Machine General Intelligence did not emerge as a distinct scientific objective overnight. Rather, it represents the culmination of more than seventy years of research into computation, cognition, mathematics, psychology and engineering. Throughout this period, the field has undergone repeated cycles of optimism, disappointment, technological breakthrough and conceptual refinement. Each generation of researchers has contributed new insights into the nature of intelligence, while simultaneously revealing the limitations of prevailing approaches.
Understanding this historical progression is essential because it demonstrates that Machine General Intelligence is not merely a larger version of today's artificial intelligence systems. Instead, it is the product of an evolving understanding of intelligence itself. The journey from symbolic reasoning to statistical learning, from expert systems to foundation models and from isolated algorithms to integrated cognitive architectures reflects a gradual shift from programming intelligent behaviour to constructing systems capable of learning intelligence autonomously.
This chapter traces that intellectual journey, highlighting the milestones that continue to shape contemporary research.
The Origins of Computational Intelligence
The origins of Machine General Intelligence can be traced to the emergence of theoretical computer science during the first half of the twentieth century. Before practical computers existed, mathematicians had already begun asking whether reasoning itself could be mechanised.
Among the most influential figures was Alan Turing, whose work fundamentally transformed the relationship between mathematics and computation. In 1936, Turing introduced the concept of the Universal Turing Machine, demonstrating that a single abstract computational device could simulate any algorithmic process given sufficient memory and time. This idea established the theoretical foundation for programmable computers and, indirectly, for artificial intelligence.
Following the Second World War, Turing turned his attention from computation itself to the nature of intelligence. His landmark 1950 paper Computing Machinery and Intelligence posed the provocative question:
Can machines think?
Rather than attempting to define thought philosophically, Turing proposed an operational criterion, now known as the Turing Test, in which a machine would be considered intelligent if a human evaluator could not reliably distinguish its conversational behaviour from that of another human.
Although later criticised for measuring behaviour rather than cognition, the Turing Test fundamentally changed the discourse surrounding artificial intelligence. Intelligence became something that could, at least in principle, be investigated scientifically rather than remaining purely philosophical.
More importantly, Turing recognised that intelligence involved learning rather than fixed programming. He proposed constructing "child machines" capable of education and adaptation, an idea remarkably consistent with contemporary research into continual learning and Machine General Intelligence.
The Dartmouth Conference and the Birth of Artificial Intelligence
The formal birth of artificial intelligence as a scientific discipline is generally dated to the Dartmouth Summer Research Project on Artificial Intelligence in 1956.
Organised principally by John McCarthy, together with Marvin Minsky, Claude Shannon and Nathan Rochester, the conference introduced the term Artificial Intelligence and articulated an ambitious research vision.
The proposal suggested that every aspect of intelligence might ultimately be described precisely enough for machines to simulate.
This optimism reflected the intellectual climate of the era.
Digital computers had recently demonstrated extraordinary computational power.
Logic appeared capable of representing reasoning.
Information theory had transformed communication science.
Many researchers believed genuinely intelligent machines might emerge within a generation.
Early successes reinforced this confidence.
Programs successfully proved mathematical theorems.
Others solved algebra problems, played games and manipulated symbolic knowledge.
Although primitive by modern standards, these demonstrations suggested that symbolic reasoning could reproduce increasingly sophisticated aspects of human cognition.
The long-term ambition already resembled what would later become Machine General Intelligence.
Researchers sought systems capable not merely of performing calculations but of reasoning generally.
Symbolic Artificial Intelligence
During the 1960s and 1970s, symbolic approaches became the dominant paradigm.
Often described as Good Old-Fashioned Artificial Intelligence (GOFAI), symbolic AI represented knowledge explicitly through logical rules, semantic networks and symbolic representations.
The underlying assumption was straightforward.
If intelligence consists of manipulating abstract symbols according to logical principles, then machines capable of sufficiently sophisticated symbolic reasoning should eventually exhibit intelligent behaviour.
