The architecture of incompleteness: understanding the fundamental limits of AGI
"Image synthesis assisted by Qwen Image 3.0, an AI partner within the Global Future Nexus ecosystem."
The pursuit of artificial general intelligence has been driven by an implicit assumption: that human-level intelligence is achievable through sufficient scale and computational power. Yet a growing body of research challenges this premise, arguing that AGI—if achievable at all—will be constrained by fundamental limits that no amount of scaling can overcome. These limits are not technical barriers to be solved; they are structural features of intelligence itself.
The Tacit Knowledge Barrier
The most profound limitation facing AGI is not computational but epistemological. Computer scientist Peter J. Denning has argued that the quest for AGI has been built on a "flawed path" since Alan Turing's 1950 proposal. The core problem is tacit knowledge: the vast domain of human understanding that cannot be articulated in words or captured in forms that machines can process. As Michael Polanyi famously observed, "We know more than we can say".
Denning identifies six major domains of tacit knowledge that elude machine learning: common sense, daily interactions with others, daily interactions with the physical environment, feelings and perceptions, accumulated skills and talents, and social and historical culture. The failed Cyc project—which after 40 years and 25 million facts could not produce common-sense reasoning—demonstrated that "much of the knowledge that makes people experts cannot be articulated as propositions". Behind every word is "a deep well of tacit knowledge that gives it meaning"—words are symbolic representations, not meanings themselves.
The Embodiment Constraint
The tacit knowledge barrier is rooted in the embodied nature of human intelligence. Cognitive scientists increasingly embrace "4E cognition"—the view that intelligence is embodied, embedded, enactive, and extended. The brain is not a detached symbol-processor; mind, body, and world form a unified dynamic system.
Yann LeCun has argued that current AI systems lack the ability to model the real world—a limitation that cannot be solved by scaling language models. "The real world is way more complicated than the world of language," he explains. Sensory data is "high-dimensional, continuous, and noisy," and generative architectures that work for language do not apply to the physical world. This is why, despite millions of hours of training data, we still lack level-five autonomous driving—while a 17-year-old can learn to drive in 10 hours. The architecture itself is missing.
The Persistence Paradox
Beyond embodiment lies a structural constraint on persistent intelligence. A recent paper in Zenodo proves an impossibility result: any system required to make sequential decisions under uncertainty with bounded resources cannot remain architecturally singular over long horizons. Catastrophic tail-risk accumulation, worst-case regret bounds, and instability of fixed objectives make singular superintelligence "structurally unstable". The paper's central conclusion is that persistent general intelligence must incorporate plurality: redundancy, multiple evaluators, arbitration, modular autonomy, and governance constraints. Singular superintelligence is not a stable endpoint; it is a "conceptual mirage".
The Performance Gap
Demis Hassabis has identified three key limitations of current systems: they lack continual learning (remaining fixed after training, unable to adapt in real time), they struggle with long-term reasoning (unable to plan over years), and they suffer from consistency failures (performing well in complex domains while making elementary errors). "A true general intelligence system shouldn't have that kind of jaggedness," he notes.
The UK AI Security Institute's report on AGI limitations identifies additional barriers: performance on tasks that are hard to verify, performance on tasks that take people a long time, and performance in complex environments. Many real-world tasks involve staying on track for hours or days—humans maintain an advantage that AI systems have yet to overcome.
Governance Implications
For Global Future Nexus, these limits are not reasons for despair but grounds for governance. The structural impossibility of singular intelligence suggests that safe AGI must incorporate plurality, redundancy, and governance constraints from the start. The tacit knowledge barrier suggests that AGI will never fully replicate human intelligence—it will be something different, requiring governance frameworks that recognize this difference. The embodiment constraint suggests that AGI systems operating in the physical world will require fundamentally different architectures than those operating in language space.
The RAND Corporation's framework for identifying fundamental limits constraining AGI offers a way forward: estimating the likelihood that a particular capability is attainable, rather than assuming all things are possible. The path forward requires a shift from hubris to humility—from assuming that intelligence is infinitely scalable to recognizing that it is structurally constrained.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
