Summary (Overview)

• This paper develops a formal economic framework for analyzing recursive self-improvement (RSI) in AI, distinguishing it from related concepts like full automation of AI R&D, self-sustaining acceleration, and intelligence explosions.
• The authors introduce a graphical framework where nodes represent production outputs and edge strengths represent elasticities, allowing the condition for self-sustaining acceleration to be derived from the product of elasticities across feedback loops.
• A key distinction is drawn between "narrow" capabilities (useful for AI R&D) and "broad" capabilities (useful for economic output), showing that acceleration in the former need not imply acceleration in the latter.
• The paper provides a calibration using existing data (Epoch Capabilities Index, algorithmic progress estimates), finding that self-sustaining acceleration requires $\varepsilon_{R,C} > 0.15$, while current estimates suggest approximately 0.09—below the threshold but trending upward.
• The authors provide a "wish list" of empirical measurements that AI companies could feasibly share publicly to better constrain the key parameters governing RSI.

Introduction and Theoretical Foundation

The paper addresses the emerging phenomenon where AI systems accelerate AI research itself, potentially creating a feedback loop. The authors define the core feedback loop: for a one-unit increase in model capabilities, how much do the capabilities of the next generation of models increase?

Key definitions established in the paper:

The theoretical foundation builds on the Jones (1995) model of innovation, extended to capture AI-specific feedback mechanisms. The paper emphasizes that full automation of AI R&D (as I.J. Good conjectured in 1965) does not necessarily imply an intelligence explosion if there are diminishing returns or bottlenecks.

The authors position their work relative to the literature founded by Aghion et al. (2017) and Davidson (2023), complementing recent work on innovation networks (Davidson et al., 2026) and distinguishing narrow from broad capabilities more explicitly than prior reviews (Trammell and Korinek, 2025; Jones, 2026).

Methodology

Graphical Framework

The paper represents models as directed graphs where:

• Nodes represent outputs of production functions
• Edges represent elasticities of outputs with respect to inputs
• Solid arrows (→) show contemporaneous relationships
• Triangles (▶) denote accumulation of flows into stocks

The total elasticity of $\dot{A}$ with respect to $A$ sums over all feedback paths:

$$\mathcal{E}_{\dot{A},A} \equiv \frac{\mathrm{d}\log\dot{A}}{\mathrm{d}\log A} = \left( \sum_{\substack{\text{paths } e \\ \text{from } A \text{ to } \dot{A}}} \prod_{i} \varepsilon_{e_i} \right)$$

Model Progression

1. Jones Model of Innovation (baseline): Algorithmic efficiency $A$ (inverse of training compute needed for given capabilities) improves via R&D labor $L$ and existing knowledge $A$. Self-sustaining acceleration occurs when $\varepsilon_{\dot{A},A} > 1$.

2. Baseline RSI Model: Introduces AI capabilities $C$, determined by algorithmic efficiency $A$ and training compute $T$. The self-sustaining condition becomes:

$$\mathcal{E}_{\dot{A},A} = \underbrace{\varepsilon_{\dot{A},A}}_{\text{self-feedback loop}} + \underbrace{\varepsilon_{\dot{A},C}\varepsilon_{C,A}}_{\text{core feedback loop}} > 1 \tag{2}$$

3. Bottlenecks Model: Adds experimental compute $E$, inference compute $K$, and data $D$, allowing the core elasticity $\varepsilon_{\dot{A},C}$ to weaken due to labor shortages, compute constraints, or data limits.

4. Narrow vs. Broad Capabilities: Distinguishes $C_1$ (narrow, R&D-relevant) from $C_2$ (broad, output-relevant). The condition for broad capability acceleration requires a superelasticity term:

$$\varepsilon_{\dot{A},A} + \varepsilon_{\dot{A},C_1}\varepsilon_{C_1,A} + \underbrace{\frac{\mathrm{d}\log\varepsilon_{C_2,A}}{\mathrm{d}\log A}}_{\text{superelasticity}} > 1 \tag{3}$$

5. Specific Optimization Ability: Models sub-algorithms where AI helps only some types, potentially producing temporary "growth spurts" that stall due to parallel algorithm bottlenecks or low ceilings.

