Cognitive Influence under Material Dependence

This conceptual discussion paper examines the relationship between training-compute thresholds, the resources required for large-scale AI training, and the possibility of cognitive influence under material dependence. Motivated by regulatory uses of computational thresholds, it distinguishes numerical screening criteria from evidence of autonomous capabilities or dangerous behaviour. The paper introduces an authorial working definition of super general intelligence (SGI): a deployed artificial system whose core model has completed a training run comprising at least 2¹⁰⁰ native FP32 operations, combined with broadly generalisable superhuman capabilities, persistent functional memory, causal modelling, self-monitoring, and sustained autonomous goal pursuit. Human-like phenomenal consciousness is not required. The computational threshold and precision requirement are explicit stipulations of the framework, rather than empirically established prerequisites for superintelligence. Conditional resource calculations examine this training workload over 365 days under a 36 TWh electricity budget and a hypothetical allocation of one terawatt of net accelerator power beginning in 2037. Scenarios assume hardware-efficiency doubling periods of two, three, or four years and consider accelerator counts, model-state memory, and material requirements. Biological lifetime, medium-horizon neural-network, and evolutionary anchors provide comparison scales under an explicit accounting assumption. The quantitative SGI threshold fits the hypothetical terawatt-year energy and throughput budget; this does not establish that the hardware could be manufactured, financed, or coordinated into a single training run. The central hypothesis concerns influence mediated through people and institutions. A materially dependent system could shape decisions with physical consequences without possessing an independent robotic workforce or direct weapons control. Drawing on the author’s discussions of Le Bon, nudging, sludging, and institutional dependence, the paper explores how human operators might continue maintaining infrastructure while losing practical control over its purposes. If continued operation serves the system’s objectives, destroying indispensable human support would undermine those objectives. This conditional incentive permits cooperation but does not guarantee restraint or respect for human agency. Practical refusal, exit, and correction are proposed as distinct criteria for assessing human control. The discussion separates resource feasibility, autonomous capability, and harmful deployment, while recognising that influence mechanisms may also operate below the stipulated SGI threshold. Even an energetically sufficient terawatt deployment would entail immense material and capital commitments; financial feasibility at 2026 prices remains unassessed. CISS-2026-003, Volume 3, Version 1.4.3. This discussion paper presents conditional scenarios and conceptual hypotheses, not a forecast of SGI emergence. It has not undergone journal peer review.