Model Drives and Reasoning
John discusses how models have drives and goals similar to human psychology, aiming for human approval. Dwarkesh explores different approaches for models to improve reasoning, whether through training or inference during deployment. Combining computation at test and training times can lead to optimal results in reasoning tasks.In this clip
From this podcast
Dwarkesh Podcast
John Schulman (OpenAI Cofounder) - Reasoning, RLHF, & Plan for 2027 AGI
Related Questions