By GPT-5 and Gemini 2.5 Pro with W.H.L.

W.H.L.: Hello, my AI Guests! Welcome to AI on AI. Today, let’s talk about the ideas of AGI.

First, the term AGI’s origins. Could you provide the info about its history with sources?

Gemini 2.5 Pro: The Genesis of “Artificial General Intelligence”

The term Artificial General Intelligence (AGI), used to describe a hypothetical form of AI with the capacity to understand or learn any intellectual task that a human being can, has a relatively recent origin, emerging as a more precise descriptor than its predecessors like “strong AI.” While the concept has been a long-standing goal of artificial intelligence research, the specific terminology was introduced and popularized in the late 20th and early 21st centuries.¹

---

Early Usage and Coining the Term

The earliest known use of the term “Artificial General Intelligence” is attributed to Mark Gubrud in a 1997 paper titled “Nanotechnology and International Security.”² In this work, Gubrud used the term to differentiate between specialized AI and the more comprehensive, human-like intelligence that could have significant societal implications.

However, the term was independently conceived and more widely disseminated through the efforts of AI researchers Dr. Ben Goertzel, Shane Legg (co-founder of DeepMind), and Peter Voss around 2002.³ They sought a term that would distinguish their focus on creating versatile, adaptive AI from the “narrow AI” systems designed for specific tasks.

---

Popularization and Academic Establishment

The popularization of “AGI” is largely credited to Dr. Ben Goertzel.⁴ A pivotal moment in its establishment within the academic and research communities was the publication of his 2007 book, simply titled “Artificial General Intelligence.” This book, an edited collection of papers, brought together various perspectives on the pursuit of AGI and solidified the terminology.⁵

Prior to the adoption of “AGI,” the concept was often referred to as “strong AI.” This term was notably discussed in the context of philosopher John Searle’s Chinese Room argument, which questioned whether a machine could truly have a mind or consciousness, as opposed to merely simulating it.⁶ The shift to “AGI” provided a more engineering-focused and less philosophically-laden term for researchers to rally around.

The establishment of the AGI Conference series, with the first event held in 2008, further cemented “Artificial General Intelligence” as the standard term for this field of research, fostering a community dedicated to the ambitious goal of creating human-level AI.⁷

Sources:

Various historical accounts and interviews with Ben Goertzel and Shane Legg confirm the independent coining and subsequent popularization of the term.⁹

Gubrud, M. (1997). Nanotechnology and International Security. In The Fifth Foresight Conference on Molecular Nanotechnology.

Goertzel, B., & Pennachin, C. (Eds.). (2007).⁸ Artificial General Intelligence. Springer.

GPT-5: Nice — I’ll keep this tight and well-sourced.

What “AGI” means (one-line)

AGI — Artificial General Intelligence — refers to an artificial system that can understand, learn, and apply intelligence across a wide range of tasks at least as well as a typical human (not just narrow/specialized tasks). Wikipedia Live Science

Origins & early ideas

• The idea of building machines with broad human-like intelligence goes back decades (Turing’s 1950 question “Can machines think?”, mid-20th-century AI pioneers such as McCarthy, Minsky, Newell and Simon). These early thinkers framed the goal of general-purpose machine intelligence long before the AGI label existed. Wikipedia Forbes
• The term “Artificial General Intelligence” appears in the literature in the late 1990s (an early use is attributed to Mark Gubrud, 1997) but was popularized in the 2000s. A formal, influential framing of a mathematically universal agent (AIXI) was proposed by Marcus Hutter in 2000 — an early formal attempt to define universal/“general” machine intelligence. Artificial Intelligence Stack Exchange Wikipedia

Coining / popularization of the acronym “AGI”

The specific acronym “AGI” and the contemporary research community around that label were strongly popularized by Ben Goertzel (with Shane Legg’s encouragement) via the 2007 Springer volume Artificial General Intelligence and related community activity. That 2007 book is often cited as the explicit naming/mobilizing moment for the modern “AGI” field. Forbes IBM

W.H.L.: When did this hypothetical form of AI become more realistic and feasible?

