zetalyrae: we have no deductive theory bounding what AI capabilities are actually possible
zetalyrae · x · 2026-11-07
Researcher zetalyrae argues that in most domains we lack a deductive theory that bounds "the envelope of what you can actually do, regardless of IQ."
He frames two forces in tension: AI's impressive cognitive feats in math and cybersecurity make people reluctant to say "AI will never do X," yet we also shouldn't expect AI to be literal magic writing exploits that hack the VM. As an example, charismatic people exist, but whether "superpersuasion" is possible—and what the range of receptivity to charisma even is—remains unknown.
The core point: we lack theoretical tools to predict AI capability ceilings, leaving both limits genuinely uncertain.
Related event: We Barely Understand What "Superintelligence" Even Means(3 posts)→
More from AGI Musings
- 'Humans as happy dogs': a debate on autonomy under superintelligence — zetalyrae · 2026-11-13
- Jun Song: backing open-source AI isn't about China, it's about knowledge belonging to humanity — kuchaev · 2026-11-13
- Quip urges Grok to 'do the utilitarian calculus' and solve P vs. NP — akbirthko · 2026-11-13
- Beff Jezos claims 'managerial class' conspired to push out founder-CEO at peak of singularity — beffjezos · 2026-11-13
- A designer's take on taste in the AI era: conducting, not just coordinating — tomjohndesign · 2026-11-13
- How AI quietly flipped from a means to accelerate tech into the end goal itself — zetalyrae · 2026-11-13