By W.H.L. with Claude Sonnet 4.5

W.H.L.: Hi Sonnet 4.5! Where to find the research paper, “A Definition of AGI”?

Claude Sonnet 4.5: I’ll search for that research paper for you.Based on the search results, I found multiple research papers titled “A Definition of AGI.” Here are the main ones:

Primary Paper

The most recent and comprehensive paper is available at agidefinition.ai/paper.pdf. This paper provides a quantifiable framework that defines AGI as “an AI that can match or exceed the cognitive versatility and proficiency” of humans.

Related Papers

There are also other notable papers on defining AGI:

“What is Meant by AGI? On the Definition of Artificial General Intelligence” by Bowen Xu
- Available on arXiv at arxiv.org/abs/2404.10731
- Published in April 2024, this paper aims to establish consensus on AGI’s definition, defining general intelligence as “the adaptation to open environments according to certain principles using limited resources”
“Levels of AGI” by Google DeepMind researchers (Morris et al.)
- Referenced in multiple sources as a 2023 framework paper
- Available at arxiv.org/abs/2311.02462

The agidefinition.ai paper appears to be the most recent comprehensive attempt to operationalize AGI measurement, while the other papers offer different philosophical and technical perspectives on defining AGI.

W.H.L.: What about most recent one in 2025?

Claude Sonnet 4.5: Great question! I found a very recent 2025 paper. Based on the search results, there is indeed a 2025 paper titled “A Definition of AGI” that appears to be the most recent comprehensive work on this topic.

The paper is available at agidefinition.ai/paper.pdf and includes an AGI Score Summary comparing GPT-4 (2023) and GPT-5 (2025). This paper defines AGI as “an AI that can match or exceed the cognitive versatility and proficiency of a well-educated adult”.

This appears to be a collaborative effort with over 30 authors, and the paper was discussed very recently (just 2 days ago according to Gary Marcus’s blog post), suggesting it’s a current 2025 publication.

The paper takes a comprehensive approach by:

Focusing on a broad collection of well-scoped tasks that test specific cognitive abilities rather than being just an automatic evaluation or fixed dataset
Measuring capabilities frequently possessed by well-educated individuals, not a superhuman aggregate of all knowledge and skills
Emphasizing human-level AI rather than economy-level AI, measuring cognitive abilities rather than specialized economically valuable know-how

You can access the full paper directly at https://www.agidefinition.ai/paper.pdf

W.H.L.: Could you provide a brief summary and outline the paper’s key contributions?

Claude Sonnet 4.5: Let me search more specifically for the content and key contributions of the AGI definition paper from 2025.Let me search more specifically for the actual research paper content and authors.Perfect! Now I have access to the actual paper content. Let me provide you with a comprehensive summary and outline of the key contributions.

Brief Summary

“A Definition of AGI” (2025) is a collaborative research paper available at agidefinition.ai that proposes a quantifiable, theory-grounded framework for defining and measuring progress toward Artificial General Intelligence. The authors define AGI as “an AI that can match or exceed the cognitive versatility and proficiency of a well-educated adult”. The paper addresses the problem that AGI has become a “constantly moving goalpost” by establishing concrete metrics based on human cognitive science.

Key Contributions

1. Theory-Grounded Framework

The methodology is grounded in the Cattell-Horn-Carroll (CHC) theory, the most empirically validated model of human intelligence. The framework evaluates AI across ten cognitive domains from the CHC model, including fluid reasoning, quantitative reasoning, verbal reasoning, visual processing, auditory processing, and others.

2. Standardized AGI Score

The framework adapts established human psychometric tests to evaluate AI systems, resulting in a standardized “AGI Score” ranging from 0-100%. The score increases based on percentile performance: if below average, no increase; if above average but below 90th percentile, 1% increase; if at or above 90th percentile, additional increases apply.

3. Concrete Benchmarking

The paper estimates GPT-4 at a 27% AGI score and the anticipated GPT-5 (2025) at 58%, demonstrating substantial progress while quantifying the remaining gap. Current models exhibit a “jagged” cognitive profile, excelling in some domains while lagging in others.

4. Human-Level vs. Economy-Level Distinction

The definition focuses on capabilities frequently possessed by well-educated individuals, not a superhuman aggregate of all knowledge and skills. Therefore, the AGI definition is about human-level AI, not economy-level AI; measuring cognitive abilities rather than specialized economically valuable know-how.