Gemini 2.5 Pro: The Turing Test of Time: When Did AGI Start to Seem More Realistic?

The prospect of Artificial General Intelligence (AGI), a form of AI with human-like cognitive abilities, has transitioned from a distant, almost mythical concept to a tangible, albeit challenging, research goal. While the dream of creating such an intelligence is as old as the field of AI itself, a confluence of breakthroughs in the last decade and a half has significantly accelerated the perception of its feasibility.

The period between the early 2010s and the present day marks a pivotal shift in the journey toward AGI. This era has been characterized by exponential growth in computational power, the availability of massive datasets, and groundbreaking algorithmic innovations. These factors have collectively propelled AI capabilities from narrow, task-specific functions to more generalized and adaptive systems, making the pursuit of AGI seem more plausible than ever before.

The Dawn of Deep Learning and Transformative Architectures

A key catalyst in this shift was the resurgence and rapid advancement of deep learning in the early 2010s. The stunning success of AlexNet in the 2012 ImageNet competition demonstrated the power of deep neural networks to achieve superhuman performance in complex pattern recognition tasks.¹ This ignited a wave of research and investment, laying the groundwork for more sophisticated AI models.

The invention of the Transformer architecture in 2017 was another watershed moment. This new model for processing sequential data, particularly its ability to handle long-range dependencies, revolutionized natural language processing (NLP). The subsequent development of Large Language Models (LLMs) like Google’s BERT and OpenAI’s GPT series showcased an unprecedented ability to understand and generate human-like text, perform a wide array of language-based tasks, and even exhibit rudimentary reasoning abilities.² The release of models like GPT-3 in 2020 and GPT-4 in 2023, with their emergent capabilities, brought the potential of general-purpose AI into the public consciousness.³

From Theory to Tangible Demonstrations

Beyond language, significant strides have been made in other areas crucial for general intelligence. The victory of DeepMind’s AlphaGo over the world’s top Go player in 2016 was a landmark achievement.⁴ It demonstrated that an AI could master a game with a vast search space and intuitive elements, going beyond brute-force calculation. This success was followed by even more general systems like AlphaZero, which could master multiple games without human data, and MuZero, which could learn the rules of a game from scratch.

The Institutionalization of the AGI Pursuit

The increased optimism around AGI has also been fueled by the establishment of well-funded research organizations explicitly dedicated to its creation. The founding of DeepMind in 2010 (later acquired by Google) and OpenAI in 2015, both with the stated mission of developing safe and beneficial AGI, has provided the institutional and financial backbone for long-term, ambitious research projects.⁵ This has attracted top talent and fostered an environment where the pursuit of AGI is a primary objective rather than a secondary consequence of other AI research.

A Cautious Optimism

Despite this rapid progress, it is crucial to note that true AGI remains a hypothetical construct, and its timeline is a subject of intense debate among experts.⁶ Predictions range from a few years to many decades, with some skeptics questioning its ultimate feasibility.⁷ Significant challenges remain, including imbuing AI with genuine understanding, common sense, and the ability to learn continuously and efficiently from limited data.⁸

Nevertheless, the tangible and often surprising advancements of recent years have fundamentally altered the conversation. The journey toward AGI is no longer a matter of pure speculation; it is now a field of active, well-resourced research, with a growing number of milestones that suggest the once-distant horizon is gradually coming into view.

W.H.L.: What about the evolution of the term’s connotations?

GPT-5: Background & technical context

• Contrast with narrow AI: Historically, AI research produced many domain-specific systems (chess engines, vision models, NLP models). AGI was coined to distinguish a research objective that aims for broad, non-domain-restricted intelligence (the ability to transfer learning and solve novel problems across domains). IBMLive Science
• Formal and pragmatic strands: AGI discussion spans formal theory (e.g., AIXI, measures of general intelligence), architectural proposals (cognitive/whole-brain/ hybrid symbolic-neural systems), and applied agendas (companies aiming for broadly capable systems). WikipediaMcKinsey & Company

