Here are the 10 most important references from “From AGI to ASI” (arXiv:2606.12683)

Here are the 10 most important references from “From AGI to ASI” (arXiv:2606.12683), ranked by conceptual centrality to the paper’s core arguments:

Tier 1 — Foundational to the Paper’s Theoretical Spine

• Legg & Hutter (2007a) — “Universal Intelligence: A Definition of Machine Intelligence” Provides the Legg-Hutter score, the formal backbone of the entire paper’s intelligence continuum. The paper’s definitions of AGI, ASI, and Universal AI all rest on this framework. Without it, the paper lacks its formal grounding entirely.
• Hutter (2005) — “Universal Artificial Intelligence: Sequential Decisions Based on Algorithmic Probability” Introduces AIXI, the theoretical endpoint of machine intelligence that defines Universal AI in the paper. The authors explicitly use AIXI as the formal upper bound above ASI.
• Kaplan et al. (2020) — “Scaling Laws for Neural Language Models” (OpenAI) The paper’s entire treatment of how compute growth translates into capability relies on scaling laws. Cited as the key empirical basis for predicting frontier model improvement trajectories.

Tier 2 — Directly Shaping the Four Pathways and Bottleneck Analysis

• Good (1965) — “Speculations Concerning the First Ultraintelligent Machine” The original source for the intelligence explosion / recursive self-improvement concept. The paper’s Pathway 3 (recursive improvement) traces directly back here. First to articulate the idea that a machine smarter than humans could design its own successors.
• Bostrom (2014) — Superintelligence: Paths, Dangers, Strategies Cited multiple times as the canonical reference for fast AI takeoff scenarios and alignment risks. Represents the dystopian side of the AGI-to-ASI discourse that the paper explicitly engages with.
• Ho et al. (2025) — “Benchmark Stitching” (Epoch AI) A newer empirical paper that the authors lean on heavily as the methodologically sound approach for extrapolating capability improvements. Cited more frequently than nearly any other recent work; described as offering a framework that goes beyond simple scaling laws.
• Morris et al. (2024) — “Levels of AGI” (Google DeepMind) Defines five levels of AGI capability. The paper uses this taxonomy directly when characterizing where AGI and ASI sit, and repeatedly refers to the “Competent AGI” and top-level definitions from this work.

Tier 3 — Key Context-Setters and Framing References

• Turing (1950) — “Computing Machinery and Intelligence” Used as the paper’s epigraph. More than decorative — it anchors the historical legitimacy of the inquiry itself and signals the paper’s ambition to place itself in the long arc of AI philosophy.
• Bloom et al. (2020) — “Are Ideas Getting Harder to Find?” Critical to the paper’s friction analysis. Cited to establish the empirical pattern that sustaining research progress traditionally requires exponentially increasing inputs — a key tension against recursive self-improvement optimism.
• Solomonoff (1985) — “The Time Scale of Artificial Intelligence: Reflections on Social Issues” Identified as the first to formally discuss hyperbolic growth in the context of AI self-improvement. The paper credits Solomonoff explicitly as the origin of the singularity-via-AI-recursion argument, before Kurzweil popularized it. A historically underappreciated citation.

A note on prioritization: The Legg-Hutter / Hutter / AIXI cluster is unambiguously most important because the paper’s entire definition of intelligence and the AGI→ASI→UAI continuum is formally grounded in those works. The scaling law and efficiency measurement papers (Kaplan, Ho et al.) are the empirical engine. The intelligence-explosion classics (Good, Bostrom, Solomonoff) provide the intellectual lineage the paper is positioning itself within.