The strongest version
Current transformer-based systems are missing organs. The strong version is not “transformers will plateau.” It is the specific claim that at least three architectural primitives — native long-term memory, learnable world models, and continual learning — are absent from current frontier systems, that the absences are not quantitative engineering problems, and that closing them requires structural changes that pretraining-plus-RL pipelines do not produce as side effects.
Who holds it
Demis Hassabis is the most public defender: the “75% of the way to AGI” claim is interesting for what it puts in the remaining 25% — not more compute, but components not yet built and integrated. Yann LeCun’s JEPA family is the academic exemplar, motivated by the claim that autoregressive next-token prediction cannot produce systems that learn the way humans and animals do.
The load-bearing assumption
That the missing pieces are structural — that no amount of scaling the current substrate produces persistent memory or a grounded world model as an emergent property.
Falsifiers
A frontier system that demonstrates durable, updatable memory and causal world-modeling without a bolted-on external component — as a genuine property of the trained network — would move health sharply down. So far, the gains have come from scaffolding, not from the substrate growing the organ.