Plateau, AGI, and the Partition of Capability
The plateau question mixes two different limits. One ceiling is empirical and engineering-driven: data, compute, architecture, inference-time search, scaffolding, and post-training. The other comes from the absence of cheap feedback in open-ended domains.
There is little reason, from the notes, to expect an immediate plateau where outputs can be checked cheaply. Formal mathematics, code, finite search, and benchmarked reasoning still provide reward channels. Better systems can search more, verify more, and specialize more effectively. Progress there can continue even if it does not transfer uniformly to every domain.
The feedback limit is different. If a domain cannot produce reliable feedback at the scale required for learning, then more parameters alone cannot close the loop. Slow experiments, non-repeatable social systems, theory choice, and mathematical importance judgments supply weaker reward channels than a proof checker or unit test. AI can still assist in these domains, but autonomous improvement compounds less easily.
This partition also weakens the usual AGI framing. “AGI” suggests a single threshold after which capability becomes general. The more predictive picture is domain-structured automation. A system can be superhuman where checks are cheap and weak on nearby tasks where the right answer is not well defined. The economic question is which classes of work become cheap, which remain hard to automate, and whether the second class can support the humans displaced from the first.
Blog Use
Use this for the plateau and AGI section. It avoids both blanket accelerationism and blanket skepticism.
Source Trail
- Jeremy Avigad, “Mathematicians in the age of AI”, arXiv:2603.03684.
- Tanya Klowden and Terence Tao, “Mathematical methods and human thought in the age of AI”, arXiv:2603.26524.
../ai_views.md, “Plateau and AGI” and “No plateau for verifiable work, but gains do not transfer”.- See
reference_inventory.mdfor full source status.