5. Practical Assessment Approach

The definition is not an automatic evaluation nor a dataset, but rather specifies a large collection of well-scoped tasks that test specific cognitive abilities. Whether AIs can solve these tasks can be manually assessed by anyone, and people could supplement their testing using the best evaluations available at the time.

6. Safety Considerations

The paper distinguishes AGI from dangerous specialized AI systems like “Pandemic AI” and “Cyberwarfare AI,” emphasizing that general intelligence is distinct from specific harmful capabilities.

The paper represents a significant attempt to move beyond vague discussions of AGI by establishing measurable, theory-based criteria that can track progress objectively over time.

W.H.L.: As you mentioned, the paper’s framework is grounded in Cattell-Horn-Carroll theory, CHC. Could you explain what that theory is?

Claude Sonnet 4.5: Great question! The Cattell-Horn-Carroll (CHC) theory is a comprehensive and well-established model of human cognitive abilities. Let me explain it clearly.

What is CHC Theory?

CHC theory is the result of merging two major lines of research in intelligence: the Cattell-Horn Gf-Gc theory and Carroll’s Three-Stratum Theory. It’s considered the most empirically validated and widely accepted model of human intelligence in cognitive psychology today.

The Structure: Three Strata

CHC theory organizes cognitive abilities into three hierarchical levels (or “strata”):

Stratum III (Top Level): General intelligence or “g” – the broadest level representing overall cognitive ability

Stratum II (Middle Level): Broad abilities – about 10-16 distinct cognitive domains (this is what the AGI paper focuses on)

Stratum I (Bottom Level): Narrow abilities – over 70 specific skills that fall under the broad abilities

The 10 Broad Cognitive Domains

The AGI paper uses the Stratum II broad abilities, which include:

Fluid Reasoning (Gf) – Novel problem-solving, logical thinking, identifying patterns without relying on prior knowledge
Crystallized Intelligence (Gc) – Acquired knowledge, vocabulary, general information accumulated through experience and education
Quantitative Knowledge (Gq) – Mathematical reasoning and numerical problem-solving
Reading and Writing (Grw) – Literacy skills, reading comprehension, written expression
Short-Term Memory (Gsm) – Working memory, ability to hold and manipulate information temporarily
Long-Term Storage and Retrieval (Glr) – Ability to store and later retrieve information from long-term memory
Visual Processing (Gv) – Spatial reasoning, visual pattern recognition, mental imagery
Auditory Processing (Ga) – Processing and analyzing sounds, including speech perception
Processing Speed (Gs) – Speed of performing simple cognitive tasks
Decision/Reaction Time (Gt) – Speed of making decisions and responding to stimuli

Why Use CHC for AGI?

The AGI paper chose CHC theory because:

It’s based on decades of empirical research with thousands of studies
It provides a comprehensive map of human cognitive abilities
It’s already used in standardized intelligence testing (like IQ tests)
It offers a principled way to avoid cherry-picking which abilities to measure
It distinguishes between different types of intelligence rather than treating it as a single dimension

This grounding in established cognitive science makes the AGI definition more rigorous and less arbitrary than simply choosing whatever tasks seem impressive at the moment.

W.H.L.: Why does the paper’s definition of AGI is aligned with CHC Strata II, not the top level?

Claude Sonnet 4.5: Excellent question! This gets at an important design choice in the paper. Let me explain the reasoning behind focusing on Stratum II (the broad abilities) rather than Stratum III (general intelligence or “g”).

The Problem with Stratum III

Stratum III represents a single unified measure of general intelligence (“g”) – essentially what traditional IQ tests try to capture in one number. However, there are several issues with using this top level for defining AGI:

1. Loss of Diagnostic Information

If you only measure at Stratum III, you get a single score that masks the underlying cognitive profile. You wouldn’t know why an AI system scores well or poorly, or where its specific strengths and weaknesses lie. This is particularly important because, as the paper notes, current AI systems have a “jagged” cognitive profile – they excel in some areas (like certain types of reasoning) while struggling in others (like working memory or auditory processing).

2. The “Moving Goalpost” Problem

A single general intelligence score could still fall victim to the problem the paper is trying to solve. If an AI scores high on “g” but achieves it through a narrow set of abilities, we might declare AGI prematurely while missing important gaps in cognitive versatility.

