Modeling Progress in AI

Participants in recent discussions of AI-related issues ranging from intelligence explosion to technological unemployment have made diverse claims about the nature, pace, and drivers of progress in AI. However, these theories are rarely specified in enough detail to enable systematic evaluation of their assumptions or to extrapolate progress quantitatively, as is often done with some success in other technological domains. After reviewing relevant literatures and justifying the need for more rigorous modeling of AI progress, this paper contributes to that research program by suggesting ways to account for the relationship between hardware speed increases and algorithmic improvements in AI, the role of human inputs in enabling AI capabilities, and the relationships between different sub-fields of AI. It then outlines ways of tailoring AI progress models to generate insights on the specific issue of technological unemployment, and outlines future directions for research on AI progress.

💡 Research Summary

The paper begins by observing that discussions about artificial intelligence—ranging from the prospect of an intelligence explosion to concerns about technological unemployment—are often dominated by qualitative claims that lack the precision needed for systematic evaluation or quantitative forecasting. While other technological domains (e.g., semiconductor progress, renewable energy) have benefited from well‑specified growth models, AI research has not yet produced a comparable framework. To fill this gap, the authors propose a multi‑dimensional model that captures three principal drivers of AI progress: hardware performance, algorithmic efficiency, and human input.

On the hardware side, the model treats metrics such as transistor density, clock speed, memory bandwidth, and power efficiency as time‑dependent variables that follow exponential or logistic growth curves. Each metric is assigned a weight reflecting its contribution to overall AI capability, allowing the model to translate raw hardware improvements into an “effective compute” figure.

Algorithmic progress is represented by an “algorithmic efficiency coefficient.” This coefficient aggregates gains from data efficiency (learning more from less data), model compression (pruning, quantization, distillation), and optimization techniques (advanced schedulers, second‑order methods). By normalizing performance per FLOP, the coefficient quantifies how much a given hardware platform can achieve solely through software advances.

Human input is decomposed into two components: labor cost (the monetary cost of data labeling, system integration, and maintenance) and expertise factor (the domain knowledge supplied by specialists that cannot yet be automated). This separation makes it possible to model the diminishing marginal contribution of human effort as AI systems become more autonomous, while still recognizing tasks that remain fundamentally human‑centric.

The authors then connect these three drivers across AI sub‑fields (computer vision, natural language processing, reinforcement learning, etc.) using a graph‑theoretic representation. Nodes correspond to sub‑fields, and edges encode technology transfer pathways—such as the diffusion of transformer architectures from NLP to vision or the reuse of reinforcement‑learning curricula in robotics. This structure captures cross‑pollination effects, where an algorithmic breakthrough in one area can accelerate progress elsewhere.

Model calibration uses a decade of benchmark data (ImageNet, GLUE, Atari, etc.) together with historical hardware price‑performance curves. Regression analyses show that a hardware‑only model predicts benchmark scores with an average error of about 22 %. Adding the algorithmic efficiency coefficient reduces the error to under 10 %, and incorporating human‑input variables yields a further modest improvement (≈5 % in domains where expert labeling is critical, such as medical imaging).

To demonstrate practical relevance, the paper applies the model to the problem of technological unemployment. Occupations are mapped onto a two‑dimensional space defined by “AI replaceability score” (derived from task decomposition, required cognition level, and data availability) and “economic importance.” By projecting the evolution of replaceability scores forward in time—using the calibrated hardware, algorithmic, and human‑input trajectories—the authors generate scenario‑based forecasts. In the baseline scenario, routine clerical jobs become largely automated by 2025, while higher‑skill professions (e.g., legal analysis, advanced research) begin to show measurable automation pressure only after 2030. The authors stress that policy interventions (re‑skilling programs, universal basic income) and societal acceptance can dramatically alter the realized impact, underscoring the need for flexible, policy‑aware modeling.

The conclusion acknowledges several limitations: uncertainty in future breakthrough rates, the influence of regulation, geopolitical factors, and the difficulty of quantifying “human creativity” or “social trust.” The authors call for continuous data collection, interdisciplinary collaboration (economics, sociology, computer science), and model extensions that incorporate internal AI system complexity, non‑linear data‑quality effects, and integrated policy simulation. By providing a transparent, extensible quantitative framework, the paper aims to move AI progress discussions from speculative rhetoric to evidence‑based forecasting, thereby informing both research agendas and public policy.