Eliezer Yudkowsky: What Is Artificial Intelligence?

Defining Artificial Intelligence

• 💡 The speaker finds it challenging to give an "airtight" definition of AI but offers powerful and useful definitions based on capabilities.
• 🎯 One key aspect is an AI's predictive ability, measured by how accurately it assigns probability to future observations and how good its "map of reality" is.
• 🔑 Another core definition revolves around an AI's steering ability, which refers to its effectiveness in achieving goals within a given environment, like winning a chess game.

The Concept of Generality

• 🧠 Generality is a crucial differentiator, referring to an entity's capacity to learn across many domains and integrate diverse knowledge.
• 🐙 Unlike an octopus, which might excel in specific tasks, humans possess a more general predictive and steering capability due to superior learning.

Evolution of AI Capabilities

• 🚀 Early AIs were characterized by narrow intelligence, excelling only in specific tasks like chess without broader learning capabilities.
• 🤖 Modern AIs, such as ChatGPT, demonstrate an ability to pull together and reason across various domains, like understanding the impact of infrared light on crops.
• 💡 These advanced systems can integrate different forms of knowledge, similar to how humans reason about concrete and water to build a dam.

Current AI Limitations and Future Trends

• ⚠️ Despite impressive capabilities, current AIs can still make mistakes that no human would make, suggesting they are not yet as smart as humans, even "dumber than 12-year-olds" in some contexts.
• 📈 The speaker emphasizes that the core components of AI—predictive ability, steering ability, and learning ability—are continuously improving, indicating that AI is "getting smarter."