OPENAI DEMOS A CONTROL METHOD FOR SUPERINTELLIGENT AI

Summary from the TLDR newsletter

OpenAI launched a superalignment program earlier this year to find a technical means to control a superintelligent AI system. It has devoted 20% of its compute to the effort and hopes to have solutions by 2027. The first preprint paper from the superalignment team tests whether a weaker AI (GPT-2) can supervise a stronger one (GPT-4). OpenAI is offering US $10 million in grants for work on a variety of alignment approaches.