tszzl challenges doomer shibboleth: aligning superintelligence isn't a one-shot problem unlike today's LLMs

tszzl · x · 2026-11-21

A notable AI-safety debate: tszzl argues that doomers cling to a shibboleth — the idea that aligning superintelligence is a one-shot problem unlike anything in aligning today's models — which he says contradicts evidence accumulated during the LLM era. In the quoted context, repligate argues that developments of the past few years are important updates on x-risk, not all positive but many are; if there's always a reason not to update, you should be suspicious of yourself, and watching relevant reality unfold is an epistemic privilege. The exchange captures a core disagreement over whether today's alignment experience extrapolates to superintelligence.

Original post →

More from AGI Musings

AGI Musings channel →