WEAK TO STRONG GENERALIZATION

Summary from the TLDR AI newsletter

This new result from the OpenAI super alignment team (with code) suggests that you can use much weaker supervisor models to steer or align a far more powerful model, an analog to weak humans aligning super intelligent models. They used GPT-2 to recover much of the alignment performance of GPT-4. One important point they note is that this method has a tractable path for dramatic improvement, unlike RLHF-like methods.