Part 8/15:
Shapiro discusses the concept of generating internal tension among various models—like a Generative Adversarial Network (GAN)—not by competition but by creating a healthy "tension" that encourages the system to evaluate different perspectives critically. He envisions multiple models, fine-tuned with various datasets (GPT-3, Bloom, NeoX, etc.), contributing different viewpoints and sometimes "challenging" each other to improve overall decision-making.
This internal "disagreement" fosters robustness and sophistication, allowing the AI to reflect on its performance and adapt dynamically.