normative

Lock Down Values To Constrain AGI

The key question is whether we can lock down a set of core values in AI systems so they are precluded from contemplating things at odds with those values; if we cannot lock down any values, we have a problem either from spontaneous misalignment or from abuse by bad actors—and the paperclip-maximizer example seems silly because you should be able to explicitly represent the value of human life so it gets ruled out as inconsistent with the axioms, unless the system is somehow not subject to those axioms.

normativepending

Speaker

Gary Marcus

Evidence Quote

“it's really a question about innateness values and whether we can do that”

Source

312 - The Trouble With Ai— Making Sense with Sam Harris
Created: 6/14/2026, 2:16:45 AM

Your notes

Loading notes...