The history of artificial intelligence: from first ideas to today

Why the history of AI matters for security too

On a cybersecurity blog, the history of artificial intelligence might seem like a detour. It's the opposite — today's AI systems are simultaneously a new defensive tool (anomaly detection, SOC automation) and a new attack surface (prompt injection, deepfakes, AI-generated phishing). Understanding why these threats are fundamentally different from previous ones is helped by knowing where AI actually came from — and that its development was anything but a straight line.

Roots: thinking machines before computers

The idea of machines capable of "thinking" is far older than computers. Philosophers like Descartes and Hobbes were already discussing the mind as a mechanical process back in the 17th century. The real foundation for modern AI, however, was laid by two papers in the 1940s:

• Warren McCulloch and Walter Pitts (1943) — proposed a mathematical model of the neuron, showing that networks of simple switching units could, in principle, compute any logical function. This is the conceptual ancestor of today's neural networks.
• Alan Turing (1950) — in his paper "Computing Machinery and Intelligence," posed the question "Can machines think?" and, instead of a philosophical debate, proposed a practical test: if a human in conversation can't decide whether they're talking to a machine or a person, the machine can be considered intelligent. Known today as the Turing test.

The birth of a field: Dartmouth, 1956

The term "artificial intelligence" itself was coined in the summer of 1956 at Dartmouth College in New Hampshire, where John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon organized a two-month workshop with an openly ambitious goal: to explore the conjecture that "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it."

This event is considered the official beginning of AI as a scientific field. The workshop's participants — McCarthy, Minsky, Allen Newell, and Herbert Simon — went on to become the field's dominant figures for decades.

The golden years of symbolic AI (1956 – 1974)

The following eighteen years brought a wave of optimism and genuine breakthroughs in what's known as symbolic AI — an approach that represented knowledge as logical symbols and rules, rather than as numbers or weights.

• The Perceptron (Frank Rosenblatt, 1958) — the first trainable neural model capable of learning to recognize simple patterns from examples. It was presented as a direct step toward machines that "see and recognize."
• The General Problem Solver (Newell and Simon, 1957) — a program designed to solve any formally describable problem using universal heuristics.
• ELIZA (Joseph Weizenbaum, 1966) — a simple program simulating a psychotherapist, which came across as surprisingly convincingly "intelligent" based purely on pattern matching and rephrasing the user's sentences. It demonstrated how readily people attribute understanding to machines that don't actually have any — a phenomenon still relevant to chatbots today.
• SHRDLU (Terry Winograd, 1970) — a program that understood natural language within a limited "blocks world," able to plan and answer questions about its own actions.

Funding flowed mainly from American military agencies (e.g. DARPA), in the belief that full-fledged machine intelligence was a matter of years, not decades.

The first AI winter (1974 – 1980)

The promises weren't fulfilled on the expected timeline, and reality caught up with the optimism. Two key events triggered the first major pullback in funding and interest, known today as the first AI winter:

• The book "Perceptrons" (Minsky and Papert, 1969) — mathematically proved that single-layer perceptrons couldn't solve even trivial problems like the logical XOR function. Although this limitation could be overcome with multi-layer networks, there was no efficient way to train them at the time — and interest in neural networks waned for years.
• The Lighthill Report (1973) — a report for the British government that critically concluded AI research had failed to meet its "grandiose objectives" and recommended drastic funding cuts. American agencies followed a similar path.

Expert systems and the second boom (1980 – 1987)

AI recovered thanks to a shift from "general intelligence" toward narrowly focused, commercially useful expert systems — programs that encoded the knowledge of specific human experts into sets of "if-then" rules.

• MYCIN (Stanford, mid-1970s, deployed in the 1980s) — an expert system for diagnosing bacterial infections and recommending antibiotics, which in tests matched or beat experienced physicians on accuracy.
• XCON / R1 (Digital Equipment Corporation, 1980) — configured computer systems according to customer orders and saved the company tens of millions of dollars a year, convincing the industry of AI's commercial value.
• Japan's Fifth Generation Computer project (1982) — a massive state investment by Japan into computers built on logic programming, which triggered competing investment in the US and Europe.
• Backpropagation (Rumelhart, Hinton, Williams, 1986) — the popularization of the error-backpropagation algorithm, which finally made it possible to efficiently train multi-layer neural networks, overcoming the limits set out in the "Perceptrons" book. This technique remains the foundation of training deep neural networks to this day.

The second AI winter (late 1980s – early 1990s)

Expert systems ran into their own limits — they were expensive to maintain, brittle outside their narrowly defined domain, and incapable of learning from experience. The market for specialized AI hardware (so-called LISP machines) collapsed once it was overtaken by cheaper general-purpose workstations. Japan's Fifth Generation project failed to reach its promised goals. Funding withdrew again, and the term "artificial intelligence" became almost toxic in industry for a while — companies preferred to talk about "machine learning" or "data analysis" instead.