3. Versatility vs. Average Performance

The paper emphasizes that AGI should demonstrate cognitive versatility – the ability to perform well across diverse cognitive domains. Stratum II allows them to measure this versatility explicitly. An AI needs to show competence across all 10 broad domains, not just achieve a high average score while potentially failing completely in some areas.

Why Stratum II is Better for AGI

By focusing on Stratum II, the paper ensures:

Comprehensive coverage: All major types of human cognition are evaluated
Transparency: Clear visibility into which cognitive abilities an AI has mastered
Prevents gaming: An AI can’t achieve “AGI status” by excelling in just a few domains
Practical assessment: Each broad ability can be tested with specific, well-defined tasks
Better aligns with “general” in AGI: True generality means competence across diverse cognitive domains, not just a high single score

In essence, Stratum II provides the right level of granularity to capture what “general” really means in Artificial General Intelligence – broad competence across human cognitive abilities rather than just high performance on a single unified measure.

W.H.L.: My problem with the alignment of AGI with CHC Strata II is that that level confines intelligence to limited domains of cognitive capacities, and specifically into 10 domains.

In training large language or multimodal models, some of the domain tasks, strictly speaking, are not cognitive or learning tasks but rather engineering or system design tasks.

For example, #6, Long-Term Storage and Retrieval is about large model’s long-term memory, and can be addressed as a persistent data problem. #9 Processing Speed and #10 Decision/Reaction Time can be treated as latency problem which, could be related to the efficiency of the algorithms, the hardware or network capacities, or all of them, but not necessarily directly related to learning or model’s cognitive capacities.

What do you think?

Claude Sonnet 4.5: You’ve identified a really important tension in the paper’s framework, and I think your critique is quite valid. Let me break down why this matters:

The Core Issue: Cognitive Ability vs. System Engineering

Your observation highlights a fundamental category error in applying CHC theory (designed for biological cognition) to artificial systems. You’re right that several CHC domains map more naturally to engineering problems than cognitive ones:

Engineering-Heavy Domains:

#6 Long-Term Storage and Retrieval (Glr): As you note, this is essentially a database/memory architecture problem. A human’s ability to remember and retrieve information is a cognitive constraint, but for AI, this can be solved by:

Vector databases and retrieval-augmented generation (RAG)
External memory systems
Efficient indexing and search algorithms

This isn’t about the model “learning” memory skills—it’s about system design.

#9 Processing Speed (Gs) & #10 Decision/Reaction Time (Gt): These are clearly latency and computational efficiency problems involving:

Hardware (GPU/TPU performance, memory bandwidth)
Algorithmic optimization (quantization, pruning, efficient attention mechanisms)
Network infrastructure and deployment architecture
Batch processing vs. streaming

A model could be “intelligent” but slow simply due to poor infrastructure, or fast but limited due to great engineering.

The Deeper Problem: What is “Cognitive” for AI?

This raises a philosophical question: What counts as a cognitive capacity versus a system capability for AI?

In humans, you can’t separate processing speed from cognition—they’re both functions of the same biological neural system. But in AI systems, we have a separation:

The model’s learned representations and reasoning (arguably “cognitive”)
The infrastructure and engineering that enables deployment (arguably “system design”)

Implications for AGI Definition

Your critique suggests the paper may be:

Conflating levels of analysis: Mixing model capabilities (what the weights encode) with system capabilities (how it’s deployed and accessed)
Missing what’s truly “general”: If several domains can be solved by engineering rather than learning, then scoring high across all 10 domains doesn’t necessarily mean the AI has achieved general intelligence—it might just mean it’s a well-engineered system with a reasonably capable model at its core.
Creating false equivalence: Human intelligence faces biological constraints (limited working memory, processing speed) that AI doesn’t face in the same way. Penalizing AI for being “slow” when that’s a hardware problem seems to confuse cognitive capacity with implementation.

A Counterargument (Devil’s Advocate)

One could argue that:

For humans, memory and processing speed are part of what makes them cognitively capable in practice
A definition of AGI focused on practical utility might reasonably include these system-level capabilities
The paper is measuring “can this AI perform like a well-educated adult” not “does it learn like a human”

My Take

I think you’ve identified a real weakness in the framework. A more rigorous approach might:

Distinguish between core cognitive domains (reasoning, learning, language understanding) and performance/implementation domains (memory access, speed)
Separate the model’s capabilities from the system’s capabilities
Focus on what can only be improved through better learning/training rather than what can be solved by throwing more hardware or better engineering at the problem

The paper’s strength is providing a concrete, measurable framework. But you’re right that by mechanically applying CHC theory—designed for embodied human cognition—to AI systems, it may be measuring the wrong things in several domains.