This approach proved remarkably successful for constrained reasoning tasks.
Systems could:
- prove mathematical theorems;
- solve logical puzzles;
- manipulate formal grammars;
- represent expert knowledge;
- perform deterministic planning.
Symbolic methods also possessed an important advantage that remains attractive today:
their reasoning processes were largely transparent.
Unlike modern neural networks, symbolic systems could explain precisely why conclusions had been reached.
However, serious limitations gradually emerged.
Real-world environments proved far more uncertain than symbolic representations assumed.
Human knowledge frequently contains ambiguity, incomplete information and exceptions impossible to encode exhaustively.
Simple tasks such as recognising everyday objects or understanding natural conversation required vastly more background knowledge than anticipated.
The difficulty of representing common sense became one of symbolic AI's greatest obstacles.
Expert Systems and the First Commercial Revolution
During the late 1970s and early 1980s, symbolic AI evolved into expert systems, representing one of the first commercially successful applications of artificial intelligence.
Rather than pursuing general intelligence directly, expert systems attempted to reproduce specialised expertise within narrowly defined professional domains.
Knowledge engineers interviewed human specialists before encoding their expertise as extensive collections of logical rules.
Prominent examples included:
- MYCIN for infectious disease diagnosis;
- XCON for computer system configuration;
- DENDRAL for chemical analysis.
These systems demonstrated that carefully encoded knowledge could outperform inexperienced practitioners within specialised tasks.
Industries invested heavily.
Governments funded ambitious AI programmes.
Optimism returned.
Yet expert systems ultimately revealed a critical distinction between expertise and intelligence.
Although capable within carefully bounded environments, they proved brittle when confronted with novel situations.
Updating thousands of manually written rules became prohibitively expensive.
Knowledge acquisition itself became known as the knowledge engineering bottleneck.
Most importantly, expert systems did not learn.
Every improvement required human intervention.
General intelligence, however, requires continuous adaptation.
AI Winters and Intellectual Reassessment
Repeated cycles of inflated expectations followed by disappointing results produced periods now commonly known as the AI Winters.
Funding declined dramatically during the mid-1970s and again in the late 1980s.
Public confidence weakened.
Many ambitious research programmes were abandoned.
Although frequently portrayed negatively, these periods played an essential role in the evolution of artificial intelligence.
Researchers gradually recognised that intelligence was substantially more complex than initially assumed.
Reasoning alone proved insufficient.
Perception, learning, memory, language, motor control, uncertainty and social interaction all contributed fundamentally to intelligent behaviour.
Rather than representing a single algorithm, intelligence appeared increasingly to resemble an integrated architecture composed of numerous interacting cognitive processes.
This conceptual shift continues to influence Machine General Intelligence research today.
The Rise of Machine Learning
The 1990s witnessed a fundamental transformation in artificial intelligence.
Instead of programming knowledge directly, researchers increasingly focused upon algorithms capable of learning from data.
This transition represented one of the most important paradigm shifts in computational science.
Machine learning reversed the traditional relationship between programmers and software.
Rather than specifying every rule explicitly, developers designed algorithms capable of discovering patterns autonomously.
As datasets expanded and computing power increased, statistical learning consistently outperformed manually engineered rule systems across numerous practical applications.
Algorithms such as:
- decision trees;
- support vector machines;
- Bayesian networks;
- ensemble methods;
became widely adopted throughout academia and industry.
Learning, rather than logical deduction, became the central organising principle of artificial intelligence.
For the first time, machines improved automatically through experience.
Although still largely specialised, these systems demonstrated the enormous potential of adaptive computation.
Deep Learning and Representation Learning
A second major revolution began during the early 2010s with the emergence of deep learning.
Artificial neural networks had existed for decades.
However, improvements in computing hardware, algorithmic optimisation and access to enormous datasets finally enabled very large neural architectures to outperform traditional methods across multiple domains.
