and mobile ballistic missiles currently forming the secure nuclear triad.
- Coupling advanced surveillance with hypersonic missiles to shorten decision-making time is expected to create a much more unstable world, and autonomous weapons could cause citizens to lose protection from human decency if major states possess the willingness to build them.
- Without further algorithmic improvements, today's AI technology could dramatically increase unemployment and inequality to extreme levels.
- The possibility of highly recursive self-improvement could generate rapid changes if narrow AI capabilities allow for self-improvement in AI research.
- Hardware overhang, data overhangs in the form of the internet and historical corpora, and insight overhangs of waiting-for discoveries could allow an AGI algorithm to run on many laptop equivalents, potentially transitioning the world from no AGIs to many millions.
- A training-to-execution computing cost ratio of one to 100 million times implies that repurposing training power for deployment could result in 10 million AGIs.
- Economists are expected to increasingly model transformative AI's impact on growth, unemployment, and labor displacement, while political scientists largely view the project as important and worth pursuing.
- Demand for personnel knowledgeable in this field is expected to outstrip supply due to historical constraints on training, though the process for training individuals and placing them in roles is expected to mature over time.
- Humanity is expected to overcome the challenge of advanced AI in the coming decades through cooperation, institution building, and the construction of identities that overlook past differences, with the gains from cooperation described as tremendous.
- The potential upside includes massive increases in wealth, scientific and intellectual insight, and human happiness, while historical progress in peace and liberalism is viewed as clear barring exposure to existential risks from advanced technologies.