The strongest version
Not “language models will keep getting better with compute.” Something narrower and harder to dismiss: every capability previously believed to require specialized architecture — vision, planning, reasoning, code synthesis, multi-step problem solving — has in turn fallen to large-scale pretraining of a single generic substrate, and there is no principled reason to expect the pattern to stop.
Rich Sutton named the pattern in The Bitter Lesson (2019): across seventy years of AI research, methods that leverage computation eventually dominate methods that encode human priors. Chess fell to search at scale. Go fell to self-play. Vision fell to ImageNet plus deep nets. Language fell to next-token prediction. Scaling-sufficient says the AGI story is the same story.
Who holds it
Dario Amodei is the most public sustained defender — “scale plus the right kinds of post-training fills in the rest.” Ilya Sutskever, pre-2023, was the canonical “felt the AGI” voice; his later shift toward a more architectural framing is itself evidence about how the theory has aged.
The 2026 stress test
The claim predicts the shape of progress: smooth, compute-driven, surprisingly general, orthogonal to specific architectures. The stress test is whether the axes that have moved least — memory, embodiment, frame construction — start moving from scale alone, or only from added structure.
Falsifiers
A sustained plateau on a capability axis that added compute demonstrably fails to move, while a structural change does, is direct evidence against. Health is decelerating: recent gains have leaned increasingly on RL and scaffolding, not raw scale.