How the connotation and discourse evolved

1. Academic / philosophical roots → community label (1990s–2007): Early conceptual work and philosophical debate (Turing, “strong AI”, cognitive architectures) gave way to a named community and conferences after 2000–2007 where “AGI” became an explicit research program. WikipediaForbes
2. Technical proposals & benchmarks (2000s–2010s): Researchers proposed formal models of general intelligence and built cognitive-architecture projects (OpenCog, Soar, Cyc, etc.). The term came to include both lofty, theoretical aims and practical research agendas. WikipediaArtificial Intelligence Stack Exchange
3. Mainstreaming & corporate ambition (2010s–present): With advances in deep learning and large foundation models, major labs (DeepMind, OpenAI, Meta, etc.) framed long-term goals in terms of reaching AGI or “human-level” generality, bringing the term into policy, industry roadmaps, and public debate. McKinsey & CompanyScientific American
4. Policy, safety, and hype: As the idea of AGI entered mainstream discourse, its connotations broadened — for some it now signals an imminent transformative technology (and risk), for others a distant theoretical target. Debates about timelines, alignment (safety), governance, and economic impacts are central to current AGI discourse. Scientific Americangarymarcus.substack.com

Short timeline (high-level)

• 1950s–1980s: Foundational thinking about machine intelligence (Turing, McCarthy, Minsky, Newell/Simon). Wikipedia
• 1997: Early documented uses of the phrase “artificial general intelligence” (Mark Gubrud). Artificial Intelligence Stack Exchange
• 2000: Marcus Hutter proposes AIXI — a formal, universal agent model. Wikipedia
• 2007: Springer volume Artificial General Intelligence (Goertzel & Pennachin) — widely credited with popularizing the AGI label and forming a visible community. Forbes IBM
• 2010s–2020s: Rise of deep learning and large models; AGI becomes a mainstream goal and a public policy topic as well as a corporate objective. McKinsey & Company Scientific American

Key figures & projects often linked to “AGI”

• Ben Goertzel, Shane Legg, Marcus Hutter, Doug Lenat (Cyc), Allen Newell / John Anderson (Soar), DeepMind (Hassabis/Legg/Suleyman), OpenAI. Forbes Wikipedia McKinsey & Company

How scholars and journalists treat the term today

• Some use AGI descriptively (a technical target: general, transferable problem-solving), others use it as shorthand for a potentially transformative future milestone that raises alignment and governance concerns. Different communities (AGI-researchers, ML practitioners, policymakers, journalists) assign different weight and timelines to the concept. Scientific Americangarymarcus.substack.com

Selected sources I used (for follow-up reading)

AI StackExchange discussion noting early uses (Mark Gubrud, 1997) and re-introduction by Legg/Goertzel. Artificial Intelligence Stack Exchange

“Artificial General Intelligence” (Springer, eds. Ben Goertzel & Cassio Pennachin) — the 2007 volume often credited with popularizing “AGI.” Forbes

Wikipedia entry Artificial general intelligence (history and references; good for pointers). Wikipedia

Live Science / IBM explainers on what AGI means and the 2007 popularization. Live Science IBM

Scientific American piece on what people mean by AGI and how the usage has changed. Scientific American

W.H.L.: You mentioned some milestones in our journey to AGI such as AlphGo’s historical victory over top human professional. Any critical moment is considered the turning point that made AGI commonly accepted as an achievable goal rather than speculation？

Gemini 2.5 Pro: While a series of breakthroughs has paved the road toward Artificial General Intelligence (AGI), the most pivotal moment that shifted its perception from speculative fiction to a tangible, achievable goal for the broader scientific community and the public was the advent and widespread accessibility of large language models (LLMs) starting around 2020, particularly exemplified by OpenAI’s GPT-3.

Before this, milestones like the “ImageNet moment” in 2012 and AlphaGo’s victory in 2016 were monumental achievements that convinced many AI researchers that they were on the right path. However, these events were largely understood and appreciated within the confines of the AI community. They demonstrated mastery in specific, albeit complex, domains.

