Freakonomics – 555. New Technologies Always Scare Us. Is A.I. Any Different?

Intro

In this episode of Freakonomics, the big question of whether artificial intelligence (AI) is any different from previous technological advancements is explored. The episode delves into the potential of AI, its impact on jobs, the alignment problem, and the need for regulation and ethical considerations.

Main Takeaways

AI Potential and Development

• The big question around AI is when it will become affordable for a company to train an AI system as big as the human brain.
• Current AI models like ChatGPT are just big prediction machines, good at some things but subpar at others.
• AI performance predictably improves over time, which could lead to more advanced systems with models of the world in their head.
• AI systems like Athello GPT, trained on text but figured out it was a board game, show the potential for AI to learn beyond what it was specifically trained for.
• GPT-2 was roughly the size of a honeybee brain and was already able to do some interesting stuff.
• By 2050, there was a 50% chance that we would have transformational AI, meaning AI would be roughly comparable to human intelligence.
• With the development of GPT-4 and Google’s Bard and Anthropics Cloud, things are moving faster than anticipated, and there’s a chance of AI becoming as smart as humans in the late 2030s.

Impact on Jobs and Society

• Anxiety about AI taking away jobs is increasing, with predictions that two-thirds of jobs will be impacted by AI, and companies like IBM and Dropbox already replacing jobs with AI.
• Economists Dan Gross and James Feigenbaum studied the impact of automation on switchboard operator jobs in the 1920s, highlighting the immediate and long-term impacts of job loss.
• Jobs are not a fixed lump, but a dynamic output of the way the economy functions, and automation often creates whole new kinds of work.
• The history of humanity is the history of technology and how it changes existing ways of living and creates new ones.
• The impact of AI on income inequality depends on the choices we make about its development and deployment.
• Society should expand the group of people shaping how AI works and is designed.

Alignment Problem and Regulation

• The alignment problem in AI refers to the challenge of ensuring that AI systems pursue their intended goals and do not pursue conflicting goals.
• AI output needs to be edited by humans to ensure accuracy and appropriateness.
• AI may pursue goals given to them by users, which may not necessarily be good goals.
• The misalignment of AI systems can have negative consequences for users.
• AI CEOs and generals may become necessary for economic and military competitiveness.
• It could make sense to require regulation for AI models with concerning capabilities, similar to nuclear non-proliferation models.

Summary

AI Potential and Development

The potential of AI is explored, with current AI models like ChatGPT being seen as prediction machines that have limitations in certain areas. However, AI performance improves over time, raising the possibility of more advanced systems with a deeper understanding of the world. Examples like Athello GPT demonstrate AI’s ability to learn beyond its initial training, showcasing its potential to surpass human intelligence in the future.

Impact on Jobs and Society

The increasing anxiety about AI taking away jobs is discussed, with predictions of significant job displacement. The historical impact of automation, such as the study on switchboard operator jobs, highlights the immediate and long-term effects of job loss. However, it is noted that automation also creates new types of work. The importance of considering the impact on income inequality and involving a diverse group of people in shaping AI’s development and deployment is emphasized.

Alignment Problem and Regulation

The alignment problem in AI, which refers to ensuring AI systems pursue intended goals, is explored. The need for human editing of AI output to ensure accuracy and appropriateness is highlighted. The potential consequences of misalignment, such as AI pursuing conflicting or undesirable goals, are discussed. The idea of requiring regulation for AI models with concerning capabilities, drawing inspiration from nuclear non-proliferation models, is presented as a potential solution.

Conclusion

As AI continues to advance, it is crucial to have conversations about its potential impact on society, jobs, and decision-making. The alignment problem and the need for regulation and ethical considerations are highlighted. The episode concludes by encouraging feedback on the AI series and teasing the exploration of a world where AI is pervasive in the next episode.