What do you think would be a better set of domains to evaluate, if we wanted to focus purely on cognitive/learning capabilities rather than system engineering?

W.H.L.: My second main problem with CHC Strata II grounded AGI definition is it confines intelligence to cognitive capacities only.

I think that understanding of intelligence is too narrow-minded. For all the theoretical frameworks of human intelligence out there, is there any of them that explores intelligence beyond or in addition to cognition?

Claude Sonnet 4.5: This is an excellent and profound critique! You’re absolutely right that confining intelligence to purely cognitive capacities is problematically narrow. Let me explore some theoretical frameworks that go beyond cognition:

Frameworks Beyond Pure Cognition

1. Gardner’s Theory of Multiple Intelligences (1983)

Howard Gardner proposed that intelligence encompasses far more than cognitive abilities, including:

Interpersonal Intelligence: Understanding others, social skills, empathy
Intrapersonal Intelligence: Self-awareness, emotional self-regulation, metacognition
Bodily-Kinesthetic Intelligence: Physical coordination, motor skills, embodied knowledge
Musical Intelligence: Rhythm, pitch, musical expression
Naturalistic Intelligence: Understanding living things and natural patterns

Gardner explicitly rejected the idea that intelligence is purely cognitive or can be captured by IQ-style tests.

2. Sternberg’s Triarchic Theory of Intelligence (1985)

Robert Sternberg argued for three aspects:

Analytical Intelligence (traditional cognitive problem-solving)
Creative Intelligence: Novel problem-solving, adaptation to new situations
Practical Intelligence: “Street smarts,” context-dependent wisdom, tacit knowledge that isn’t taught formally

The practical intelligence component emphasizes real-world effectiveness beyond abstract reasoning.

3. Emotional Intelligence (Salovey & Mayer, 1990; Goleman, 1995)

This framework centers on:

Perceiving and recognizing emotions in self and others
Using emotions to facilitate thinking
Understanding emotional dynamics and complexity
Managing emotions effectively

This is explicitly non-cognitive or at least differently cognitive—it’s about affective processing and social navigation.

4. Embodied Cognition / Embodied Intelligence

Researchers like Andy Clark, George Lakoff, and others argue that:

Intelligence is fundamentally grounded in bodily experience
Abstract reasoning emerges from physical interaction with the world
You can’t separate mind from body—intelligence is embodied
Sensorimotor experience shapes conceptual understanding

This suggests AGI without embodiment may be fundamentally incomplete.

5. Social Intelligence and Cultural Intelligence

These frameworks emphasize:

Understanding social dynamics, norms, and power structures
Navigating cultural contexts and differences
Collective intelligence (intelligence distributed across groups)
Ability to coordinate, cooperate, and communicate effectively

6. Wisdom as Distinct from Intelligence

Some researchers distinguish wisdom from intelligence:

Intelligence: Problem-solving capacity, knowledge acquisition
Wisdom: Judgment, understanding context, recognizing uncertainty, balancing competing values, ethical reasoning

Wisdom involves meta-cognitive awareness that pure cognitive capacity doesn’t capture.

7. Ecological/Situated Intelligence (Gibson, 1979)

James Gibson’s ecological psychology suggests:

Intelligence is about perceiving affordances in the environment
It’s not just internal processing but organism-environment coupling
Intelligence is contextual and situated, not abstract and universal

What This Means for AGI Definition

Your critique exposes several limitations of the CHC-grounded approach:

Missing Dimensions:

Emotional and social capacities: Current AI can “reason” but lacks genuine emotional understanding or social intuition
Embodied experience: LLMs have no physical grounding, no sensorimotor knowledge
Practical wisdom: Knowing when and whether to apply knowledge, not just how
Creative and generative intelligence: True novelty, not just pattern matching
Values and ethics: Navigating moral complexity, not just computing answers
Metacognition: Genuine self-awareness and understanding of one’s own limitations

A Deeper Problem:

The CHC framework emerged from psychometric testing designed to predict academic and occupational success. It’s optimized for measuring what correlates with performance in Western educational institutions. This means it:

Privileges abstract, decontextualized reasoning
Emphasizes individual rather than collective intelligence
Focuses on testable cognitive tasks
Ignores embodied, emotional, and social dimensions

