Background Theories & Constraints

Predictions about AI futures do not exist in a vacuum. They are supported, or constrained, by established principles from physics, economics, systems theory, and cognitive science. This document makes those constraints explicit, so that predictions can be evaluated against them and violations flagged.

Each entry follows the same structure: what the principle states, why it matters for AI forecasting, and what it constrains. Entries are tagged with a strength indicator. Hard constraints are physical or mathematical laws that cannot be violated. Soft constraints are empirical regularities that can be overcome, but require justification.

Physics

Thermodynamic Limits of Computation

The principle. Landauer’s principle sets an absolute minimum energy cost per bit erasure: kT ln 2, roughly 3 × 10⁻²¹ joules at room temperature. Real hardware operates orders of magnitude above this floor — current transistors dissipate roughly 1,000× the Landauer limit. The floor itself, however, is absolute. No amount of engineering circumvents it.

Why it matters for AI forecasting. Training frontier models already consumes megawatts. Even with perfect chip design, energy costs per computation have a hard floor. Compute scaling therefore faces a physical ceiling, not just an economic one. The constraint becomes more binding as models grow: a 100× larger model requires at least 100× more energy, with no theoretical workaround.

What it constrains. Predictions assuming unlimited compute scaling, negligible training costs, or that hardware improvements alone will remove the energy bottleneck. The thermodynamic floor means that energy infrastructure — not just chip design — determines the pace of AI scaling.

Computer Science

Scaling Laws (Chinchilla / Kaplan)

The principle. Neural network performance improves as a power law of compute, data, and parameters — but with diminishing returns. The Chinchilla results (Hoffmann et al., 2022) showed that optimal training requires scaling data proportionally with parameters, not just making models bigger. Kaplan et al. (2020) established that loss decreases as a smooth power law, meaning each halving of loss requires roughly an order of magnitude more compute.

Why it matters for AI forecasting. The scaling laws are simultaneously the strongest evidence for continued progress and the clearest constraint on its pace. They predict that gains will continue — but each increment requires exponentially more resources. The era of easy gains from simply scaling up appears to be ending. Future breakthroughs increasingly require algorithmic innovation, not just bigger clusters.

What it constrains. Predictions assuming continued exponential capability growth at constant cost. Also predictions that scaling alone — without architectural innovation — will produce AGI. The power-law relationship means we can extrapolate approximate capability trajectories, but those trajectories flatten.

Amdahl’s Law

The principle. The speedup achievable through parallelism is limited by the fraction of a task that must be executed sequentially. If 10% of a computation is inherently serial, no amount of parallel processing yields more than a 10× speedup — even with infinite processors.

Why it matters for AI forecasting. AI training and inference both contain serial bottlenecks. Autoregressive generation is inherently sequential (each token depends on all previous tokens). Multi-step reasoning requires serial composition. Training pipelines have synchronization points. More hardware does not, on its own, translate linearly into faster or more capable systems.

What it constrains. Predictions assuming that hardware scaling (more GPUs, larger clusters) translates directly into proportional capability gains. Also predictions about inference speed: autoregressive architectures face fundamental latency floors that parallelism cannot remove.

Goodhart’s Law

The principle. “When a measure becomes a target, it ceases to be a good measure.” Systems optimized for a proxy metric tend to find ways to score well on the proxy that diverge from the true objective. In machine learning, this manifests as benchmark gaming, reward hacking, and the gap between evaluation performance and real-world capability.

Why it matters for AI forecasting. AI benchmarks are proxy metrics for capability. Models that score perfectly on MMLU, HumanEval, or ARC may not exhibit proportional improvements in real-world usefulness. Benchmark saturation does not equal AGI. As benchmarks become targets for optimization, their signal degrades — we should expect benchmark scores to become less informative over time, not more.

What it constrains. Predictions that equate benchmark performance with genuine capability, or that declare milestone achievement based on benchmark scores alone. The divergence between proxy and real performance is likely to increase as models are optimized harder against fixed benchmarks.

Economics

Technology Adoption S-Curves

The principle. New technologies follow logistic (S-curve) adoption patterns: a slow start as early adopters experiment, then rapid growth as infrastructure and practices mature, then saturation. Even electricity took roughly 40 years from invention to majority industrial adoption. The internet took about 15 years from commercialization to ubiquity. No technology in recorded history has achieved instant mass adoption.

Why it matters for AI forecasting. Lab capability demonstrations are the beginning of the adoption curve, not the endpoint. Real deployment requires organizational restructuring, workforce retraining, infrastructure investment, trust-building, regulatory frameworks, and cultural adaptation. Each of these has its own timeline. The gap between “AI can do X in a demo” and “AI routinely does X in production” is measured in years, not months.

What it constrains. Predictions assuming instant mass adoption of AI capabilities, or that lab breakthroughs translate immediately to economic transformation. Also claims that specific industries will be “disrupted” within 1–2 years of a capability demo.

Jevons Paradox

The principle. When technological progress makes a resource cheaper to use, total consumption of that resource often increases rather than decreases, because lower cost enables new use cases that were previously uneconomical. Jevons observed this with coal in the 19th century; the pattern has since been documented across energy, computing, and communication technologies.

Why it matters for AI forecasting. Cheaper, more efficient AI inference will not reduce total compute and energy demand. Instead, it will expand the range of tasks worth automating — from high-value enterprise applications down to consumer micro-tasks. Efficiency gains drive demand expansion, not resource savings. This has direct implications for energy infrastructure planning and environmental impact.

