Abstract
The idea that intelligence is the result of a computational process and can, therefore, be automated, is centuries old. We review the historical origins of the idea that machines can be intelligent, and the most significant contributions made by Thomas Hobbes, Charles Babbage, Ada Lovelace, Alan Turing, Norbert Wiener, and others. Objections to the idea that machines can become intelligent have been raised and addressed many times, and we provide a brief survey of the arguments and counter-arguments presented over time. Intelligence was first viewed as symbol manipulation, leading to approaches that had some successes in specific problems, but did not generalize well to real-world problems. To address the difficulties faced by the early systems, which were brittle and unable to handle unforeseen complexities, machine learning techniques were increasingly adopted. Recently, a sub-field of machine learning known as deep learning has led to the design of systems that can successfully learn to address difficult problems in natural language processing, vision, and (yet to a lesser extent) interaction with the real world. These systems have found applications in countless domains and are one of the central technologies behind the fourth industrial revolution, also known as Industry 4.0. Applications in analytics enable artificial intelligence systems to exploit and extract economic value from data and are the main source of income for many of today’s largest companies. Artificial intelligence can also be used in automation, enabling robots and computers to replace humans in many tasks. We conclude by providing some pointers to possible future developments, including the possibility of the development of artificial general intelligence, and provide leads to the potential implications of this technology in the future of humanity.
Introduction: The Philosophical Roots of AI
The quest to automate intelligence is not a modern phenomenon but a centuries-old pursuit. In their comprehensive review, Arlindo L. Oliveira and Mário A. T. Figueiredo trace the lineage of "thinking machines" from 17th-century philosophy to the current era of Large Language Models (LLMs). By examining the contributions of pioneers like Thomas Hobbes, Charles Babbage, and Ada Lovelace, the authors provide a grounded perspective on the transition from rigid logic to fluid, data-driven learning.
From Symbols to Learning: Overcoming Moravec’s Paradox
Early AI research was dominated by "Good Old Fashioned AI" (GOFAI), which focused on symbol manipulation and discrete rules. While successful in specific domains like the Logic Theorist or Expert Systems, these models were notoriously "brittle"—unable to handle the messy complexity of the real world.
This era highlighted Moravec’s Paradox: the discovery that high-level reasoning (like playing chess or solving calculus) is computationally easier for machines than low-level sensorimotor skills (like walking or recognizing a face). This paradox necessitated a shift toward statistical and numerical methods where systems adapt behavior based on input-output pairs rather than pre-programmed rules.
The Deep Learning Revolution
The transition from "programming" intelligence to "learning" it from data was fueled by the rise of neural networks and backpropagation. The authors highlight several key architectural breakthroughs and milestones:
- Convolutional Neural Networks (CNNs): Designed with an "inductive bias" for spatial invariance, CNNs revolutionized computer vision by assuming an object’s identity is independent of its location.
- Transformers: These architectures have become the backbone of modern Natural Language Processing (NLP), enabling models to process trillions of words and achieve near-human levels of text-based interaction.
- Large-Scale Benchmarks: The success of these models was accelerated by landmark datasets like ImageNet, which provided the massive volume of data necessary to train robust classifiers.
Key Results and Economic Impact
The qualitative leaps in AI are most visible in two areas:
- Superhuman Performance in Games: Moving beyond the brute force of Deep Blue, systems like AlphaZero utilize reinforcement learning to develop "alien" strategies that human experts now study.
- The NLP Breakthrough: The evolution from simple rephrasing tools like ELIZA to sophisticated models like ChatGPT represents a near-fulfillment of Alan Turing’s 1950 vision for the "Imitation Game."
Beyond research, these technologies are the engine of Industry 4.0. Analytics-driven AI is now the primary revenue source for global tech giants and is increasingly replacing humans in complex automation tasks.
Limitations and the Future of AGI
Despite these advances, significant hurdles remain. Current neural networks are "narrow"—they lack the autonomy to decide which problems are worth tackling. Furthermore, the relationship between digital intelligence and biological consciousness remains a mystery. Turing famously sidestepped the "consciousness" objection by focusing on behavior, but the debate over whether a digital substrate can replicate human experience persists.
As we look toward Artificial General Intelligence (AGI), the authors suggest that the path forward is as much a policy challenge as a technical one. Frameworks like the EU Artificial Intelligence Act will be essential to manage the risks of high-stakes automation and data access in the coming decades.
Summary updated