Artificial general intelligence (AGI)

Working Definition

Artificial general intelligence (AGI) refers to a machine system capable of learning and performing a broad range of cognitive tasks at or above the level of a competent human without being re-engineered for each new domain. All three components of this definition are contested independently: breadth (how wide a range of tasks counts as general), depth (measured against an average worker, a specialist, or the pinnacle of human science), and autonomy (whether the system selects its own long-horizon goals or executes prompted instructions). Whereas popular usage equates AGI with “an AI that can do everything,” technical frameworks such as Google DeepMind’s 2023 taxonomy treat it as a multi-level matrix on which existing foundation models already occupy the lower rungs of general capability. In any serious analysis, one must first ask which of those three axes a speaker has in mind, because most disputes over whether AGI is imminent are actually disputes over definitions.

Genesis and Global Lineage

The question of a generally thinking machine is as old as computer science: Alan Turing (1950) framed it through the imitation game, the 1956 Dartmouth workshop assumed every aspect of learning could be mechanized, and I. J. Good (1965) formulated the “ultraintelligent machine.” Yet the compound term artificial general intelligence appeared late and from the margins. Physicist Mark Gubrud used it in 1997 in an essay on automated nanotechnology and military systems, and around 2002 Shane Legg proposed the name to Ben Goertzel and Cassio Pennachin for an edited volume published by Springer in 2007. For Goertzel and the AGI conference series launched in 2008, “general” was a manifesto breaking with post-1990s “narrow AI,” which solved isolated tasks like chess or spam filtering. In 2007, Legg and Marcus Hutter formulated universal intelligence (weighting performance across all computable environments by Kolmogorov/Solomonoff complexity), before Legg co-founded DeepMind in 2010 with the slogan “solve intelligence.” Philosopher Nick Bostrom (Superintelligence, 2014) then cast AGI as the trigger for an existential control crisis, while OpenAI wrote AGI into its 2015 founding charter—turning a philosophical horizon into a clause in multi-billion-dollar commercial contracts, where AGI increasingly means an economic threshold (automating most economically valuable cognitive labor).

Central European Lineage

Half a century before the modern AGI debate crystallized in Silicon Valley and Oxford, the Polish writer and philosopher Stanisław Lem explored the same civilizational horizon in his philosophical treatise Summa technologiae (Wydawnictwo Literackie, Kraków 1964). In Chapter IV, “Intelektronika” (Intellectronics), Lem analyzed a future in which cybernetic “intelligence amplifiers” equal and surpass human cognitive capacity across scientific, administrative, and strategic domains. Dismissing pulp tropes of a melodramatic robot rebellion, Lem diagnosed a subtler epistemic trap: human civilization becoming structurally dependent on black-box cybernetic regulators whose solutions surpass the capacity of human scientists to verify or comprehend. Born in interwar Lwów (then Poland, now Lviv, Ukraine) in 1921, where he began medical studies in 1940 and witnessed Soviet and Nazi occupations before relocating to Kraków in 1945, Lem brought a skeptical Central European sensibility toward grand teleological systems to cybernetics. Because Summa technologiae was not translated into English until Joanna Zylinska’s 2013 edition (University of Minnesota Press), the anglophone AGI tradition—from Gubrud and Goertzel to Legg and Bostrom—developed without direct contact with Lem’s work, making this a striking historical anticipation rather than a chain of textual transmission.

Why Now

Until 2020, “AGI” was confined to rationalist blogs, a niche academic conference, and existential-risk books; three shifts moved it to the center of global politics and finance. First, scaling autoregressive Transformers from GPT-3 (2020) to GPT-4 (2023) and multimodal reasoning models demonstrated cross-domain transfer that narrow AI deemed impossible: a single pretrained network could pass bar exams, write code, translate languages, and debug proofs, prompting Microsoft researchers in March 2023 to claim “sparks of AGI” (sparking a controversy over test-set contamination). Second, AGI became a binding legal and geopolitical term: the OpenAI–Microsoft partnership tied profit caps and IP licenses to whether OpenAI’s board declared AGI achieved (reportedly benchmarked in 2025 negotiations to a $100 billion profit threshold), while the March 2023 pause letter and the November 2023 Bletchley Park AI Safety Summit placed general-purpose frontier systems on the agenda of heads of state. Third, measurement moved to structured taxonomies: Google DeepMind’s Levels of AGI (Meredith Ringel Morris et al., November 2023) and François Chollet’s ARC-AGI benchmark forced researchers to distinguish rote template memorization from on-the-fly skill acquisition, even as lab CEOs began quoting AGI timelines in years rather than decades.

Explanatory Power

Unpacking the concept separates three dimensions that public commentary tangles together. First is breadth versus depth: Stockfish or AlphaFold is superhumanly deep within a single formal ruleset while having zero competence outside it, showing why superhuman performance on one benchmark says nothing about general intelligence. Second is cognitive competence versus operational autonomy: a model can score in the 99th percentile on medical licensing or olympiad math problems in a single turn while lacking the ability to pursue a week-long engineering project across changing environments without human scaffolding; safety and labor policy must track autonomy as closely as test scores. Third is cognitive architecture versus economic substitutability: defining AGI as outperforming humans at most economically valuable work is a statement about labor markets and software interfaces, not minds, and could be satisfied by a patchwork of narrow tools without solving general reasoning. Through this lens, debates over whether LLMs “truly understand” become concrete empirical questions, contract clauses referencing AGI stand revealed as revenue triggers, and forecasts of “AGI by 2027” usually mean automating ML engineering workflows rather than universal intelligence.

Axes of Debate

Underappreciated Angle

The least appreciated dynamic surrounding AGI is a self-consuming measurement paradox: every benchmark that gains authority as a test of “general intelligence” immediately becomes an optimization target for training pipelines, destroying its validity as a measure of generality. Tasks universally cited forty years ago as hallmarks of general human intellect—grandmaster chess, idiomatic translation, code generation, bar exams—are routinely reclassified as “mere narrow pattern matching” the week after a model masters them (the AI effect). Consequently, the boundary of AGI recedes because humanity unconsciously defines “general intelligence” negatively—as whatever machines still cannot do today. Stanisław Lem saw the end of this loop in 1964: the decisive crisis arrives not when a machine passes a philosophical test of thinking “like a human,” but when human institutions begin acting on machine-generated strategies whose validity no human committee can check.