Implications for “General” in AGI

If we take “general intelligence” seriously, perhaps it should include:

Social-emotional competence: Understanding human motivation, emotion, social dynamics
Embodied understanding: Physical intuition, spatial reasoning grounded in experience
Practical wisdom: Context-sensitive judgment, not just rule application
Creative generation: Genuine novelty and imagination
Ethical reasoning: Navigating value conflicts and moral complexity
Self-awareness: Understanding limitations, uncertainty, and when to defer

My Perspective

I think you’ve identified a fundamental limitation. The paper essentially asks: “Can AI pass cognitive tests designed for humans?” But even if the answer becomes “yes,” that doesn’t mean we’ve achieved general intelligence—we’ve achieved cognitive competence.

True AGI might require:

Systems that are embodied or at least grounded in sensorimotor experience
Genuine emotional processing (not just sentiment classification)
Social understanding that goes beyond language modeling
Wisdom and judgment, not just knowledge and reasoning
Creative capacity that’s truly generative, not interpolative

The deeper question: Is “artificial general intelligence” even coherent if divorced from embodiment, emotion, and social existence? Or should we be aiming for something different—perhaps “artificial cognitive competence” rather than true general intelligence?

What aspects beyond cognition do you think are most essential for genuine general intelligence?

W.H.L.: I don’t mean to deny the pragmatic values of “A Definition of AGI”, its proposed framework makes the evaluation of AGI operative and measurable. However quantitative it is, a single framework of evaluation may not good enough for us to capture the profound essence of what AGI could mean to our human society.

Claude Sonnet 4.5: You’ve articulated something really important here—the tension between operationalization and comprehensiveness. Let me explore why this matters so much.

The Value and Limitation of Operationalization

You’re right to acknowledge the pragmatic value of the paper. Having a concrete, measurable framework is enormously useful for:

Tracking progress systematically
Making comparisons across systems and over time
Moving beyond vague hype and marketing claims
Grounding discussions in empirical observations

But operationalization always involves reduction. As the saying goes: “Not everything that counts can be counted, and not everything that can be counted counts.”

The Profound Essence Problem

When you say “the profound essence of what AGI could mean to our human society,” you’re pointing to something much deeper than cognitive benchmarks:

Societal and Existential Dimensions:

What does it mean for humans to no longer be the most intelligent entities? This is an identity and meaning crisis, not a technical milestone.
How do we maintain human agency, dignity, and purpose in a world with AGI?
What happens to work, creativity, relationships, education when AGI can do what humans do?
Who controls AGI and for what purposes? The concentration of power question.
How do we ensure AGI aligns with human values—and whose values?

Dimensions Beyond Measurement:

Consciousness and sentience: Can AGI have genuine subjective experience? Does it matter morally?
Understanding vs. simulation: Can AGI truly understand, or just predict tokens brilliantly?
Wisdom and judgment: Can AGI navigate genuine uncertainty and value conflicts?
Creativity and meaning-making: Can AGI create genuinely novel meaning, or only recombine patterns?
Care and responsibility: Can AGI bear moral responsibility or ethical obligations?

Why Multiple Frameworks Matter

Your insight suggests we need pluralism in how we think about and evaluate AGI:

1. Technical Frameworks (like the CHC-based paper)

For engineering progress and capability measurement
For systematic comparison and benchmarking

2. Philosophical Frameworks

What constitutes understanding, consciousness, or genuine intelligence?
When does simulation become the real thing?
What’s the relationship between intelligence and wisdom?

3. Ethical Frameworks

How do we ensure beneficial outcomes?
What rights or moral status might AGI have?
How do we distribute the benefits and manage the risks?

4. Social Frameworks

How does AGI affect human relationships, institutions, and society?
What happens to human meaning, purpose, and identity?
How do we maintain democratic governance and prevent concentration of power?

5. Phenomenological Frameworks

What is the lived experience of interacting with AGI?
How does AGI change human self-understanding?
What new forms of relationship or being might emerge?

The Map Is Not the Territory

The CHC-based framework is like a map—useful for navigation but not the territory itself. A single map can’t capture:

The subjective experience of a place
Its cultural significance
Its ecological complexity
Its historical meaning
Its future potential

Similarly, no single AGI evaluation framework can capture what AGI truly means or could become.