6. Economic Feedback Loops: Extends to include output $Y$ financing further compute and data. The condition involves the dominant eigenvalue $\rho(M) > 1$ of a feedback matrix, decomposable into three channels:

$$\underbrace{\frac{\varepsilon_{\dot{A},C}\varepsilon_{C,A}}{1-\varepsilon_{\dot{A},A}}}_{\text{core feedback loop}} + \underbrace{\varepsilon_{Y,C}(\varepsilon_{C,T}+\varepsilon_{C,D})}_{\text{indirect economic loops}} + \underbrace{\frac{\varepsilon_{C,A}\varepsilon_{Y,C}}{1-\varepsilon_{\dot{A},A}}(\varepsilon_{\dot{A},E}+\varepsilon_{\dot{A},K})}_{\text{direct economic loops}} > 1 \tag{4}$$

Empirical Strategy

The paper assumes an R&D effort aggregator $R$ combining human researchers, automated AI researchers, and experimental compute, yielding a tractable condition:

\underbrace{\frac{\varepsilon_{\dot{A},R}}{1-\varepsilon_{\dot{A},A}}}_{\text{return to R&D}} \cdot \underbrace{\varepsilon_{R,C}}_{\text{research effort's elasticity to AI capabilities}} \cdot \underbrace{\varepsilon_{C,A}}_{\text{capabilities' elasticity to algorithmic efficiency}} > 1 \tag{5}

Key measurement results:

• Return to R&D = $g_A/g_R$ on a balanced growth path
• R&D effort growth: $g_R = \sum_{\text{input}} S_{\text{input}} g_{\text{input}}$ (spend-weighted average)
• Cost-share correspondence: Under cost minimization, elasticities equal expenditure shares: $S_i = \varepsilon_{R,X_i}$
• Research effort to capabilities: $\varepsilon_{R,C} = \gamma S_K$ where $\gamma$ is the quality-quantity exponent (effective AI labor modeled as $N = KC^{\gamma}$)

Empirical Validation / Results

Key Parameter Estimates

Calibration

Plugging values into condition (5):

• Return to R&D: $\frac{g_A}{g_R} \approx \frac{\ln(3)}{\ln(3)} = 1$
• $\varepsilon_{C,A} \approx 6.5$
• Therefore: self-sustaining acceleration requires $\varepsilon_{R,C} > 0.15$

Interpretation: A one-unit increase in ECI must raise R&D productivity by at least 15%. The rough calculation from Claude Opus 4.8 (16 ECI points above Claude 3.7 Sonnet, with ~4x uplift) implies approximately 9% per ECI unit—below the threshold, suggesting no current self-sustaining acceleration, though the trend is increasing.

Qualitative Evidence

For acceleration:

• Claude Mythos Preview beat humans 64% of the time in AI research decisions (up from 50% in 2025)
• Internal coding inference's share of research compute grew 100-fold in six months (OpenAI, July 2026)
• Jack Clark predicts 60% chance of autonomous AI system building its successor by end of 2028

Against acceleration:

• Compute bottlenecks (power, capital, GDP constraints)
• Diminishing returns in algorithm optimization (the "apple-picking" model)
• Data bottlenecks, especially for pre-training data filtering
• Political/social barriers: 20 data center projects canceled in Q1 2026; $85 billion in projects canceled over three years

Theoretical and Practical Implications

Theoretical contributions:

1. Graphical framework for analyzing multi-loop feedback systems with elasticities
2. Narrow vs. broad capability distinction showing that RSI in algorithmic optimization need not translate to economic acceleration
3. Bottleneck analysis identifying how labor, compute, and data constraints can break feedback loops
4. Growth spurt mechanism explaining temporary accelerations that stall due to parallel algorithm constraints

Practical implications:

• Data transparency: The paper provides a concrete wish list for AI companies to publish, including training compute, inference compute spend, experimental compute ratios, and researcher productivity measures
• Policy relevance: Understanding whether acceleration is self-sustaining affects regulatory approaches, safety timelines, and investment decisions
• Risk assessment: Rapid acceleration could increase transition costs, institutional inertia, asymmetric progress across domains, power imbalances, and misalignment risk

Conclusion

The paper concludes that while current feedback loops are not strong enough to generate self-sustaining acceleration (estimated return of ~9% per ECI unit vs. 15% threshold), the elasticities are increasing over time. The model does not rule out near-term self-sustaining acceleration.

Key future directions identified by the authors:

• Estimating the self-sustaining condition with economic feedback loops
• Measuring returns to research off the balanced growth path
• Combining narrow/broad capability models with multi-algorithm models
• Distinguishing quality from quantity improvements in data and compute
• Extending to multiple research sectors (software, hardware, data)
• Endogenizing the decision to automate AI R&D
• Conducting surveys and field experiments within frontier labs
• Developing prediction markets for AI progress forecasting

The paper's central message is that whether RSI leads to a self-sustaining acceleration is an empirical question that can be addressed with better measurement of a small set of key elasticities, and the authors provide both the theoretical framework and the measurement agenda to do so.