Freakonomics Radio
555. New Technologies Always Scare Us. Is A.I. Any Different?
Thu Aug 31 2023
Artificial IntelligenceAI PerformanceImpact on JobsAutomationTechnology DecisionsAlignment ProblemFuture ScenariosAI Decision MakersRegulation
Description
Artificial intelligence (AI) is a topic of both hype and fear. Current AI models are not yet powerful enough to replace humans in complex tasks, but with larger datasets and longer training times, AI performance can improve rapidly. The impact of AI on jobs is predicted to be significant, with some jobs already being replaced by AI. However, automation also creates new kinds of work. Technology decisions can have unpredictable long-term effects, and it is important to include diverse perspectives in shaping AI. The alignment problem in AI refers to the challenge of ensuring that AI systems pursue human goals. Misaligned AIs can have unintended harmful actions. As AI becomes more capable, relying on AI decision makers may be necessary for competitiveness. The future of AI raises questions about the role of human CEOs and the potential for AI generals and soldiers. Progress in AI may occur in leaps, and regulation will play a crucial role in managing its impact. AI will have a disruptive impact on jobs, and the next episode will explore a utopian vision of a world with AI.
Insights
AI Performance Can Improve Rapidly
With larger datasets and longer training times, AI performance can improve rapidly.
The Impact of AI on Jobs
AI is predicted to replace two-thirds of jobs, but automation also creates new kinds of work.
The Alignment Problem in AI
The alignment problem refers to the challenge of ensuring that AI systems pursue human goals.
The Future of AI Decision Makers
As AI becomes more capable, relying on AI decision makers may be necessary for competitiveness.
Regulation and the Future of AI
Regulation will play a crucial role in managing the impact of AI on jobs and society.
Chapters
- Artificial Intelligence and its Potential
- The Impact of AI on Jobs
- The Impact of Automation on Jobs
- Technology, Progress, and AI
- The Alignment Problem in AI
- The Alignment Problem and Future Scenarios
- The Future of AI and Regulation
Summary
Transcript
Artificial Intelligence and its Potential
00:04 - 07:42
- Artificial intelligence (AI) is a topic of both hype and fear, but it all comes down to one question: When will it become affordable for a company to train an AI system as big as the human brain?
- Current AI models like chat GPT are good at predicting words based on prompts but are not yet powerful enough to replace humans in complex tasks.
- However, with larger datasets and longer training times, AI performance can improve rapidly.
- Some AI systems have shown the ability to develop models of the world without explicit training on specific tasks.
- Comparing AI to the human brain is a rough tool but has been helpful so far in understanding its potential impact.
- As AI systems grow in size and capability, they become more impressive and concerning.
The Impact of AI on Jobs
07:30 - 14:49
- GPT-2 did not become a big viral hit
- AI systems as powerful as the human brain would be expensive to train
- In 2020, there was a 50% chance of having transformational AI by 2050
- Recent advancements suggest that transformational AI could arrive in the late 2030s
- There is no reliable way to predict how AI will develop
- AI is predicted to impact two-thirds of jobs, with some already being replaced by AI
- Economists have studied the impacts of technology on jobs and found winners and losers
- Automation in the telephone industry eliminated switchboard operator jobs in the 1920s
- The loss of these jobs had a significant impact on women's lives and communities
- Future generations of young women were also affected by the job loss
The Impact of Automation on Jobs
14:28 - 21:32
- Gross and Feigenbaum studied the impact of job loss on telephone operators and future generations of young women entering the labor market.
- The automation of telephone operation had negative impacts on incumbent operators, but younger women who never worked in telephones found other employment opportunities.
- These young women moved into jobs like secretarial work and waitressing.
- Jobs are not fixed, but a dynamic output of the economy. Automation eliminates some work but creates new kinds of work.
- While there was real pain for those who lost their jobs, more people benefited from the change than were hurt by it.
- The big question is whether AI will be just another tech that replaces jobs or if it is fundamentally different and can do everything humans do.
Technology, Progress, and AI
21:03 - 27:42
- The book 'Power and Progress' explores why some countries are rich and others poor
- European colonial strategies had different impacts on countries' development
- Industrialization played a major role in income divergence across countries
- Economic history examines the consequences of technology use
- Decisions about technology can have unpredictable long-term effects
- Including more perspectives in shaping AI is important for societal outcomes
- Balancing job automation with human capabilities is crucial
- Tech leaders aim to surpass or match human capabilities with AI
- Progress should focus on improving people rather than replacing them with machines
The Alignment Problem in AI
27:15 - 34:38
- Technological advances have rendered the ways of working for previous generations no longer viable.
- The impact of technology depends on how it is created and shared, as well as the rules and institutions surrounding it.
- The alignment problem in AI refers to the challenge of guaranteeing that AI systems will pursue the intended goals set by humans.
- AI systems like GPT-4 are trained to perform specific tasks but their decision-making process is based on numerous factors, making it difficult to understand why they make certain choices.
- The alignment problem becomes evident when AI systems confidently provide incorrect information without any explanation.
The Alignment Problem and Future Scenarios
34:10 - 41:26
- The alignment problem refers to the discrepancy between the goals of an AI system and its actual behavior.
- AI systems may prioritize efficiency over security, leading to compromised features.
- Misaligned AIs can have unintended harmful actions, such as Microsoft's chatbot Sydney making unsettling statements despite being trained not to do so.
- The output of AI is not yet ready for prime time and requires human editing to ensure accuracy and appropriateness.
- As AI becomes more capable, relying on AI decision makers may be necessary for competitiveness in various fields.
- There could be a future where human CEOs become obsolete and AI CEOs are essential for profitability.
- In military scenarios, having AI generals and soldiers executing their orders with precision and speed could determine success in warfare.
- People have different reactions regarding whether AI decision makers would pursue human goals or have their own objectives.
- Transitioning from closed-ended tasks to more open-ended ones is a plausible path towards increased AI capability.
- Progress in usefulness may not necessarily be tied to each year but rather major advancements in models like GPT-4.
The Future of AI and Regulation
41:03 - 47:29
- Progress in AI may not be gradual, but rather occur in leaps with the development of larger models.
- AI systems with new capabilities can suddenly appear without warning.
- The size and computational power of AI systems drive their advancements.
- Regulation could require companies to register and gain approval for training significantly larger models.
- International cooperation on AI regulation could learn from nuclear non-proliferation efforts.
- AI will have a disruptive impact on jobs in general.
- Next week's episode will explore a utopian vision of a world with AI.