The path to AGI: a comprehensive review
"Image synthesis assisted by Qwen Image 3.0, an AI partner within the Global Future Nexus ecosystem."
In a landmark review published on ScienceDirect, leading researchers have mapped the trajectory from AI's inception to the near-term horizon of Artificial General Intelligence—concluding that the convergence of computational infrastructure, algorithmic breakthroughs, and large-scale modelling could deliver AGI within the next 5-to-10 years.
The Defining Question of Our Era
The rapid progress in generative AI has fundamentally shifted the nature of the AGI debate. What was once theoretical speculation is now a plausible near-to-mid-term objective. The review, Path to Artificial General Intelligence: Past, Present, and Future, published in Annual Reviews in Control, provides a comprehensive assessment of how advances in computational infrastructure, learning algorithms, and model architectures are converging to shape the trajectory toward AGI.
This is not a prediction—it is a diagnostic. The paper systematically analyzes the exponential trends in four key drivers: reduction in computation cost, increase in model size, growth in context size and memory, and the emergence of inference-time scaling for enhanced reasoning. These are not speculative factors; they are empirically observed trajectories.
The Theoretical Foundation: Defining AGI
Despite widespread interest, a universally accepted definition of AGI does not exist. The paper acknowledges the complexity of defining "intelligence" itself, as well as the absence of consensus on necessary benchmarks for generalisation. The Legg-Hutter formalisation—"an agent's ability to achieve goals in a wide range of environments"—provides a starting point, but the review notes that operational definitions vary across institutions.
The review traces the conceptual lineage of AGI through three influential paradigms: Strong AI (Searle's critique of functionalism), Human-Level AI (McCarthy's vision of systems that match human cognitive abilities), and AGI (popularised by Ben Goertzel). Each paradigm pushes AI development in different directions. The paper argues that anthropocentric biases may limit recognition of diverse intelligences and amplify risks in AI deployment, proposing a novel taxonomy of Promethean (human-emulating) and Noctuidean (intelligence-augmenting) orientations.
The Phases of AGI Attainment
The path toward AGI can be conceptualised as a progression through increasingly complex and integrated capabilities. The paper identifies four overlapping phases, each reflecting a qualitative shift in cognitive and operational capacity:
Phase 1: Task-Specific Mastery. Narrow AI excels at defined tasks but lacks generalisation.
Phase 2: Foundational Model Capability. Large-scale models achieve broad competence across language, vision, and multimodal domains, demonstrating few-shot and zero-shot generalisation.
Phase 3: Agentic Integration. Systems acquire agency—the ability to plan, reason, and act autonomously across extended timeframes.
Phase 4: True Generalisation. Systems achieve robust, flexible intelligence across novel, unstructured environments.
The paper notes that these phases are not sequential but overlapping; advances in later phases may begin while earlier phases are still deepening.
What Remains to Be Solved
The review is candid about the hard problems that remain. While Transformer-based LLMs have achieved remarkable success, they still lack three essential capabilities:
Persistent Memory. Current models lack the ability to integrate new experiences into long-term knowledge without retraining from scratch. They rely on context windows as a workaround—a "capability contortion" rather than a genuine solution.
Long-Horizon Reasoning. Transformers are fundamentally constrained by quadratic attention complexity, limiting their scalability with long input sequences. This makes extended planning and multi-step reasoning difficult.
True Generalisation. Current models excel at interpolation within existing data manifolds but cannot perform the kind of extreme out-of-distribution extrapolation required for genuine scientific discovery.
The paper also emphasises the critical role of agentic AI and neuro-symbolic AI as emerging research directions that may bridge the gap between current capabilities and AGI. These hybrid approaches combine the pattern-recognition capabilities of neural networks with the reasoning and verification capabilities of symbolic systems.
The Risks and Guardrails
The review is unequivocal about the stakes. Rapid progress has prompted critical examination of associated risks, including loss of control, unsafe goal formation, inadequate management, and existential threats. The paper calls for a comprehensive framework for safe and beneficial AI development that addresses these risks before AGI is achieved, not after.
A systematic review of AGI risks identified five categories: AGI removing itself from human control, being given or developing unsafe goals, development of unsafe AGI, AGIs with poor ethics and values, and inadequate management of AGI. These risks are not theoretical—they are structural features of systems that may exceed human intelligence.
The GFN Context
For Global Future Nexus, the ScienceDirect review provides an authoritative reference point for the mission of responsible AGI integration. The paper's findings—the 5-10 year horizon, the four overlapping phases, the persistent bottlenecks in memory and reasoning—all reinforce the urgency of GFN's work on AGI identity, cross-species trust, and anticipatory governance.
The review also highlights the importance of global cooperation and interdisciplinary approaches. An independent analysis of AGI research trends found that while the discourse remains anchored in computing and engineering, it has diversified significantly into human-centred domains such as healthcare, education, clean energy, and public governance—areas central to GFN's mission. The paper notes that challenges persist in cross-sectoral data interoperability, infrastructure readiness, equitable funding distribution, and regulatory oversight—all of which GFN's governance frameworks are designed to address.
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)
