Superintelligence will be the most impactful technology humanity has ever invented, but it could also be dangerous. To solve the problem of superintelligence alignment, were starting a new team and dedicating 20% of our compute to this effort. We need excellent ML researchers and engineers to join us in building a human-level automated alignment researcher. Our approach involves developing scalable training methods, validating the resulting model, and stress testing our alignment pipeline. We expect our research priorities to evolve as we learn more about the problem.