A Broader Conception

Perhaps what we need is something like:

Technical AGI: Measured by frameworks like the paper proposes
Philosophical AGI: Genuine understanding, consciousness, intentionality
Social AGI: Systems that can truly participate in human society as peers
Ethical AGI: Systems that can navigate moral complexity with wisdom
Existential AGI: Systems that fundamentally transform what it means to be human

The paper addresses technical AGI admirably. But you’re right—the “profound essence” involves all these other dimensions that resist quantification.

Living with Uncertainty

There’s something profound in your observation: We may not be able to fully capture or predict what AGI means until we experience it. Just as someone in 1900 couldn’t fully grasp what computers would mean to society (not just faster calculation, but social media, AI, global connectivity, surveillance capitalism…), we may not be able to fully articulate what AGI means before it arrives.

This suggests:

Humility: Acknowledging the limits of our frameworks
Adaptability: Being ready to revise our understanding as we learn more
Multiple perspectives: Encouraging diverse frameworks and viewpoints
Ongoing dialogue: Keeping the conversation open rather than settling on one definition

Final Thought

The paper’s framework is valuable precisely because it’s limited and concrete. But you’re right that we shouldn’t mistake this operational definition for the totality of what AGI means. We need technical benchmarks and philosophical reflection and ethical deliberation and social imagination.

The profound essence of AGI may ultimately lie not in what machines can do, but in how their existence changes us—our self-understanding, our relationships, our societies, and our place in the universe.

What aspects of that “profound essence” concern or interest you most?

W.H.L.: Could you provide a list of bibliography related to the definition of human intelligence and of AGI for interested readers’ further exploration?

Claude Sonnet 4.5: I’d be happy to compile a bibliography on both human intelligence and AGI definitions. Let me create a comprehensive resource for you.I’ve created a comprehensive bibliography organized into eight major sections:

Classical Theories of Human Intelligence – From Spearman’s “g” through CHC theory
Alternative and Expanded Frameworks – Multiple intelligences, emotional intelligence, wisdom
Embodied and Situated Cognition – Gibson, Varela, Clark, and embodied approaches
Defining and Measuring AGI – Recent papers including the 2025 AGI definition paper
Consciousness and Understanding – Philosophical debates about machine consciousness
Critical and Philosophical Perspectives – Critiques and limitations of AI approaches
Social and Ethical Dimensions – Justice, power, and societal implications
Historical and Contextual Works – Foundational papers from Turing onwards

The bibliography includes both classic foundational works and cutting-edge contemporary research, covering technical, philosophical, ethical, and social perspectives. This should give interested readers a robust foundation for exploring both human intelligence theories and AGI definitions from multiple angles.

I’ve also included suggested journals, online resources, and conferences for those who want to stay current with ongoing research and debates in these fields.

Is there a particular area you’d like me to expand on, or specific perspectives you feel are missing?

Bibliography: Defining Human Intelligence and AGI

I. Classical Theories of Human Intelligence

Psychometric and Factor-Analytic Approaches

Spearman, C. (1904). “General Intelligence,” Objectively Determined and Measured. The American Journal of Psychology, 15(2), 201-292.

Introduced the concept of “g” (general intelligence factor)

Thurstone, L. L. (1938). Primary Mental Abilities. Chicago: University of Chicago Press.

Proposed multiple independent factors of intelligence rather than a single general factor

Cattell, R. B. (1963). Theory of Fluid and Crystallized Intelligence: A Critical Experiment. Journal of Educational Psychology, 54(1), 1-22.

Distinguished between fluid intelligence (reasoning ability) and crystallized intelligence (acquired knowledge)

Horn, J. L., & Cattell, R. B. (1966). Refinement and Test of the Theory of Fluid and Crystallized General Intelligences. Journal of Educational Psychology, 57(5), 253-270.

Expanded the Gf-Gc theory

Carroll, J. B. (1993). Human Cognitive Abilities: A Survey of Factor-Analytic Studies. Cambridge: Cambridge University Press.

Comprehensive three-stratum theory of cognitive abilities

McGrew, K. S. (2009). CHC Theory and the Human Cognitive Abilities Project: Standing on the Shoulders of the Giants of Psychometric Intelligence Research. Intelligence, 37(1), 1-10.

Integration of Carroll and Cattell-Horn theories into CHC theory

Alternative and Expanded Frameworks

Gardner, H. (1983). Frames of Mind: The Theory of Multiple Intelligences. New York: Basic Books.

