TL;DR
An AI training paradigm and alignment method where a less capable model is used to supervise, guide, and train a significantly more powerful model.
This paradigm relies on the principle that a stronger model possesses latent capabilities and representations that can be unlocked and guided by a weaker model acting as a teacher. During training, the weaker model provides soft constraints or labels that steer the larger model toward desired behaviors, preventing it from optimizing for local minima or harmful biases. Empirical research shows that the stronger model can correct the weak teacher's errors and successfully generalize to tasks beyond the teacher's capability.
Why this matters for your business
It offers a scalable method to align superhuman AI systems using existing, easily understandable human and machine supervision, solving a core bottleneck of the AI safety transition.
