Humans dominate the Earth not due to physical strength or speed, but (largely) because of a unique and transformative trait: general intelligence. The evolution of human intelligence was a long and complex journey. For most species, the development of general intelligence is constrained by its high energy demands —and inefficacy. As a result, animals typically evolve narrow intelligence, which is optimized for survival within specific ecological spaces.
General intelligence is the ability to learn, reason, and apply knowledge across diverse domains, even those without immediate survival benefits. While most animals excel in narrow intelligence (e.g., a spider spinning a web or a bird building a nest), they lack the cognitive flexibility to apply these skills outside their niche. Humans, on the other hand, can generalize knowledge and use it to tackle entirely new challenges.
The development of human general intelligence has led to transformative milestones, starting with the discovery of fire, the first tools all the way to the digital age. Today, humanity is on the verge of creating "that which makes it the most powerful species in the world"—intelligence. Could this be humanity's final invention?
Artificial Intelligence (AI) is everywhere—both in obvious forms and more subtle applications. From the algorithms behind personalized media recommendations, like those suggesting music, movies, or news article, to self-driving cars navigating (some of) our highways using computer vision, advanced sensors, and actuators, AI has already become deeply integrated into our lives. And, of course, there are large language models like ChatGPT, Claude, and DeepSeek, which are reshaping how we interact with information and technology.
But this raises the question: What is artificial intelligence? Could a series of simple if-then statements qualify as AI? If so, how many statements would it take? Defining what AI truly is—and what it isn’t—is complex for several reasons.
One challenge is that AI as a field of research lacks a fixed definition. Like fields like biology or chemistry, which study specific phenomena, AI focuses on developing methods to replicate or approximate intelligent behavior. These methods vary widely, from simple decision trees to complex neural networks.
Another complication comes from the legacy of science fiction. Popular culture often depicts AI as a form of general intelligence—a machine capable of independent thought, emotion, and creativity. Think of the robotic monotones delivering overly detailed facts in movies or the humanoid robots that wrestle with existential dilemmas (hmm.. hmm..). In reality, AI today is far from this fictional ideal. The vast majority of AI systems are narrow AI—highly specialized tools designed for specific tasks, such as recognizing faces, translating languages, or washing dishes (surprisingly without hands).
There’s also a disconnect in how people perceive the difficulty of creating AI systems. What seems hard to humans is often easy for AI, and what seems easy is surprisingly hard.
Defining artificial intelligence (AI) has always been a challenge, partly because the definition itself tends to shift as the field progresses. Here are three definitions that highlight different perspectives from Elements of AI:
At its core, machine learning involves creating algorithms that learn from data, allowing systems to improve performance over time without being explicitly programmed. ML is often categorized into three types:
Classification is a core aspect of machine learning, allowing algorithms to categorize inputs into predefined classes based on patterns learned from labeled data. Essentially, it’s about teaching a system to assign labels to new inputs based on the characteristics of the training data.
A common example of classification is the nearest neighbor classifier. This algorithm classifies a new input by comparing it to the training data and selecting the class of the closest data point. In other words, if you want to predict the category of a new item, the classifier looks for the nearest "neighbor" in the training set and assigns the same category. This method is widely used in various applications, including media recommendation systems. For example, platforms like Netflix or Spotify use classification algorithms to suggest movies, music, or shows. These systems typically classify users into groups based on preferences and behaviors. Once a user is classified, the system suggests media that other similar users have enjoyed but that the current user has yet to discover.
Neural networks are the foundation of deep learning, a subset of machine learning (ML) that has revolutionized fields like image recognition, natural language processing, and autonomous driving. Unlike traditional computing systems, neural networks are inspired by the structure and function of the human brain, enabling them to process information in a fundamentally different way.
In traditional computers, information is processed in a central processing unit (CPU), which can only handle one task at a time. The CPU retrieves data from memory, processes it, and stores the result back in memory. This separation of data storage and processing creates a bottleneck, especially for tasks requiring large-scale computations. Neural networks, on the other hand, consist of a large number of neurons (or nodes) that can process information simultaneously. Each neuron operates independently, allowing the network to handle vast amounts of data in parallel. This parallel processing capability is one of the key reasons neural networks excel at tasks like image recognition and language translation
In traditional computers, data storage (memory) and processing (CPU) are separate components. Data must be retrieved from memory, processed by the CPU, and then stored back in memory. This back-and-forth can slow down computation, especially for large datasets. In neural networks, storage and processing are integrated. Neurons both store and process information. Short-term data is stored in the activation state of neurons (whether they "fire" or not), while long-term data is stored in the weights of the connections between neurons. These weights determine the strength of the connections and are adjusted during training to improve the network’s performance.
A simple neural network consists of three main components:
Each neuron in a layer is connected to neurons in the next layer via weighted connections. During training, the network adjusts these weights using optimization techniques like gradient descent to minimize errors in its predictions.
Regression techniques play a crucial role in training neural networks. They are used to extract the weights that define the connections between neurons.backpropagation is the primary method for adjusting weights. It works by: Calculating the error between the network’s prediction and the actual result. Propagating this error backward through the network to update the weights. Repeating this process iteratively to improve accuracy.
The history of artificial intelligence (AI) has been marked by various paradigms—dominant trends or approaches that gain widespread acceptance among researchers and drive optimism about the field’s potential. These paradigms shape how AI is developed and influence the expectations for its future.
For example, in the 1960s, neural networks emerged as a paradigm. Researchers believed that mimicking the human brain's learning processes could solve all AI problems. This optimism faded as limitations in computing power and understanding of neural networks became apparent, leading to what is now referred to as an AI Winter—a period of reduced funding and interest in AI research.
By the 1980s, a new paradigm emerged with expert systems, which relied on logic and human-coded rules to replicate decision-making processes. These systems showed promise in specialized fields but ultimately proved inflexible and unable to scale effectively for broader applications.The current AI Summer—a period of rapid progress and enthusiasm—was brought about by advancements in processing power, the availability of vast datasets, and breakthroughs in machine learning, particularly deep learning. These factors have enabled AI to make significant strides in areas like computer vision, natural language processing, and robotics.
The future of AI is the subject of much speculation, with questions arising about its economic, social, and political implications. Predictions range from utopian visions—such as universal basic income, personalized education, and improved healthcare—to dystopian fears, like AI overlords as popularized by science fiction (e.g., The Terminator).
While AI is expected to become increasingly pervasive in our lives, it is the possibility of Artificial General Intelligence (AGI) that raises the most concern. AGI refers to an AI system capable of performing any intellectual task that a human can, with the ability to reason, learn, and adapt across a wide range of domains.AGI remains a theoretical concept and is not yet achievable with current technology. Most AI systems today are narrow AI, designed for specific tasks.
Mustafa Suleyman, co-founder of DeepMind, has described AGI as not merely a new technology but a culmination of humanity itself. In his view, AGI will embody the full spectrum of human traits—the good and the bad. Suleyman argues that the challenge of this century is not merely developing AGI but ensuring that it reflects humanity’s empathy, love, and wisdom.