Is alignment now the bottleneck for capabilities? Leo Gao debate ignites
JacquesThibs · x · 2026-10-01
A debate sparked by Jacques Thibs questions the classic orthogonality claim: alignment, historically seen as separate from capabilities, now appears to be a bottleneck for capabilities.
Leo Gao offers one possibility: weak AIs are bottlenecked by "alignment," after which the two increasingly become orthogonal. Thibs agrees with @ohabryka's worry that prosaic alignment and control techniques may just defer problems until ASI, with nothing that reliably scales past that point.
Related event: Alignment Researchers Debate Whether Current Techniques Scale to ASI(2 posts)→
More from AGI Musings
- Gary Marcus says Altman's UN AI warnings ring hollow, déjà vu from 2023 — GaryMarcus · 2026-10-01
- AI Race Dynamics Are Not Always Bad: Commercial Competition Beats Arms-Race Framing — jessi_cata · 2026-10-01
- Echoing Hassabis: Lab insiders see the singularity foothills, outsiders can't predict 3 months out — haider1 · 2026-10-01
- DeepMind researchers propose markets where AI scientists compete for compute and lab time — VraserX · 2026-10-01
- College Board redesigns AP CS courses to embed GenAI starting 2027-28 school year — lmoroney · 2026-10-01
- "AIs architecturally cannot have writer's block": why AI writing disappoints — QuintinPope5 · 2026-10-01