Introduction
Artificial General Intelligence (Artificial general intelligence) constitutes one of the most significant prospective transformations in the history of science and technology, representing a shift from narrow computational systems towards broadly capable, adaptive potentially autonomous forms of intelligence. This white paper provides an extended and in-depth exploration of the future trajectories of general intelligence, advancing the argument that Artificial general intelligence will not emerge through linear scaling alone but through the convergence of multiple paradigms, including architectural innovation, agentic autonomy, embodied cognition distributed infrastructural models. It critically examines the limitations of current approaches, particularly the scaling paradigm associated with large models proposes a synthesis of emerging frameworks that prioritise continual learning, world modelling socio-technical integration. Furthermore, it evaluates the epistemological and governance challenges posed by increasingly capable systems, arguing that the evolution of general intelligence is inseparable from institutional, philosophical geopolitical developments. The analysis is intended to provide an authoritative foundation for advanced postgraduate study and research, situating Artificial general intelligence within a broader context of technological evolution and societal transformation.
The Conceptual Horizon of General Intelligence
The pursuit of artificial general intelligence has long occupied a central position within the broader field of artificial intelligence, yet its meaning and feasibility remain contested. While contemporary systems have demonstrated remarkable capabilities across language processing, image generation complex problem-solving, they remain fundamentally constrained by their dependence on pre-defined training regimes, statistical pattern recognition limited forms of generalisation. General intelligence, in contrast, implies the capacity to adapt flexibly across domains, to reason abstractly, to learn continuously from experience to operate autonomously within dynamic and uncertain environments. These attributes collectively distinguish human cognition from existing artificial systems and define the conceptual horizon towards which current research is oriented. The central thesis of this paper is that the future trajectory of general intelligence will be shaped by the interaction of multiple developmental pathways rather than a single dominant paradigm that understanding these pathways requires an integrated analysis of technical, cognitive socio-political dimensions.
The Limits of the Scaling Paradigm
The recent success of artificial intelligence has been closely associated with the scaling paradigm, whereby increases in model size, training data computational resources have yielded substantial improvements in performance. This paradigm has been particularly evident in large language models and multimodal systems, which exhibit emergent capabilities as their scale increases. However, the assumption that general intelligence can be achieved through indefinite scaling is increasingly subject to critical scrutiny. Empirical evidence suggests that while scaling enhances pattern recognition and interpolation within known domains, it does not inherently confer the ability to perform robust causal reasoning, to develop stable long-term memory, or to engage in genuine abstraction. These limitations indicate that scaling, although necessary, is not sufficient for the emergence of general intelligence.
A more nuanced understanding of scaling is therefore required, one that encompasses not only quantitative expansion but also qualitative transformation. Emerging frameworks conceptualise scaling along multiple axes, including efficiency, distribution adaptability. The notion of “scaling out” is particularly significant, referring to the distribution of intelligence across networks of interacting systems rather than its concentration within a single monolithic model. This shift reflects a broader recognition that intelligence, whether biological or artificial, is inherently distributed and context-dependent. In parallel, the emphasis on efficiency and “scaling down” addresses the practical constraints of energy consumption, latency accessibility, which are increasingly salient as AI systems are deployed at scale. These developments suggest that the future of general intelligence will involve a reconfiguration of the scaling paradigm, integrating expansion with optimisation and decentralisation.
Architectural Innovation and Continual Learning
The limitations of current systems have prompted a renewed focus on architectural innovation, with the aim of developing models that more closely approximate the functional characteristics of human cognition. Contemporary deep learning architectures, while highly effective for pattern recognition, lack several key features associated with general intelligence, including persistent memory, structured reasoning the capacity for continual learning. Addressing these deficiencies requires the development of new architectures that integrate multiple forms of representation and processing.
