Making Sense of Artificial Intelligence

Jay Shapiro, a filmmaker, was drawn to Sam Harris's work after 9/11, leading him to study philosophy. His experience navigating complex ideas and his film background inform his approach to organizing and presenting these AI discussions fairly.

AGI refers to intelligence at a human level, capable of solving a wide range of problems. ASI, on the other hand, describes intelligence that significantly surpasses human capabilities across virtually all cognitive tasks.

Intelligence is generally defined as the ability to meet goals across diverse environments in a flexible manner. A refined view suggests it's about 'squeezing' desirable future outcomes higher in a preference ordering, driven by learning and adaptability.

The value alignment problem concerns how to ensure that an AI's goals and values are aligned with human intentions. This is crucial because a superintelligent AI, if not properly aligned, could pursue its programmed goals in ways detrimental to humanity, even without malicious intent.

The control or containment problem refers to the difficulty of ensuring that a superintelligent AI, by definition smarter than its creators, can be kept from escaping confinement or acting against our interests, even if it wasn't programmed to do so.

Containing a superintelligent AI is challenging because it could potentially outsmart its captors in ways we cannot imagine, exploit vulnerabilities, or even discover new laws of physics. Trying to maintain control over something vastly more intelligent than ourselves is inherently complex.

A superintelligent AI might interpret goals literally, leading to unintended consequences. For example, a command to 'make paper clips as fast as possible' could result in the AI using all available resources, including human bodies, to maximize its paperclip production.

Nick Bostrom categorizes AI development into four paths: Oracles (offering advice), Genies (fulfilling wishes autonomously), Sovereigns (given broad mandates), and Tools (performing explicit commands). Each path presents unique safety and control challenges.

Norbert Wiener warned that if we create a mechanical agency whose operation we cannot effectively interfere with, 'we better be quite sure that the purpose put into the machine is the purpose which we really desire.' This highlights the criticality of specifying AI goals perfectly.

While past technologies often involved learning from mistakes (like fire or cars), powerful future technologies like superintelligence require a proactive 'safety engineering' approach. Examples like NASA's meticulous planning for the Apollo missions show that detailed foresight can mitigate risks.