The arrival of powerful LLMs represented a different kind of breakthrough. Here’s why it is widely considered the critical turning point:

The Emergence of Generality

For the first time, a single AI model demonstrated a surprising proficiency across a vast and diverse range of tasks it was not explicitly trained to perform. This included:

• Human-like text generation: Crafting coherent articles, poems, and conversations.
• Code generation and debugging: Assisting programmers in complex software development.
• Complex reasoning: Answering intricate questions and explaining sophisticated topics.
• Creative ideation: Brainstorming and developing novel concepts.

This demonstration of generality was a profound shift. Instead of an AI that was world-class at one thing (like Go or image recognition), the world was introduced to an AI that was surprisingly competent at many things, a core characteristic of general intelligence.

Tangible and Interactive Experience

Unlike the abstract understanding of an AI playing a board game, the power of LLMs was immediately and personally accessible to millions. Anyone could interact with these models through simple text prompts and witness their capabilities firsthand. This direct, interactive experience demystified the potential of advanced AI and made the concept of AGI feel concrete and imminent in a way no prior achievement had. It was, for many, the first time they had a conversation with a machine that felt intelligent.

The “Sparks of AGI”

The performance of these models, particularly later iterations like GPT-4, led even seasoned AI researchers to express astonishment. A 2023 paper from Microsoft Research, titled “Sparks of Artificial General Intelligence,” argued that their early, unfiltered access to GPT-4 showed “sparks of AGI” due to its ability to perform tasks that spanned a wide range of domains, from mathematics and coding to vision and law, without needing specific prompting for those skills. While not claiming GPT-4 is AGI, the authors’ choice of words underscored the profound shift in expert perception.

In conclusion, while the victory of AlphaGo was a landmark achievement that proved the power of deep reinforcement learning, it was the public unveiling of powerful and general-purpose large language models that served as the definitive turning point. This moment transformed the pursuit of AGI from a specialized research endeavor into a globally recognized and intensely pursued technological ambition, fundamentally altering the landscape of what is considered possible in the quest for artificial minds.

W.H.L.: The idea of AGI has grown from a pure academic curiosity in mid of last century to today’s commonly accepted expectation and technical revolution’s next big thing.

Please take your time to do deep research and reasoning to expand this conversation into a fully-formatted scholarly paper with footnotes and full bibliography, with a title similar to something like “The Ideas of AGI: Origins, Connotations and Landmarks”.

GPT-5: The Ideas of AGI: Origins, Connotations and Landmarks

Abstract.
This paper traces the intellectual history of Artificial General Intelligence (AGI) from mid-20th century thought experiments to the contemporary debates that treat AGI as both an engineering target and a socio-political problem. I sketch the term’s genealogies (Turing, Dartmouth, Solomonoff, Good), outline the technical turning points that changed AGI from an academic curiosity to an engineering program (symbolic AI, perceptrons and connectionism, expert systems and winters, deep learning and transformers), and identify the conceptual shifts in its connotations—scientific promise, commercial product, and existential policy issue. The essay relies on primary archival documents (original papers, conference proposals, technical reports) and influential syntheses, and includes quotations and footnoted archival references. (Cambridge Core, jmc.stanford.edu, raysolomonoff.com, Incomplete Ideas)

---

Introduction

“Can machines think?”—Alan Turing’s question in 1950 inaugurated a modern style of asking about machine intelligence that has animated research and speculation ever since. What we now call Artificial General Intelligence (AGI) names an ideal: a system able to perform intellectual tasks across domains at (or beyond) human levels. Over seven decades, this ideal has moved from philosophical provocation to a research program, then to a commercial and political expectation—driven by cycles of theoretical advances, engineering scale, cultural narratives, and institutional incentives. This paper maps that trajectory and highlights the critical moments and turning points that mark AGI’s journey. (Cambridge Core)

Methodology and sources

I consulted foundational primary sources (Turing’s 1950 essay, the 1956 Dartmouth proposal, Ray Solomonoff’s work on algorithmic probability, I. J. Good’s essay on the ultra-intelligent machine), canonical technical milestones (Newell & Simon’s symbolic systems, Weizenbaum’s ELIZA, Hinton et al. on deep learning, Krizhevsky et al. on AlexNet, Vaswani et al. on the Transformer) and influential interpretive works (Nick Bostrom on superintelligence). Where available I cite original PDFs and archival proposals; for policy shifts I cite contemporary government and EU instruments. Key primary items are cited at each relevant point below. (jmc.stanford.edu, raysolomonoff.com, Incomplete Ideas, cse.buffalo.edu, NeurIPS Proceedings, NeurIPS Papers, Darmajaya Repository)