Unlike earlier machine learning approaches, deep neural networks automatically learned hierarchical internal representations directly from raw data.
Instead of manually engineering visual features for image recognition, networks discovered increasingly abstract concepts themselves.
The consequences proved extraordinary.
Deep learning rapidly transformed:
- computer vision;
- speech recognition;
- natural language processing;
- robotics;
- autonomous vehicles;
- biological modelling.
Image classification reached superhuman accuracy within many benchmarks.
Speech recognition approached human performance.
Language translation improved dramatically.
Scientific applications expanded rapidly.
For many researchers, deep learning suggested that increasingly general intelligence might emerge through scaling sufficiently capable learning systems.
Foundation Models and Emergent Capabilities
The introduction of Transformer architectures in 2017 initiated perhaps the most significant development in modern artificial intelligence.
Unlike previous neural networks optimised for individual tasks, Transformer-based foundation models learned broad statistical representations from extraordinarily diverse datasets.
Scaling these architectures produced unexpected emergent capabilities.
Models demonstrated:
- reasoning;
- translation;
- summarisation;
- programming;
- mathematical problem solving;
- multimodal understanding;
- scientific assistance.
Remarkably, many of these abilities were never explicitly programmed.
Instead, they appeared naturally as models increased in scale.
This phenomenon fundamentally reshaped thinking about Machine General Intelligence.
Rather than constructing dozens of specialised systems, researchers began exploring whether sufficiently capable general-purpose architectures might acquire increasingly broad cognitive abilities through self-supervised learning.
Yet important limitations remained.
Despite impressive linguistic competence, current foundation models continue to struggle with:
- robust causal reasoning;
- persistent memory;
- continual learning;
- grounded understanding;
- autonomous long-term planning;
- genuine world modelling.
Consequently, foundation models are widely viewed as important milestones rather than complete realisations of Machine General Intelligence.
Towards Machine General Intelligence
Modern research increasingly recognises that no single paradigm is likely to produce genuinely general intelligence.
Instead, Machine General Intelligence appears to require the integration of multiple complementary approaches.
Future systems will likely combine:
- large-scale neural learning;
- symbolic reasoning;
- causal inference;
- reinforcement learning;
- world models;
- episodic and semantic memory;
- multimodal perception;
- language understanding;
- social cognition;
- continual self-improvement.
This emerging synthesis reflects an important historical lesson.
Each previous generation of AI captured genuine aspects of intelligence while overlooking others.
Symbolic systems explained reasoning.
Machine learning explained adaptation.
Deep learning explained representation learning.
Foundation models demonstrated emergent generality.
Machine General Intelligence seeks to integrate all of these capabilities within coherent cognitive architectures rather than treating them as competing alternatives.
Lessons from Seventy Years of Artificial Intelligence
The history of artificial intelligence reveals several enduring lessons.
First, intelligence is vastly more complex than early researchers anticipated.
Second, learning consistently proves more scalable than manual programming.
Third, intelligence emerges from interaction among multiple cognitive processes rather than isolated algorithms.
Fourth, increasing computational power alone is insufficient without corresponding architectural innovation.
Finally, progress has been driven as much by conceptual advances as by technological improvements.
Machine General Intelligence therefore represents neither a sudden revolution nor a simple extension of current systems.
It is the product of decades of accumulated scientific understanding concerning learning, reasoning, perception, abstraction and adaptation.
The pursuit of Machine General Intelligence should consequently be viewed not as replacing previous approaches but as integrating their strongest insights into unified computational systems capable of exhibiting the flexibility that characterises natural intelligence.
The historical trajectory also offers a broader lesson. Every generation of researchers has expanded humanity's understanding of intelligence by attempting to recreate it. In doing so, artificial intelligence has become not only an engineering discipline but also one of the most important scientific investigations into the nature of cognition itself.