Intelligence is the ability to abstract information and retain it as knowledge, for application within a certain context. Today, I'm researching artificial intelligence (or AI for short). Artificial Intelligence refers to the simulation of human intellectual processes by machines, often mimicking the human brain itself. AI processes are ordinarily carried out by computer systems. They include learning, reasoning, and self-correction. This can be useful when it comes to performing tasks that typically require human cognition, such as understanding natural language, recognizing patterns, solving problems, and making decisions.
How AI Works
Artificial intelligence uses algorithms and large amounts of data to help machines perform tasks that typically require human intelligence. At its core, AI relies on machine learning, where systems learn from examples rather than following strict rules. There are three main types of machine learning: supervised learning (trained on labeled data to make predictions), unsupervised learning (finds patterns in unlabeled data), and reinforcement learning (learns through trial and error with rewards). Modern AI often uses deep learning, which employs neural networks inspired by the human brain to process complex information. While AI has made remarkable progress, it remains a tool for pattern recognition and prediction, not true understanding or reasoning like humans.
A Brief History of Artificial Intelligence
1940s-1956: The Birth of AI
Roots in cybernetics (Norbert Wiener), neural models (Warren S. McCulloch and Walter Pitts), and Turing's 1950 paper 'Computing Machinery and Intelligence'. The pivotal moment: the 1956 Dartmouth Conference, organized by McCarthy, Minsky, Rochester, and Shannon, officially coins the term 'Artificial Intelligence'.
1957-1974: Early Optimism & First Golden Age
Sybolic AI and 'good old-fashioned AI' (GOFAI) dominate: Logic Theorist (1956), General Problem Solver (1959), perceptrons (Rosenblatt, 1958). Governments pour money in. ELIZA (1966) and SHRDLU (1970) impress with natural-language and block-world reasoning. Overhype leads to unrealistic promises.
1974-1980: First AI Winter
Funding dries up after the Lighthill Report (UK) and DARPA cuts. Perceptrons book (Minsky & Papert, 1969) exposes limitations of single-layer neural nets.
1980-1987: Expert Systems Boom
Second wave fueled by commercial expert systems (XCON, MYCIN). Japan's Fifth Generation project and U.S. response spark huge investment. Lisp machines flourish.
1987-1993: Second AI Winter
Expert systems prove brittle and expensive to maintain; Lisp machine market collapses; funding crashes again.
1993-2011: Rise of Machine Learning
Computing power and data explode. Key milestones:
- 1997 Deep Blue beats Kasparov
- 2000s: Support Vector Machines, Random Forests, kernel methods
- 2006: Geoffrey Hinton's 'deep belief networks' revive neural nets
- 2011: IBM Watson wins Jeopardy
2012-2020: Deep Learning Revolution
AlexNet (2012 ImageNet victory) marks the deep learning tsunami. GPUs + big data + ReLU + dropout unleash convolutional and recurrent networks. Milestones:
- 2014: GANs (Goodfellow)
- 2016: AlphaGo beats Lee Sedol
- 2017: Transformer architecture ('Attention is All You Need')
- 2018-2020: BERT, GPT-2, GPT-3 scale language models to hundreds of billions of parameters
2021-2025: The Generative AI Era & AGI Debate
ChatGPT (Nov 2022) brings AI to the masses. Explosion of multimodal models (GPT-4, Gemini, Grok-1/2, Claude 3, Llama 3, Midjourney, Sora). Key themes:
- Scaling laws hold (billion > trillion-parameter models)
- RLHF aligns LLMs with human values
- Open-source vs closed-source wars (Meta Llama series vs OpenAI/Anthropic)
- Heavy investment in AGI labs (OpenAI, Anthropic, xAI, Google DeepMind)
- Regulatory efforts (EU AI Act, U.S. executive orders)
- Rapid integration into every industry and daily life
2025 Snapshot
AI research is no longer confined to narrow tasks; frontier labs openly pursue Artificial General Intelligence (AGI). Reasoning capabilities in models are improving fast (o1-style chain-of-thought, test-time compute), robotics is accelerating (Figure, Tesla Bot), and agentic systems (AI that can plan and execute multi-step tasks) are emerging. The field is simultaneously the most exciting and most controversial it has ever been.
In ~75 years, AI has gone from philosophical thought experiment, to rule-based systems to statistical learning, to deep neural scaling, to today's race toward general intelligence. The next chapter is being written right now.
Broad Categories of Artificial Intelligence
Artificial Narrow Intelligence (ANI) / Weak AI / Narrow AI
The bulk of AI in use today falls into the Narrow AI category. Usually, these machines are designed with a specific task in mind. Some example use cases could be; recommendation algorithms on streaming platforms, image/sound recognition, voice assistants, email smart replies, vehicle autopilot, writing code, deepfakes, and fraud detection. You can think of Narrow AI as an incredibly complex mathematical calculator. It can solve incredibly complex problems in almost an instant, but it can't write poetry or taste food. It's specialist (narrow), not a generalist, and represents almost 100% of the AI you interact with as of 2025.