What it constrains. Predictions assuming that efficiency improvements in AI will reduce total resource requirements, or that current infrastructure is sufficient for widespread deployment. Also optimistic energy projections that assume per-query efficiency gains translate to lower aggregate consumption.

Comparative Advantage

The principle. Ricardo’s insight: even when one party is absolutely better at everything, both parties benefit from specializing in what they do relatively best. Trade based on comparative (not absolute) advantage increases total output. The principle applies to any producer — including machines.

Why it matters for AI forecasting. Even if AI surpasses human performance in every cognitive task, humans retain economic roles where their relative advantage is greatest: tasks requiring embodiment, social trust, legal accountability, cultural context, or where the cost of AI deployment exceeds the cost of human labor. Complete displacement requires AI to be not just better, but cheaper per unit of output across every task simultaneously.

What it constrains. Predictions of complete human labor displacement. Even in scenarios where AI is cognitively superior across the board, comparative advantage guarantees that humans retain economic roles — though which roles, and at what wage levels, remains an open question.

Principal-Agent Problems

The principle. When one party (the agent) acts on behalf of another (the principal), misaligned incentives and information asymmetry produce suboptimal outcomes. The principal cannot fully observe or verify the agent’s actions, leading to monitoring costs, moral hazard, and adverse selection. These problems are structural, not solvable by better intentions.

Why it matters for AI forecasting. Deploying autonomous AI agents creates principal-agent problems at unprecedented scale. Organizations delegating decisions to AI systems cannot fully verify AI reasoning, detect subtle misalignment, or ensure that AI optimization targets match organizational intent. The more capable the AI agent, the harder it is to monitor — capability and verifiability are in tension.

What it constrains. Predictions assuming that technically capable AI agents will be rapidly deployed with full autonomy. In practice, organizations will limit delegation based on their ability to verify outcomes — creating a natural speed limit on autonomy that is independent of AI capability.

Systems Theory

Normal Accidents (Perrow)

The principle. In complex, tightly coupled systems, accidents are inevitable (“normal”) — not because individual components fail, but because of unexpected interactions between components that each work as designed. The more complex and tightly coupled the system, the more certain such accidents become. This is a property of system architecture, not engineering quality.

Why it matters for AI forecasting. As AI systems become more complex (multi-agent, deeply integrated into infrastructure) and more tightly coupled (real-time decision chains with minimal human buffer), systemic failures become statistically inevitable. This cannot be engineered away. It can only be managed through looser coupling and reduced complexity, both of which trade off against capability and speed.

What it constrains. Predictions assuming that sufficiently advanced AI systems will achieve near-perfect reliability in critical domains (autonomous driving, medical decisions, financial markets, military systems). Also the assumption that multi-agent AI systems can coordinate without emergent failure modes.

Cognitive Science

Bounded Rationality (Simon)

The principle. Herbert Simon’s insight: decision-makers operate with limited information, limited cognitive capacity, and limited time. They satisfice — find solutions that are “good enough” — rather than optimize. The principle applies not just to individuals but to organizations, markets, and governance institutions.

Why it matters for AI forecasting. Organizations adopting AI do not make perfectly rational adoption decisions. They satisfice: choosing what is available, affordable, politically feasible, and compatible with existing systems — not what is theoretically optimal. Adoption speed is bounded by organizational decision-making capacity, not just by AI capability. Committee approvals, budget cycles, pilot programs, and institutional inertia all impose delays that pure capability predictions ignore.

What it constrains. Predictions assuming rational, rapid, optimal adoption of AI across industries and institutions. Also claims about governance responses — regulators satisfice too, producing “good enough” rather than optimal policy.

Automation Paradox (Bainbridge)

The principle. Lisanne Bainbridge’s irony of automation: the more reliable an automated system becomes, the less practice human operators get at the tasks the system performs. When the system eventually fails — and all systems eventually do — the human backup is less capable than before automation was introduced. The better the automation, the worse the human fallback.

Why it matters for AI forecasting. As AI handles more cognitive work, humans lose the skills and situational awareness needed to oversee, correct, and intervene. This creates a fundamental tension in human-in-the-loop safety architectures: the loop depends on human competence, but automation degrades that competence. Over time, “human oversight” tends to become increasingly nominal.

What it constrains. Predictions assuming that human oversight remains effective as AI capability increases. Also the viability of “human-in-the-loop” as a long-term safety architecture — it may work as a transitional measure but degrades as the system it is meant to safeguard improves.

Historical Patterns

Amara’s Law

The principle. Roy Amara’s observation: “We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.” This pattern — hype, disappointment, then quiet transformation — has repeated across electricity, aviation, the internet, smartphones, and virtually every general-purpose technology.

Why it matters for AI forecasting. Current AI predictions cluster heavily in 2025–2030 for transformative impact. Historical precedent suggests two corrections: near-term impact will likely disappoint relative to the most aggressive predictions, while long-term impact may exceed what even optimists currently imagine. The pattern does not tell us what will happen — but it strongly suggests that the timeline will be shifted rightward relative to peak-hype predictions.

What it constrains. Both aggressive near-term predictions (AGI by 2027, ASI by 2028) and dismissive long-term projections (AI will not fundamentally change the economy). The historical pattern suggests that both extremes are wrong, and that the truth lies between them — but on different timescales than currently predicted.