Every few years the field re-litigates a question that should be settled by now: is intelligence learned statistics or structured symbol manipulation? Our answer is the unfashionable one — yes.
The critiques of pure deep learning have been on the record for years and have aged well: data hunger measured in orders of magnitude beyond any human learner, brittleness when the test distribution drifts from training, and a persistent shallowness of compositionality — systems that master 'red cube' and 'blue sphere' yet stumble over 'blue cube.' These are not bugs to be patched with scale. They are signatures of architectures that interpolate over a training manifold rather than operating over variables.
The critiques of pure symbolic systems aged equally well, in the opposite direction: hand-built structure is rigid, perception-blind, and helpless before the noisy, statistical texture of real data.
Neural components propose; symbolic components compose and check. Each covers the other's failure modes.
So we build hybrids, with a principled division of labor. Learned components do what statistics does best — perceive, retrieve, generate candidate structure from raw experience. Symbolic components do what structure does best — bind values to variables, compose operations systematically, and verify inference step by step. The neural half proposes; the symbolic half composes and checks.
The payoff we care most about is verifiability. When reasoning is explicit, a conclusion arrives with a derivation, and a derivation can be audited — by a person, or by a checker whose soundness does not depend on the system that produced it. For systems that must be trusted, this property is not a luxury.
Scale is a genuine force, and we use it. But the walls it keeps hitting — composition, abstraction, extrapolation — are exactly the walls structure was built for. The interesting research was never at either pole. It is in the joinery.