AI Experts Disagree on How Hard It Is to Make AI Safe
This debate could decide whether AI helps or harms us—here's the simple version.
safety is described as largely pre-paradigmatic, with noticeable camps and wide variation, and as rapidly evolving. The article uses a fitness-landscape-like 3D surface to visualize these views, replacing the usual axes with variance between digital minds and the difficulty of reaching those minds using current methods, intersecting with a higher “alignment” space.
It condenses the perspectives into five main camps represented by Yann LeCun, Paul Christiano, Eliezer Yudkowsky, Daniel Kokotajlo, and Roman Yampolskiy, and uses P(doom) as a heuristic for parsing levels of concern. In the article’s framing, “alignment” loosely means any mind that would not cause extinction or an s-risk while also having substantial power; misalignment would lead to extinction or an s-risk.
The article describes one perspective as a global alignment attractor and an engineering problem, another as prosaic alignment needing resources to navigate harder-to-reach aligned places, and another as a steep hill that is hard to hill climb. It also says a cliff represents something basically non-navigable with current methods.
- Experts disagree on whether making AI safe is easy or nearly impossible.
- A 3D map helps visualize different views: some see gentle hills, others see steep cliffs.
- The stakes are high: AI safety affects healthcare, finance, and information you see online.
Why It Matters
AI safety debates shape how AI is built, affecting your privacy, safety, and trust in technology.