"The Cognitive Revolution" | AI Builders, Researchers, and Live Player Analysis Inference Scaling, Alignment Faking, Deal Making? Frontier Research with Ryan Greenblatt of Redwood Research
81 snips
Feb 20, 2025 
Ryan Greenblatt, Chief Scientist at Redwood Research, dives into the complex world of AI safety and alignment. He discusses alignment faking and innovative strategies for ensuring AI compliance, including negotiation techniques. The conversation addresses the balancing act between AI progress and safety, emphasizing the need for transparency and ethical considerations. Ryan stresses the importance of international cooperation for effective AI governance, highlighting the potential risks of advancing technology without proper alignment with human values.
AI Snips
Chapters
Transcript
Episode notes
Explicit Reasoning and Clear Instructions
- To improve language model performance, make reasoning explicit and demonstrate the desired thought process.
- Use clear, concise instructions, and consider simplifying them for better understanding.
Shifting Importance of Demonstrations
- Demonstrating reasoning may become less important as models improve and learn to reason autonomously.
- Reinforcement learning can teach models effective reasoning strategies beyond human demonstrations.
Inference Scaling Laws
- Ryan Greenblatt used up to 5,000 generations per prompt in his work on the ARC-AGI challenge, revealing scaling laws.
- These scaling laws showed a relationship between increased generations and improved accuracy.