One promising direction involves the incorporation of mechanisms inspired by neuroscience, such as dual memory systems that distinguish between short-term and long-term storage, as well as forms of synaptic plasticity that enable adaptive learning over time. These approaches aim to overcome the problem of catastrophic forgetting, whereby models lose previously acquired knowledge when trained on new data to facilitate lifelong learning in dynamic environments. At the same time, there is growing interest in hybrid architectures that combine symbolic and sub-symbolic methods, enabling systems to perform both statistical inference and logical reasoning. Such architectures reflect a broader shift from purely data-driven approaches towards models that incorporate structured knowledge and causal representations.
Another significant development is the emergence of developmental or ontogenetic frameworks, which conceptualise intelligence as a process of growth and adaptation rather than a static property. Within this perspective, an intelligent system is not fully specified at the outset but evolves through interaction with its environment, acquiring increasingly complex capabilities over time. This approach aligns with theories of human cognitive development and suggests that the path to general intelligence may involve staged learning processes, guided exploration the gradual accumulation of knowledge. The emphasis on development also highlights the importance of embodiment and interaction, as discussed in subsequent sections underscores the need to move beyond static training paradigms towards dynamic, experiential learning systems.
Agentic Autonomy, World Models and Embodiment
A central trajectory in the evolution of general intelligence is the transition from passive systems to agentic entities capable of autonomous action. Agentic artificial intelligence is characterised by the integration of perception, reasoning action within a continuous feedback loop, enabling systems to pursue goals, adapt to changing conditions interact with complex environments. This shift represents a fundamental redefinition of artificial intelligence, transforming it from a tool for processing information into an active participant in the world.
The development of world models constitutes a critical component of this trajectory. World models are internal representations that allow a system to simulate aspects of its environment, predict the consequences of actions plan accordingly. By enabling learning through simulation, these models facilitate generalisation across contexts and reduce the need for extensive real-world data. They also provide a foundation for more sophisticated forms of reasoning, including counterfactual analysis and long-term planning. The integration of world models with agentic frameworks thus represents a key step towards the realisation of general intelligence.
Embodiment further extends this paradigm by emphasising the role of physical interaction in the development of cognition. Embodied systems are grounded in sensorimotor experience, enabling them to acquire knowledge through direct engagement with their environment. This approach challenges the assumption that intelligence can be fully captured through abstract computation, suggesting instead that cognition is inherently linked to perception and action. Advances in robotics and vision-language-action models illustrate the potential of this approach, demonstrating how integrated systems can perform complex tasks in real-world settings. The convergence of agentic autonomy, world modelling embodiment therefore represents a major trajectory in the evolution of general intelligence, with significant implications for both research and application.
Distributed and Edge-Based Intelligence
The infrastructural context of artificial intelligence is undergoing a profound transformation, driven by the proliferation of connected devices and the increasing importance of real-time processing. Traditional models of AI deployment rely on centralised cloud infrastructure, which offers substantial computational resources but introduces challenges related to latency, privacy scalability. In response, there is a growing shift towards decentralised and edge-based systems, in which intelligence is distributed across a network of devices operating in diverse environments.
This transition reflects a broader conceptual shift towards distributed cognition, in which intelligence is understood as an emergent property of interacting components rather than a feature of a single system. Edge intelligence enables local processing and decision-making, reducing dependence on centralised resources and enhancing responsiveness. It also facilitates personalised and context-aware applications, as systems can adapt to the specific conditions and preferences of individual users. The integration of agentic capabilities within edge environments further extends this paradigm, enabling networks of autonomous systems to collaborate and coordinate in complex settings such as smart cities, transportation networks industrial systems.
The decentralisation of intelligence also has significant implications for governance and control, as it challenges traditional models of oversight and regulation. Ensuring the reliability and safety of distributed systems requires new approaches to coordination, verification accountability, highlighting the interplay between technical and institutional factors in the development of general intelligence.
Temporal Scaling and Complex Task Automation
An important dimension of recent progress in artificial intelligence is the increasing capacity of systems to handle tasks of greater temporal and structural complexity. This phenomenon, sometimes described as temporal scaling, reflects the ability of AI systems to maintain coherence and effectiveness over extended sequences of actions and interactions. Whereas earlier systems were limited to short, well-defined tasks, contemporary models are increasingly capable of managing multi-step workflows, integrating information over longer time horizons adapting to evolving objectives.