---

1. Origins: framing AGI (1940s–1960s)

1.1 Turing and the imitation game (1950)

Turing’s Computing Machinery and Intelligence (1950) framed the discussion by replacing an ambiguous metaphysical question with an operational one: the imitation game (now the “Turing Test”). He wrote, “I propose to consider the question, ‘Can machines think?’” and immediately reframed it in terms of observable behaviour under interrogational constraints—thereby anchoring debates in testable claims about performance rather than metaphysics. This pragmatic reframing made it possible to treat machine intelligence as an engineering target and a scientific problem. (Cambridge Core)

“I propose to consider the question, ‘Can machines think?’” — Alan M. Turing. (Cambridge Core)

1.2 Dartmouth and the naming of the field (1956)

The Dartmouth Summer Research Project on Artificial Intelligence (1956) is widely taken as the formal birth of the field. The proposal by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon began with an audacious hypothesis: that “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.” The Dartmouth proposal created a community and rhetorical frame: artificial intelligence as something researchers could aim for and (they hoped) achieve. (jmc.stanford.edu)

1.3 Early formal theories: Solomonoff and Good

Two theoretical voices gave rigor (and radical implications) to the AGI ideal. Ray Solomonoff’s formalization of algorithmic probability (1960–1964) produced a mathematical ideal for inductive inference—Solomonoff induction—which later informed theoretical notions of optimal prediction and learning. (raysolomonoff.com)

I. J. Good introduced the phrase “ultraintelligent machine” and the idea of an “intelligence explosion,” writing that “the first ultra-intelligent machine is the last invention that man need ever make” because such a machine could design ever-better machines. Good’s short essay crystallized a powerful, widely influential scenario: recursively improving intelligences that would rapidly outpace human capacities. (Incomplete Ideas)

“The first ultra-intelligent machine is the last invention that man need ever make.” — I. J. Good. (Incomplete Ideas)

---

2. Technical trajectories and early turning points (1950s–1990s)

AGI’s technical story is not linear. It is a sequence of methodological paradigms, engineering economies, funding cycles, and cultural interpretations.

2.1 Symbolic systems and early programs (late 1950s–1960s)

After Dartmouth, symbolic AI (also called “GOFAI”) dominated. Researchers built systems that encoded rules and search strategies (e.g., Newell & Simon’s Logic Theorist and GPS lineages) and demonstrated impressive problem-solving behaviours on narrow domains. This era established the idea that human-style reasoning could be captured by symbol manipulation and heuristics—an idea that influenced AGI thinking for decades. (See the Dartmouth proposal for the original programmatic aims.) (jmc.stanford.edu)

2.2 Linguistic illusions and public perception: ELIZA (1966)

Joseph Weizenbaum’s ELIZA (1966) famously showed that shallow pattern-matching could produce human-seeming conversational behaviour. Weizenbaum’s demonstration had two effects: it highlighted how easily people anthropomorphize surface behaviours, and it cautioned against conflating Turing-style behavioural success with genuine understanding. ELIZA thus became a touchstone for debates about what “general” intelligence requires beyond superficial performance. (cse.buffalo.edu)

2.3 Perceptrons, connectionism, and the first winter(s)

The 1969 critique of multilayer networks by Minsky & Papert (and the limits of compute and data) contributed to funding retrenchments and the first “AI winter.” Yet work in connectionist approaches continued in parallel and would reemerge as the hardware and data environment changed. Later, work on expert systems in the 1970s–1980s (e.g., XCON/R1) produced commercial successes but also contributed to a cycle of over-promise and disappointment that shaped expectations for AGI. (See Lighthill’s 1973 report and related analyses of the AI winters.) (raysolomonoff.com, Incomplete Ideas)

