Research map / Large language models / AI and society
Increased Compute Efficiency and the Diffusion of AI Capabilities
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
A conceptual policy paper formalizing "compute investment efficiency" (hardware price performance times algorithmic efficiency) and arguing that its access and performance effects cause AI capabilities to diffuse while leading compute investors keep a performance advantage.
Problem
How falling training costs from hardware price-performance gains and algorithmic efficiency gains affect the diffusion of AI capabilities, competitive advantage, and dangerous-capability proliferation.
Approach
Introduces compute investment efficiency, f_c(i) = f_a(f_h(i)), and decomposes its growth into an access effect (less compute needed for a given performance) and a performance effect (more performance for a given investment), arguing large compute investors stay ahead as capabilities diffuse absent performance ceilings.
Key results
- Cites that training an ImageNet classifier to 93% accuracy fell from over $1,000 in 2017 to $5 in 2021, a reduction of over 99%.
- Reports that the compute required for ImageNet 93% accuracy halved every nine months from 2012 to 2022 (cited analysis).
- Argues that the access effect and performance effect together let leaders keep a performance advantage as capabilities diffuse, unless performance ceilings or diminishing returns dampen the performance effect.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase compute efficiency through hardware price-performance improvements | AI accelerators · AI accelerator hardware, 2006–2021 | price performance: doubled approximately every two years | Helped |
| Improve image classification algorithms | Image classification training · Image classification, 2012–2022 | compute required for 93% classification accuracy: halved every nine months | Helped |
| Improve image classification algorithms | Image classification training · Image classification, 2012–2019 (AlexNet performance) | compute required: 97.7\% reduction | Helped |
| Increase compute efficiency over time | AI model training · Given compute investment | model performance | Helped |
| Increase compute efficiency over time | AI model training · Given performance level | required compute investment | Helped |
Surprising findings
- Frontier developers have become reluctant to publish architecture and compute details, so algorithmic advances diffuse mainly through leaks or employees switching companies.
Limitations
- Model is conceptual and illustrative; figures do not assert empirical curve slopes.
- Assumes constant investment, equal access to efficiency gains, and that actors only access capabilities of models they train, relaxed only in discussion.
- Assumes compute is the main determinant of capabilities, ignoring talent and data bottlenecks in detail.
Experimental setup
Builds on: Social and Governance Implications of Improved Data Efficiency
Closest papers by meaning
- AI Progress Should Be Measured by Capability-Per-Resource, Not Scale Alone: A Framework for Gradient-Guided Resource Allocation in LLMs (NeurIPS 2025 · similarity 0.86)
- Resource Democratization: Is Compute the Binding Constraint on AI Research? (AAAI 2024 · similarity 0.84)
- Path Dependence under Adaptive AI Delegation (NeurIPS 2026 · similarity 0.83)
- Position: AI Scaling: From Up to Down and Out (ICML 2025 · similarity 0.82)
- The Elicitation Game: Evaluating Capability Elicitation Techniques (ICML 2025 · similarity 0.82)
- The Leaderboard Illusion (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.