Abstract
<title>Abstract</title> <p>Current artificial intelligence systems draw their training signals from a narrow set of machine-accessible data environments, and despite rapid growth in model scale and training compute they show persistent ceilings in causal reasoning, social cognition, and compositional symbolic reasoning. We argue that these ceilings are failures of data environment rather than of scale. The signals the missing capabilities require exist only in environments that current systems engage shallowly or not at all. We adopt the Cyber-Physical-Social-Thinking (CPST) quadspace as a taxonomy of AI data environments, in which each non-cyber space supplies one irreplaceable signal type: physical interventional signals for causal reasoning, social intentional signals for social cognition, and thinking-space symbolic signals for compositional reasoning. Organized by this taxonomy, a screened corpus of eighty systems is placed in a coverage-capability $(K, L)$ space. The multi-space closed-loop region ($K\geq 3$) is empty across all eighty systems. The survey's central finding is that no cyber-only system, across architectures and the full range of training scales, has crossed from pattern recognition into reliable planning and counterfactual reasoning. Systems that close one non-cyber loop are much more likely to reach $L1.5$ or above: twenty-eight of forty-eight, compared with two of thirty-two systems without such a loop (two-sided Fisher's exact test, $p \approx 1.23\times10^{-6}$). Two controlled same-architecture contrasts carry the causal weight. We close with a three-pathway roadmap for quadspace expansion from narrow toward general intelligence.</p>