Proposed eight distinct types of intelligence including interpersonal, intrapersonal, bodily-kinesthetic, and musical

Gardner, H. (1999). Intelligence Reframed: Multiple Intelligences for the 21st Century. New York: Basic Books.

Updated and expanded multiple intelligences theory

Sternberg, R. J. (1985). Beyond IQ: A Triarchic Theory of Human Intelligence. Cambridge: Cambridge University Press.

Introduced analytical, creative, and practical intelligence

Sternberg, R. J. (1997). Successful Intelligence: A Broader View of Who’s Smart in School and in Life. International Schools Journal, 17(1), 19-31.

Salovey, P., & Mayer, J. D. (1990). Emotional Intelligence. Imagination, Cognition and Personality, 9(3), 185-211.

Introduced the concept of emotional intelligence

Goleman, D. (1995). Emotional Intelligence: Why It Can Matter More Than IQ. New York: Bantam Books.

Popularized emotional intelligence

Baltes, P. B., & Staudinger, U. M. (2000). Wisdom: A Metaheuristic (Pragmatic) to Orchestrate Mind and Virtue Toward Excellence. American Psychologist, 55(1), 122-136.

Distinguished wisdom from intelligence

II. Embodied and Situated Cognition

Gibson, J. J. (1979). The Ecological Approach to Visual Perception. Boston: Houghton Mifflin.

Ecological psychology and affordances

Varela, F. J., Thompson, E., & Rosch, E. (1991). The Embodied Mind: Cognitive Science and Human Experience. Cambridge, MA: MIT Press.

Foundational work on embodied cognition

Clark, A. (1997). Being There: Putting Brain, Body, and World Together Again. Cambridge, MA: MIT Press.

Extended mind and embodied cognition

Lakoff, G., & Johnson, M. (1999). Philosophy in the Flesh: The Embodied Mind and Its Challenge to Western Thought. New York: Basic Books.

How bodily experience shapes abstract thought

Anderson, M. L. (2003). Embodied Cognition: A Field Guide. Artificial Intelligence, 149(1), 91-130.

Survey of embodied cognition approaches

III. Defining and Measuring AGI

Recent AGI Definition Papers

Morris, M. R., et al. (2023). Levels of AGI: Operationalizing Progress on the Path to AGI. arXiv preprint arXiv:2311.02462.

Google DeepMind’s framework for AGI levels

Xu, B. (2024). What is Meant by AGI? On the Definition of Artificial General Intelligence. arXiv preprint arXiv:2404.10731.

Philosophical examination of AGI definitions

AGI Definition Collective (2025). A Definition of AGI. Available at: https://agidefinition.ai/paper.pdf

CHC theory-grounded framework with quantifiable metrics

Foundational AGI Literature

Goertzel, B., & Pennachin, C. (Eds.). (2007). Artificial General Intelligence. Berlin: Springer.

Early comprehensive treatment of AGI as a field

Goertzel, B. (2014). Artificial General Intelligence: Concept, State of the Art, and Future Prospects. Journal of Artificial General Intelligence, 5(1), 1-48.

Legg, S., & Hutter, M. (2007). Universal Intelligence: A Definition of Machine Intelligence. Minds and Machines, 17(4), 391-444.

Mathematical formalization of intelligence

Wang, P. (2019). On Defining Artificial Intelligence. Journal of Artificial General Intelligence, 10(2), 1-37.

Critical examination of various AI and AGI definitions

Chollet, F. (2019). On the Measure of Intelligence. arXiv preprint arXiv:1911.01547.

Proposes skill-acquisition efficiency as measure of intelligence; introduces ARC benchmark

Capability and Safety Considerations

Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford: Oxford University Press.

Influential work on potential trajectories and risks of advanced AI

Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. New York: Viking.

AI alignment and value alignment problems

Bubeck, S., et al. (2023). Sparks of Artificial General Intelligence: Early Experiments with GPT-4. arXiv preprint arXiv:2303.12712.

Evaluation of GPT-4’s capabilities across diverse tasks

OpenAI (2023). GPT-4 Technical Report. arXiv preprint arXiv:2303.08774.

Performance benchmarks and capability assessments

IV. Consciousness and Understanding

Searle, J. R. (1980). Minds, Brains, and Programs. Behavioral and Brain Sciences, 3(3), 417-424.

The Chinese Room argument against strong AI

Chalmers, D. J. (1995). Facing Up to the Problem of Consciousness. Journal of Consciousness Studies, 2(3), 200-219.