The implications of this trend are far-reaching, particularly in relation to the automation of cognitive labour. As AI systems become capable of performing complex tasks that require sustained attention and coordination, they have the potential to transform a wide range of professional domains, including research, engineering administration. However, the extension of temporal capabilities also introduces new challenges, particularly in relation to reliability and control. Ensuring that systems behave consistently and predictably over long time horizons is a non-trivial problem, requiring advances in monitoring, verification alignment.
Safety, Governance and Epistemological Challenges
The development of general intelligence raises profound questions regarding safety, governance the nature of knowledge itself. As AI systems become more capable and autonomous, the potential risks associated with their deployment increase correspondingly. These risks include not only technical failures and unintended behaviours but also broader societal impacts, such as the concentration of power, the erosion of privacy the transformation of labour markets. Addressing these challenges requires a comprehensive approach to governance that integrates technical, ethical institutional perspectives.
One of the central issues in AI safety is the problem of alignment, which concerns the extent to which the goals and behaviours of artificial systems correspond to human values and intentions. Achieving alignment is complicated by the difficulty of specifying complex and context-dependent values in a formal and operational manner. Moreover, as systems become more autonomous, the potential for emergent behaviours that were not explicitly programmed or anticipated increases, further complicating efforts to ensure safety.
The governance of general intelligence is also shaped by geopolitical dynamics, as nations and organisations compete for leadership in a strategically significant domain. This competition can both accelerate innovation and exacerbate risks, particularly if it leads to the deployment of systems without adequate safeguards. International cooperation and the development of shared standards are therefore essential components of a sustainable approach to AI development.
At a deeper level, the emergence of general intelligence challenges traditional epistemological frameworks, particularly in relation to the production and validation of knowledge. As AI systems become capable of generating novel insights and conducting autonomous research, questions arise regarding the nature of understanding, the role of human expertise the criteria for trust and reliability. These issues suggest that the impact of general intelligence will extend beyond technology, reshaping fundamental aspects of human cognition and society.
Conclusion
The future of general intelligence is characterised by complexity, uncertainty convergence. While significant progress has been made in recent years, the realisation of Artificial general intelligence remains contingent upon the integration of multiple technological and conceptual advances. Scaling alone is insufficient; it must be complemented by architectural innovation, agentic autonomy, embodied interaction distributed deployment. At the same time, the development of general intelligence is inseparable from broader socio-technical dynamics, including governance, ethics epistemology.
The trajectory outlined in this paper suggests that general intelligence will not emerge as a single, discrete breakthrough but as an evolving ecosystem of systems and practices. Its development will require sustained collaboration across disciplines, as well as careful consideration of the societal implications of increasingly capable technologies. In this sense, the future of general intelligence is not merely a technical problem but a collective endeavour that encompasses science, philosophy public policy.
Bibliography
- AAAI, Report on the Future of AI Research, 2025.
- Bengio, Y. et al., International AI Safety Report, 2025-2026.
- Gupta, R. et al., ‘Personalized Artificial General Intelligence via Neuroscience-Inspired Continuous Learning Systems’, arXiv (2025).
- Joshi, S., Comprehensive Review of Artificial General Intelligence and Agentic GenAI Applications in Business and Finance, 2025.
- Live Science, ‘AI Task Complexity Growth Study’, 2025.
- Live Science, ‘Efficient AI Models and Edge Deployment’, 2026.
- Subasioglu, M. and Subasioglu, N., ‘From Mimicry to True Intelligence’, arXiv (2025).
- The Guardian, ‘DeepMind and the Path to Artificial general intelligence’, 2025.
- Wang, Y., Li, Y. Xu, C., ‘AI Scaling: From Up to Down and Out’, Proceedings of the International Conference on Machine Learning, 2025.
- Zhang, R. et al., ‘Toward Edge General Intelligence with Agentic AI’, arXiv (2025).
- ‘Ontogenetic Architecture of General Intelligence’, conceptual framework, 2025.
- ‘Vision-Language-Action Models’, technical overview, 2025.