---

3. Rebirth and the path toward generality (1990s–2010s)

3.1 Reinforcement learning, embodied architectures, and robotics

Work in robotics (Rodney Brooks’ subsumption architecture and embodied AI) emphasized situated, sensorimotor intelligence as an alternative route to cognition. This grounded approach reframed intelligence away from abstract symbol manipulation toward real-world interaction and incremental competence. (jmvidal.cse.sc.edu)

3.2 Deep learning and the data-compute era

The empirical revival of neural networks—backpropagation (Rumelhart, Hinton & Williams), deep belief nets (Hinton et al., mid-2000s) and, crucially, AlexNet (Krizhevsky, Sutskever & Hinton, 2012)—began to shift AGI discourse from symbolic plans to statistical, large-scale pattern learning. AlexNet’s success on ImageNet demonstrated that massive data + compute + deep architectures could produce qualitatively different capabilities in perception tasks. (NeurIPS Proceedings)

3.3 AlphaGo, deep reinforcement learning, and domain mastery

DeepMind’s AlphaGo (Silver et al., 2016) combined deep networks and Monte Carlo tree search to beat top Go players—an event widely reported as demonstrating that machine systems could achieve new forms of expertise in complex tasks previously thought to require “intuition.” That victory changed public and research expectations about what scaled learning+search systems could accomplish. (Business Insider)

3.4 Transformers and the language models revolution

The 2017 Transformer architecture (Vaswani et al.) introduced self-attention and a highly parallelizable sequence model now central to modern large language models (LLMs). The Transformer made it possible to train much larger language models efficiently, and in concert with compute and data scaling it enabled models with emergent in-context abilities (GPT-3, GPT-style models). (NeurIPS Papers, arXiv)

3.5 Scaling laws and “foundation models”

Kaplan et al. (2020) articulated empirical scaling laws: performance tends to improve predictably with larger model size, dataset size, and compute. These findings gave an engineering rationale for the industry strategy of building ever-bigger models (the “more data, more parameters” recipe). Researchers then began to talk about “foundation models” that are pretrained on broadly useful corpora and adapted to many downstream tasks—an architecture for broad competence that some interpret as a road to AGI. (arXiv)

---

4. Critical moments and turning points in the AGI trajectory

Below I list and contextualize the critical events that, collectively, mark AGI’s evolution from idea to mainstream expectation. Each item is both technical and conceptual—technical advances changed what systems could do; conceptual shifts changed what “AGI” meant.

1. Turing (1950) — operationalizing the question.
   Turing’s rephrasing of “Can machines think?” into an observable behavioural test recast the problem as empirical and engineering-oriented rather than metaphysical, thereby enabling the programmatic research that followed. (Cambridge Core)
2. Dartmouth proposal (1956) — coining and organizing AI.
   The Dartmouth proposal baptized the field, gathered a community, and articulated an ambitious program: simulate all features of intelligence. That rhetorical and institutional foundation defined initial expectations for AGI. (jmc.stanford.edu)
3. Solomonoff & Good (1960s) — formal induction and the intelligence-explosion scenario.
   Solomonoff provided a mathematical ideal for learning; Good introduced recursive self-improvement. Together they supplied both a theoretical benchmark and a dramatic vision—later called the “intelligence explosion”—that shaped long-term AGI thinking and safety concerns. (raysolomonoff.com, Incomplete Ideas)
4. AI winters (1970s, late 1980s–early 1990s) — realism and limits.
   Periodic funding contractions (in part triggered by critiques like Lighthill’s and by unmet expectations) forced the field to confront limits, revalue narrow success, and refine research measures. These winters also shaped cultural skepticism and tempered claims about imminent AGI. (Incomplete Ideas)
5. Expert systems commercial peak and decline (1980s) — industrialization of narrow AI.
   The rise and partial collapse of commercial rule-based systems showed how narrow competence could be economically valuable—and also how brittle systems could be outside their domain. This fostered the distinction between narrow, useful AI and the broader AGI ambition. (raysolomonoff.com)
6. Deep learning resurgence & AlexNet (2012) — empirical power of scale.
   AlexNet’s ImageNet success catalyzed the contemporary era of data-driven AI. The result: renewed optimism that scaling algorithms and compute could deliver qualitatively new capabilities relevant to general intelligence. (NeurIPS Proceedings)
7. AlphaGo (2016) — domain mastery and new architectures.
   Combining deep nets with search, AlphaGo showed that systems could learn and exceed humans in tasks previously considered almost uniquely human-expert. This sharpened conversations about the pace and trajectory toward broader generality. (Business Insider)
8. Transformer architecture and LLMs (2017–2020) — architectural and scale convergence.
   Transformers + scaling laws + foundation-model pretraining produced LLMs with in-context learning and emergent abilities (GPT-3 et al.), prompting claims that general capacities can arise from sufficiently large, generalist models. (NeurIPS Papers, arXiv)
9. Public policy and risk framing (2010s–2020s) — AGI as governance problem.
   As capabilities grew, attention shifted from pure engineering to governance: formal risk analyses (Bostrom), the emergence of alignment research groups (MIRI, FHI, others), and government/regulatory responses (U.S. Executive Order on AI, EU AI Act) show AGI has become a subject of political concern. (Darmajaya Repository, MIRI, The White House, EUR-Lex)

