New AI Research Challenges Scaling Dogma: Hybrid Architectures Seen as Key to Real-World Capability

A recent preprint published on arXiv challenges the prevailing assumption that larger AI models inevitably lead to greater capability. The paper, titled "Capability from Access Structure, Not Scale," proposes the Capability Convergence Hypothesis (CCH) as a counterpoint to the Platonic Representation Hypothesis (PRH). While PRH suggests that representations from different models converge as they scale, CCH argues that under a fixed per-token inference budget, representational convergence does not equate to capability convergence. Instead, capability is determined by the model's access structure—specifically, whether it possesses both a compressive O(1)-state channel and a scalable verbatim-index channel, a combination the authors call "access-complete hybrid."

The research anchors its argument on a witness task called the Newton's-apple problem, set in an infinite data stream. The authors identify three resource walls that any architecture must cross: a Shannon wall that blocks any model with less than O(Nb) states, a horizon wall that defeats any fixed window, and a circuit wall that limits fixed-depth attention-only models (conditional on the computational complexity assumption TC0 != NC1). The access-complete hybrid architecture crosses all three walls by paying each wall's price individually, demonstrating that capability is strictly super-additive under composition.

Crucially, the paper separates proven results from conjectures. The access-completeness principle is grounded in information-theoretic lower bounds and pre-registered experiments, while the field-level convergence trend is described as an economics-motivated conjecture. The authors report the first small-scale pre-registered tests, with criteria frozen before data collection. The predicted "scissors gap" was measured: exact-retrieval error of 0.994 for a 64-scalar state model versus 0.000 once a single global-attention layer was added. The state-tracking bifurcation occurred at the registered boundary, and a conjunction witness showed an irreducibly two-channel solution. One prediction failed with reversed direction and was reported as such.

The paper concludes that representational convergence is given freely by scale, but capability convergence must be purchased by access structure. This distinction has significant implications for the AI industry, suggesting that endless scaling of monolithic models may yield diminishing returns, while investment in hybrid architectures could unlock new practical capabilities. The authors' pre-registered experiments lend credibility to their claims, making this a potential turning point in AI architecture design.

For investors and business leaders, the takeaway is clear: the next wave of AI value may not come from larger models but from smarter, hybrid designs that combine compression and retrieval. Companies that pivot to building access-complete systems—or that supply the hardware and software needed for such architectures—could be positioned for outsized gains. The research also highlights the importance of pre-registered experimental validation, which could reduce the hype cycle around new AI claims.