Derivative Statistical Grounding and the Limits of the AGI Roadmap

The prevailing roadmap to artificial general intelligence assumes that large language models are precursors to AGI and that continued scaling, combined with architectural refinements, will eventually cross the threshold to general intelligence. This paper argues that the roadmap is misconceived. The obstacle is not scaling limits, data exhaustion, or engineering challenges, though these are real. The obstacle is grounding. LLMs instantiate what we term Derivative Statistical Grounding (DSG): they operate on statistical patterns in human-produced text rather than on the world that text describes. This derivative character constitutes a categorical limit on what scaling can achieve. AGI, understood as general intelligence capable of handling genuinely novel situations, would require endogenous truth-evaluation: the capacity to check outputs against reality rather than against patterns that correlate with reality. DSG systems cannot achieve this regardless of scale because they have no pathway to the causal ground that makes truth-evaluation possible. More sophisticated pattern-matching remains pattern-matching. The derivative layer extends but does not breach. We engage directly with recent AGI predictions from OpenAI, Anthropic, and Microsoft Research, arguing that these predictions conflate capability scaling with architectural transformation. The "sparks of AGI" visible in current systems are sparks of sophisticated pattern-matching, not sparks of grounded intelligence. The paper concludes that AGI, if achievable, requires a fundamentally different architectural approach, one that provides embodied world-contact rather than derivative access to human cognitive artifacts.