Visualizing the Difficulty of AI Alignment

The author uses a 3D fitness landscape visual representation to explore and categorize varying perspectives on the difficulty of AI alignment into five main camps, ranging from Yann LeCun's view of alignment by default to Roman Yampolskiy's view of impossibility. The author argues that recent progress and capability speedups suggest alignment is increasingly difficult, potentially harder than P vs. NP. Anthropic's safety perspectives and Chris Olah's graphs are also discussed to frame the current trajectory of alignment research. Open article →