Shane Legg (DeepMind Founder) — 2028 AGI, superhuman alignment, new architectures

This podcast episode examines the intricate challenges in defining, developing, and ensuring ethical behavior in Artificial General Intelligence (AGI). Experts Shane Legg and Demis Hassabis delved into the difficulties of measuring AGI progress, addressing the limitations of current AI models in episodic memory and sample efficiency. They explored the challenges in assessing human-level AI and the complexities of defining intelligence, emphasizing the significance of search in unlocking creativity. They also discussed the importance of aligning AI systems with ethical values, emphasizing the need for deep understanding, robust reasoning, and ethical training. The discussion extended to the impact of organizations like DeepMind in accelerating AI capabilities and safety concerns, along with potential timelines for developing AI systems that surpass human capacities.