Part 8/10:
- When high-quality, nuanced training data are needed, he plans to switch to Davinci, the most sophisticated and resource-intensive engine.
Additionally, he discusses adjusting the temperature setting, with a lower value (like 0.5) resulting in more deterministic and focused outputs, whereas a higher value (like 0.7) introduces more randomness or eccentricity.
Advancing Toward Ethical, Morally-Aware AGI
By training GPT-3 to evaluate actions based on their alignment with core moral principles, Shapiro envisions an AGI capable of moral reasoning. For example, he demonstrates how the system can flag an ethically problematic action—such as proposing to euthanize all in pain—and reject it as inconsistent with the fundamental objective of reducing suffering.