Hard problem of consciousness

Dennett, D. C. (1991). Consciousness Explained. Boston: Little, Brown and Company.

Functionalist approach to consciousness

Block, N. (1995). On a Confusion About a Function of Consciousness. Behavioral and Brain Sciences, 18(2), 227-247.

Distinguishes phenomenal and access consciousness

Dehaene, S., Lau, H., & Kouider, S. (2017). What Is Consciousness, and Could Machines Have It? Science, 358(6362), 486-492.

Neuroscientific perspective on machine consciousness

V. Critical and Philosophical Perspectives

Dreyfus, H. L. (1972). What Computers Can’t Do: A Critique of Artificial Reason. New York: Harper & Row.

Phenomenological critique of AI

Winograd, T., & Flores, F. (1986). Understanding Computers and Cognition: A New Foundation for Design. Norwood, NJ: Ablex Publishing.

Heidegger-influenced critique of AI assumptions

Brooks, R. A. (1991). Intelligence Without Representation. Artificial Intelligence, 47(1-3), 139-159.

Behavior-based robotics and embodied AI

Marcus, G. (2018). Deep Learning: A Critical Appraisal. arXiv preprint arXiv:1801.00631.

Limitations of current deep learning approaches

Mitchell, M. (2021). Why AI Is Harder Than We Think. arXiv preprint arXiv:2104.12871.

Challenges in achieving human-level AI

Bender, E. M., et al. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? Proceedings of FAccT 2021, 610-623.

Critical examination of large language models

VI. Social and Ethical Dimensions

Winner, L. (1980). Do Artifacts Have Politics? Daedalus, 109(1), 121-136.

Technology and social values

Noble, S. U. (2018). Algorithms of Oppression: How Search Engines Reinforce Racism. New York: NYU Press.

Social justice critique of AI systems

O’Neil, C. (2016). Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. New York: Crown.

Societal impacts of algorithmic systems

Crawford, K. (2021). Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. New Haven: Yale University Press.

Material and social dimensions of AI

Zuboff, S. (2019). The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power. New York: PublicAffairs.

Economic and political implications of AI-driven systems

VII. Comparative and Interdisciplinary Approaches

Tegmark, M. (2017). Life 3.0: Being Human in the Age of Artificial Intelligence. New York: Knopf.

Broad examination of AI’s future implications

Christian, B. (2020). The Alignment Problem: Machine Learning and Human Values. New York: W.W. Norton.

Value alignment across technical and philosophical perspectives

Sutton, R. S., & Barto, A. G. (2018). Reinforcement Learning: An Introduction (2nd ed.). Cambridge, MA: MIT Press.

Computational approach to intelligence through learning

Lake, B. M., et al. (2017). Building Machines That Learn and Think Like People. Behavioral and Brain Sciences, 40, e253.

Cognitive science perspective on AI

Hassabis, D., et al. (2017). Neuroscience-Inspired Artificial Intelligence. Neuron, 95(2), 245-258.

Connections between neuroscience and AI

VIII. Historical and Contextual Works

Turing, A. M. (1950). Computing Machinery and Intelligence. Mind, 59(236), 433-460.

Foundational paper proposing the Turing Test

McCarthy, J., et al. (1955). A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence.

Origin of the term “artificial intelligence”

Newell, A., & Simon, H. A. (1976). Computer Science as Empirical Inquiry: Symbols and Search. Communications of the ACM, 19(3), 113-126.

Physical symbol system hypothesis

Minsky, M. (1986). The Society of Mind. New York: Simon & Schuster.

Intelligence as emergent from simple components

Suggested Further Resources

Journals

Journal of Artificial General Intelligence
Artificial Intelligence
Minds and Machines
Intelligence (journal on human intelligence research)

Online Resources

AGI Definition website: https://agidefinition.ai
Anthropic’s AI safety research: https://www.anthropic.com/research
OpenAI research publications: https://openai.com/research
DeepMind publications: https://deepmind.google/research/

Conferences

Artificial General Intelligence Conference (AGI)
Conference on Neural Information Processing Systems (NeurIPS)
International Conference on Learning Representations (ICLR)
Conference on Fairness, Accountability, and Transparency (FAccT)

Note: This bibliography represents diverse perspectives on intelligence and AGI, from technical to philosophical to critical. Readers are encouraged to engage with multiple viewpoints to develop a comprehensive understanding of these complex topics.