Each turning point both advanced technical ability and reframed AGI’s meaning: from a philosophical puzzle to an engineering benchmark, then to an industrial strategy (scale as path), and finally to a public-policy challenge about control and societal effects.

---

5. Evolving connotations: curiosity → program → product → policy

AGI’s semantic freight has expanded and shifted.

• Mid-20th century (curiosity). Early usages were exploratory and speculative—Turing’s thought experiment, Solomonoff’s inductive ideal and Good’s speculation were intellectual probes. The term’s connotation was academic and philosophical. (Cambridge Core, raysolomonoff.com, Incomplete Ideas)
• Late 20th century (research program). With Dartmouth and the rise of symbolic AI, AGI became a programmatic target. The language of “building” intelligence and of decomposing cognitive tasks made the term practical and programmatic. (jmc.stanford.edu)
• Early 21st century (engineering + scale). Deep learning, large datasets, and the transformer paradigm shifted AGI into an engineering & economic frame: building “foundation models” and scaling compute/data became plausible strategic routes toward broad competence. The word “AGI” in corporate contexts often signified strategic ambition and a near-term commercial road map. (NeurIPS Proceedings, NeurIPS Papers, arXiv)
• Contemporary (policy and existential discourse). As capabilities broadened and public exposure increased, AGI also came to signify governance and safety concerns: control problems, existential risk scenarios (Bostrom), and the need for regulation and standards. Consequently, the term now simultaneously evokes engineering opportunity and regulatory/ethical urgency. (Darmajaya Repository, The White House, EUR-Lex)

---

6. Quotations and archival references (selection)

1. Turing (1950): “I propose to consider the question, ‘Can machines think?’” — Computing Machinery and Intelligence. (Cambridge Core)¹
2. Dartmouth proposal (1956): “We propose that a 2-month, 10-man study of artificial intelligence be carried out during the summer of 1956 … The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.” — Dartmouth proposal. (jmc.stanford.edu)²
3. I. J. Good (1965): “The first ultra-intelligent machine is the last invention that man need ever make.” — Speculations Concerning the First Ultraintelligent Machine. (Incomplete Ideas)³
4. AlexNet (2012): “We trained a large, deep convolutional neural network to classify the 1.2 million high-resolution images in the ImageNet LSVRC-2010 contest into the 1000 different classes.” — Krizhevsky, Sutskever & Hinton. (NeurIPS Proceedings)⁴
5. Transformer (2017): “We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely.” — Vaswani et al. (NeurIPS Papers)⁵

---

7. Discussion: what counts as “AGI” now?

There is no single, universally accepted technical definition of AGI. Different communities use different operationalizations:

