"The Cognitive Revolution" | AI Builders, Researchers, and Live Player Analysis 2028 AGI, New Architectures, and Aligning Superhuman Models with Shane Legg, Deepmind Founder, on The Dwarkesh Podcast
32 snips
Jan 4, 2024 
Shane Legg, Co-founder and Chief AGI Scientist at DeepMind, delves into the timeline for Artificial General Intelligence (AGI) and the new architectures required to achieve it. He discusses the challenges of measuring AI progress, particularly regarding human-like episodic memory. The conversation highlights DeepMind's diverse research initiatives, the importance of ethical reasoning in AI alignment, and a future shift toward multimodal capabilities, revolutionizing how AI interacts with different types of data.
AI Snips
Chapters
Transcript
Episode notes
Architectural Solutions for AI Memory
- Architectural changes, not just scale, are key to solving episodic memory issues in deep learning models.
- The brain's separate mechanisms for rapid and slow learning suggest architectural solutions for AI.
DeepMind's Diverse Projects
- DeepMind's AlphaFold, while significant, doesn't directly contribute to AGI development.
- DeepMind pursues diverse projects, including fusion, sustainability, and deforestation monitoring.
Foundation Models and AGI
- Powerful foundation models, like LLMs, act as effective sequence predictors, compressing world data.
- Extending these models with search and reinforcement learning could unlock powerful AGI systems.