
The very welcome @DarioAmodei post calling for an AI slowdown touches on some of the concepts I referred to a couple of days ago. "Recursive self improvement" is perhaps the AI capability we need to be the most careful about, because it could lead to a superintelligence explosion. If that happens without "alignment", it is an existential threat. And it is tough to judge "alignment" without interpretability - ie the ability to monitor what the AI is thinking. He has clearly identified the problem. Now let's hope @openai and @GoogleDeepMind join him in pushing for a solution. For more on these concepts do watch this. 👇 youtu.be/GlDinj6E4Js?si=GOBk…




