Artificial General Intelligence (AGI) / Strong AI / General AI
This is the future goal of AI. It can do any intellectual task a human can, and is not a current reality - at least as far as the general public is aware. This conceptual form of AI would posses human-like cognitive abilities, allowing it to understand, learn, reason, creativity, and apply its intelligence to a broad spectrum of tasks. For a machine to be deemed AGI, it must be able to complete any intellectual task that a human can with absolute autonomy. Many experts predict the arrival of AI between 2030-2040.
Artificial Super Intelligence (ASI) / Super AI / Super-intelligent AI
Further to AGI, exists Super AI. A form of intelligence that not only matches human intellect, but surpasses it by an order of magnitude. Hypothetically, this machine would outperform humans in almost (if not all) cognitive domains. Including fields such as scientific creativity, general wisdom, strategic thinking, social skills, as well as every other intellectually relevant field.
Sub-fields of AI Systems
Various sub-fields of AI systems have seen an explosion since 2020. Fuelled by massive data availability, computational power, and investment. As far as I can make of it, there are 5 distinct subsets worthy of note:
Machine Learning
The most important sub-field to date and serves as the core engine of modern AI. In essence, you give a the given system data and let it figure out the patterns autonomously. Within this field, there are three main categories:
- Supervised learning: An algorithm is trained on a dataset where each input is linked to a correct answer or label. The model then learns to match inputs to outputs (or data to labels).
- Unsupervised learning: An algorithm that runs counterpart to supervised learning. It's all about finding hidden structures or anomalies within datasets. It differs from supervised learning, labels are not provided. This means there are no right answers to guide you; the algorithm explores the data on its own. As one might imagine, this has enormous impact on discovery tasks. Especially where labels are expensive, unavailable, or unknown. Experts estimate 90% of real-world data is unlabeled.
- Reinforcement learning: In practice, this method is akin to training a dog with treats (i.e. teaching an AI to learn by doing). The core idea is simple:
- Agent observes the current state
- Agent takes an action
- Environment gives a reward (positive or negative) and moves to a new state
- Repeat until the episode ends (game over, task complete, etc.)
Deep learning / Neural Networks
A subset of ML using neural networks with multiple layers to process complex data, often inspired by the human brain. The 'Deep' in Deep Learning, refers to the multiple layers of a given model (a software program that learns from data to perform tasks like classification, prediction, analysis, and generation. Often used in autonomous vehicles and facial recognition).
Natural Language Processing (NLP)
A subfield of Artificial Intelligence that enables computers to understand, interpret, generate, and interact with human language in a way that is both meaningful and useful. In 2011, it was a niche academic field. But has since become a core part of AI, powering search engines, virtual assistants, legal discovery, medical diagnosis, coding copilots, and almost every modern software product. Some notable breakthroughs were:
- Word embeddings: In 2013 word embeddings were created. They have quickly become the foundation of almost all modern NLP. They're dense vectors (e.g. 300 numbers) that represent the meaning of a word in a way computers can understand. They map words to points in high-dimensional space where distance reflects semantic similarity, learned automatically from patterns in text. Transformers use self-attention to let every word directly look at every other word in the sentence and decide 'how much should I care about this word right now?'
- Transformers: In 2017, transformer architecture made it so that machines could execute near-human performance on many tasks. Pioneered in the paper 'Attention is All You Need', the architecture allows neural networks to process entire sequences (like sentences) all at once, instead of word-by-word like older models (RNNs or LSTMs). For each word, the model computes three vectors: Query (Q) 'what am I looking for?', Key (K) 'what do I have to offer?', Value (V) 'here's the actual information.' It then calculates attention scores.
- Foundation Models: Large neural networks (usually transformers) trained on enormous, broad datasets in a self-supervised or unsupervised way. Once trained, they serve as a 'foundation' that can be adapted to many different tasks with little or no additional training. Some examples of foundation models are Large Language Models (LLMs), Vision Foundation Models, Multimodal (text and vision), Image Generation (text > image/image > image), Audio/Speech Models, Music Foundation Models, and Code Foundation Models to name a few. Most of the models used these days are some combination of various models by default.
Computer Vision
This involves teaching machines to interpret and understand visual information from the world, this could be either images or videos. It uses algorithms to detect objects, recognize faces, or segment scenes, relying on techniques like edge detection, feature matching, and learning models like CNNs for high accuracy. Everyday uses include autonomous vehicles analyzing road scenes or medical imaging for tumor detection.
Robotics
Combines AI with mechanical engineering to create intelligent physical agents. It includes perception (using computer vision and sensors), planning (deciding actions), and control (moving in real world environments). Applications range from factory automation and warehouse robots (Boston Dynamics, Tesla Bot) to surgical robots and household assistants.
Applications and Impact
AI is integrated into everyday life and industries:
- Healthcare: Diagnosing diseases from scans or predicting outbreaks.
- Finance: Fraud detection and algorithmic trading.
- Entertainment: Generating art, music, or personalized content.
- Transportation: Optimizing routes or enabling autonomous drones.
While AI offers efficiency and innovation, it raises ethical concerns like job displacement, bias in algorithms, privacy issues, and the need for regulation to ensure safe development.