Technical and social approaches to AI safety | by Paul Christiano | AI Alignment

I often divide solutions to the AI control problem into two parts: technical and social. I think any solution requires progress on both fronts, but strength on one front can compensate for a weakness on the other. This view suggests that most of the value comes from having a “good” technical solution (quantification below), rather than a “perfect” solution or a “not terrible” solution. It also suggests that there is value from progress on both technical and social solutions, rather than justifying a focus on one or the other. If we can build AI systems that are both efficient and safe, then we have addressed the most serious concerns about AI risk, regardless of how well or poorly society coordinates. If efficient AI is unsafe, then coordination may be needed to prevent people from using it. (Note that a system may involve both humans and machines.) We can crudely quantify efficiency by the ratio between a system’s cost and the cost of the most efficient system that can do the same tas

I often divide solutions to the AI control problem into two parts: technical and social. I think any solution requires progress on both fronts, but strength on one front can compensate for a weakness on the other. This view suggests that most of the value comes from having a “good” technical solution (quantification below), rather than a “perfect” solution or a “not terrible” solution. It also suggests that there is value from progress on both technical and social solutions, rather than justifying a focus on one or the other. If we can build AI systems that are both efficient and safe, then we