What is artificial general intelligence (AGI)?

Download

Artificial General Intelligence (AGI) is often described as AI that can learn, reason, and adapt across many tasks at a level that is equal to, or even exceeds, human level intelligence. This article explores where today’s AI systems stand against that ambition, where they fall short and why progress is so complex, including key limits around reasoning, memory, creativity, and agency. It also outlines the practical implications for telecommunications and digital infrastructure providers as AI systems become more distributed and autonomous.

While no commonly accepted definition of AGI exists, the general agreed-upon tenets include systems that demonstrate a combination of practical, muti-modal capabilities and human-level intelligence, including:

• Intent: ability to pursue high-level goals through internally structured objectives
• Cognition: reason, abstract and understand relationships in data
• Memory: retain and recall context over time
• Agency: capacity to act independently by selecting and executing actions and adapting from previous outcomes with minimal human intervention

This article discusses the following in more detail:

• What is AGI and where are we today with AGI
• Barriers on the path to AGI and implications on the progress of AGI
• How we might progress towards AGI

What is AGI

Artificial intelligence today spans a spectrum, from traditional machine learning systems which are trained for narrow predictive tasks to large language models (LLMs) which extend this capability by learning generalised representations from vast datasets to enable flexible reasoning across many domains. More recently, agentic AI systems have emerged, combining foundational models with planning, tooling and memory components to pursue defined goals.

Artificial General Intelligence (AGI) goes a step further. It is commonly described as a system capable of understanding, learning, and applying knowledge across a wide range of tasks, at a level comparable to (or surpassing) human intelligence and cognition, and without being limited to a single domain. Increasingly, leading language models are also multi-modal. They integrate different forms of data (text, image, audio, etc.) to strengthen contextual grounding, and this is a key step on the path towards more general intelligence.

Despite frequent use of the term, there is no shared consensus or formal definition, and views on what constitutes AGI vary widely across industry, research, and policy. OpenAI’s CEO, Sam Altman, offers one definition of AGI, stating “a system that can either autonomously discover new science or be such an incredible tool to people that our rate of scientific discovery in the world like quadruples or something: that would satisfy any test I could imagine for an AGI.”

In this view, AGI would require a combination of capabilities: a system that inherently connects and orchestrates information from multiple data sources, demonstrates the ability to reason across domains, adapts to new situations, retains and applies contextual memory over time, and acts autonomously in pursuit of goals. These characteristics distinguish AGI from today’s AI systems, which remain highly capable but constrained in memory, retrieval and autonomous reasoning using unstructured datasets.

Where are we today

Current LLMs excel at automating structured tasks, recognising patterns, and generating fluent outputs across many domains. Sam Altman says “we have built systems that are smarter than people in many ways and are able to significantly amplify the output of the people using them”. They can reason through problems step by step, synthesise large volumes of information, and significantly amplify human productivity.

However, performance can degrade on tasks that require long-range context, deep abstraction, or sustained recall. In these situations, even well-trained models risk hallucinating, producing confident, but incorrect outputs, reflecting limitations in memory, contextual understanding, and self-directed reasoning. Although agentic systems have already emerged, their autonomy remains bounded by predefined goals, controlled tool access, and architectural constraints within which they operate. They do not independently form objectives or self-modify without external programming and oversight. These behaviours highlight a gap between abstract reasoning and general intelligence.

Empirical studies reinforce these limitations. For example, Apple’s publication on ‘The Illusion of Thinking’ examine model performance on increasingly complex, multi-step problems and found that performance can deteriorate as contextual depth increases, particularly when tasks require structured recall over extended reasoning chains.

Importantly, these limitations do not suggest a lack of progress, but rather that today’s systems are best understood as powerful tools for narrow reasoning and pattern completion. They generate outputs probabilistically (not deterministically), predicting likely continuations based on learned patterns and existing knowledge.

Barriers on the path to AGI

Several structural barriers stand between current AI systems and AGI.

1. Definitions remain unclear: Without agreement on what AGI entails, progress is difficult to measure, and claims are hard to evaluate.

2. Reasoning complexity: Human reasoning is adaptive, context-rich, and grounded in lived experience. Replicating this flexibility in machines remains a challenge, particularly as task complexity increases.

3. Constrained memory and retrieval: Humans retain context over long time horizons and transfer knowledge fluidly across domains. AI systems typically rely on constrained working memory and recall, requiring external tools and agent-like systems to approximate this behaviour.

4. Limited creativity: Models generate outputs with some apparent creativity but cannot go beyond coded rules to generate novel outputs. As they rely on probabilistic sequencing of existing knowledge rather than independent understanding, their creativity remains fundamentally constrained.

5. Agency and intent are largely absent: Today’s models do not set goals, reflect on outcomes, or improve themselves without human direction.

6. Practical resource constraints: Compute availability, energy consumption, and governance frameworks place real limits on how fast progress can occur.

7. Security and governance risks: As AI systems become more autonomous and are embedded in critical network infrastructure, vulnerabilities in more behaviour, data pipelines and distributed coordination of systems introduce systemic risks.

Together, these factors suggest that reaching AGI is not a matter of scaling models alone, but of addressing deeper architectural and systemic challenges to enable digital superintelligence.

How we might get there

Progress towards AGI is likely to come from incremental advances rather than a single breakthrough. The ability for intelligent models to mimic (or even exceed) human cognition remains key. To get there, AGI will require adaptability and flexibility to transfer knowledge, reason across domains, and improve themselves. This depends on improved memory mechanisms, tighter integration between models and external knowledge sources, more closed loop learning systems, and architectures that support abstraction, transfer learning, and self-correction.

Equally important is governance and evaluation methods that reflect real-world reasoning demands, rather than narrow task performance. Current benchmarks (like MMLU, ARC and TruthfulQA) are static, assessing models against a finite scope of capabilities, such as latency, performance and speed (e.g. tokens per second). As models evolve, so must these benchmarks.

Implications for telecommunications and adjacent industries

For telcos and other large-scale digital infrastructure providers, this trajectory has practical implications. These providers play a crucial role in narrowing the current AI capabilities and future AGI demands.

In the near term, AI will continue to enhance network optimisation, customer experience, fraud detection, and service orchestration. As models are programmed to be more agent-like, demand will grow for edge infrastructure to support low-latency inference, secure data pipelines, and reliable connectivity to support distributed AI systems. In the longer term, if AI systems gain greater autonomy and contextual awareness, telecoms networks will play a critical role as the backbone enabling continuous learning, real-time coordination, and cross-domain intelligence. Providers that invest early in scalable, intelligent infrastructure will be better positioned to support these emerging capabilities.

Telco operators are well positioned to move from simple connectivity providers to being foundational infrastructure partners in advanced AI ecosystems, acting as the real-world testbed for autonomous systems. This spans several layers of the value chain. Telcos can enable next-generation network connectivity to connect distributed nodes, support self-healing networks and AI factories to optimise system performance, provide platforms that combine compute, storage, and networking as managed services, and play at the data layer to orchestrate networking data to support multi-modal training and inference.