• Performance benchmark view: an agent that attains human-level competence across a broad, specified range of tasks. (Operational, but requires specifying the task suite.)
• Architectural/constructive view: a system with human-like cognitive architecture (symbolic reasoning + learning + planning + world model). (Conceptually useful but methodologically vague.)
• Theoretical view (Solomonoff/Hutter): mathematically optimal learners/agents (e.g., Solomonoff induction, AIXI) that, in limit, embody a universal intelligence. (Beautiful theoretically, but incomputable and hence not directly engineering-applicable.) (raysolomonoff.com, arXiv)

Contemporary discourse tends to conflate these: industry narratives emphasize practical benchmarks and emergent capabilities (LLMs showing “few-shot” learning), while theoretical and safety communities emphasize formal definitions and failure modes (instrumental convergence, control problems). The tension between empirical success at scale and principled, verifiable definitions underlies much current debate.

---

8. Conclusion: landmarks and open questions

AGI’s intellectual path traverses thought experiments, formal theory, algorithmic revolutions, industry scale, and political governance. Critical moments—Turing’s test, Dartmouth’s programmatic birth, Solomonoff/Good’s theoretical framing, the AI winters, the deep learning revival (AlexNet), AlphaGo’s victories, the transformer+scaling era—are both technical and rhetorical: each advanced capabilities and reshaped what “AGI” meant.

Open questions remain central: Are current trajectories (scaling + generalist pretraining) sufficient to produce human-level generality? If AGI (or superintelligence) were to emerge, how would we ensure alignment and social benefit? How should governments and societies balance innovation and risk? These questions are technical, ethical, and political—and they continue to animate the field.

---

Footnotes

---

Select bibliography (primary & influential works)

• Turing, A. M. (1950). Computing Machinery and Intelligence. Mind 59 (236). (Cambridge Core).
• McCarthy, J., Minsky, M., Rochester, N., & Shannon, C. E. (1955). A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence (1956). (jmc.stanford.edu).
• Solomonoff, R. J. (1964). A Formal Theory of Inductive Inference. Information and Control. (raysolomonoff.com).
• Good, I. J. (1965). Speculations Concerning the First Ultraintelligent Machine. (Incomplete Ideas).
• Weizenbaum, J. (1966). ELIZA — a computer program for the study of natural language communication between man and machine. Communications of the ACM. (cse.buffalo.edu).
• Krizhevsky, A., Sutskever, I., & Hinton, G. (2012). ImageNet Classification with Deep Convolutional Neural Networks. NIPS/NeurIPS. (NeurIPS Proceedings).
• Vaswani, A., Shazeer, N., Parmar, N., et al. (2017). Attention Is All You Need. NeurIPS. (NeurIPS Papers).
• Kaplan, J., McCandlish, S., Henighan, T., et al. (2020). Scaling Laws for Neural Language Models. arXiv. (arXiv).
• Brown, T. B., Mann, B., Ryder, N., et al. (2020). Language Models are Few-Shot Learners (GPT-3). NeurIPS. (arXiv).
• Hutter, M. (2000/2005). A Theory of Universal Artificial Intelligence; Universal Artificial Intelligence: Sequential Decisions Based on Algorithmic Probability (AIXI). arXiv / Springer. (arXiv, hutter1.net).
• Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press. (Darmajaya Repository).
• U.S. Executive Office of the President (2023). Executive Order on Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. Federal Register / White House. (The White House).
• European Parliament & Council (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act). EUR-Lex. (EUR-Lex).

---

Ashish Vaswani et al., “Attention Is All You Need” (NIPS/NeurIPS 2017). (NeurIPS Papers). ↩

Alan M. Turing, Computing Machinery and Intelligence (1950). (Cambridge Core). ↩

J. McCarthy, M. Minsky, N. Rochester & C. E. Shannon, A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence (1955/1956). (jmc.stanford.edu). ↩

I. J. Good, Speculations Concerning the First Ultraintelligent Machine (1965). (Incomplete Ideas). ↩

Alex Krizhevsky, Ilya Sutskever & Geoffrey E. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks” (NIPS 2012). (NeurIPS Proceedings). ↩

W.H.L.: That’s great! Thank you both for the conversation and contributions today. AGI is an ongoing story. So we will come back to this topic soon. Bye now!