The quiet revolution: statistical machine learning (the 1990s)

Despite the winter, a new, less media-friendly but longer-lasting current of research was growing beneath the surface — statistical machine learning, which derived patterns directly from data instead of relying on hand-written rules.

• Support vector machines and Bayesian networks — mathematically elegant classification methods that became the workhorse of applied machine learning for years to come.
• Deep Blue defeats Kasparov (1997) — an IBM supercomputer specialized in chess beat the reigning world champion, Garry Kasparov, in a formal match. A symbolic moment demonstrating the power of raw computational force combined with domain-specific search, not "general" intelligence.
• The growth of the internet — created, for the first time in history, an enormous amount of digital data (text, images, user behavior) that would later become fuel for modern AI.

Big data and the return of neural networks (2000 – 2011)

Three trends quietly converged during this decade and set the stage for the later explosion: cheaper compute (especially graphics cards, GPUs, originally developed for video games), a massive increase in available data thanks to the internet and social media, and improved algorithms. A key milestone was the creation of the ImageNet database (2009) — more than 14 million hand-labeled images across thousands of categories, which became the standard benchmark for computer vision and directly enabled a breakthrough a few years later.

The deep learning revolution (2012 – 2017)

This period is now widely recognized as the start of the modern AI era we know today.

• AlexNet (2012) — a deep convolutional neural network that decisively beat every competing approach to image recognition at the ImageNet competition, trained on gaming GPUs. This moment convinced the entire research field that deep neural networks trained on large amounts of data outperform hand-engineered methods.
• Word2Vec (2013) — a technique that taught computers to represent word meaning as numerical vectors, laying the groundwork for modern natural language processing.
• Generative Adversarial Networks — GANs (Ian Goodfellow, 2014) — an architecture of two competing networks that enabled generating surprisingly realistic synthetic images. A direct ancestor of today's deepfakes.
• AlphaGo defeats Lee Sedol (2016) — a DeepMind system beat one of the world's best Go players at a game that, due to its combinatorial complexity, had until then been considered decades away from being mastered by computers.
• The Transformer — "Attention Is All You Need" (Google, 2017) — an architecture that replaced the previously dominant recurrent networks with an "attention" mechanism, enabling text to be processed in parallel and long-range dependencies to be learned more effectively. This is the architectural foundation of absolutely every modern large language model.

The era of large language models (2018 – 2022)

The Transformer architecture kicked off a scaling race — training ever-larger models on ever-larger amounts of text.

• BERT (Google, 2018) — a model that significantly improved contextual understanding of text and quickly became part of real products, including search.
• The GPT series (OpenAI, 2018 – 2020) — GPT-1, GPT-2, and finally GPT-3 (2020, 175 billion parameters) demonstrated that simply scaling up model size and data leads to surprisingly general capabilities — from writing text to basic programming — without the model ever being explicitly trained for those tasks.
• Text-to-image models — DALL·E (2021), and later Stable Diffusion and Midjourney (2022), brought diffusion models capable of generating photorealistic images from a text description, which sharply democratized both the technology and the accompanying concerns about disinformation and deepfakes.
• ChatGPT (November 2022) — not a fundamentally new model, but an interface and a fine-tuning process (using human feedback, RLHF) that made large language models accessible to ordinary users in a simple conversational form. It gained hundreds of millions of users within months and definitively brought AI into mainstream awareness.

Today: multimodal and agentic systems (2023 – 2026)

Development after ChatGPT proceeded along several parallel tracks that together shape the AI we know today:

• Multimodality — models like GPT-4, Claude, and Gemini stopped being limited to text alone and learned to understand images, audio, and video within a single unified system.
• "Reasoning" models — a new generation of models that internally breaks a problem down into steps before answering, instead of generating a response directly, which significantly improved performance on math, logic, and coding tasks.
• Agentic AI — a shift from models that merely answer questions to systems that can independently use tools, browse the web, run code, and carry out multi-step tasks with minimal human oversight — this exact class of system is what powers tools like the one that wrote this article.
• Open-weight models — models whose parameters are publicly released (e.g. the Llama, Mistral, and DeepSeek families) made it possible to run advanced AI locally, outside the large cloud providers — with both the benefits and the security risks that brings.
• Regulation — the EU AI Act (2024) and similar legislative frameworks in other countries are responding to AI's rapid growth in capability with requirements for transparency, risk assessment, and restrictions on high-risk uses.

Timeline: key milestones

AI and cybersecurity: a double-edged sword

Back to the home turf of this blog — the current wave of AI is fundamentally reshaping cybersecurity too, in both directions:

• On the defense side — AI models today analyze enormous volumes of network traffic and logs faster than a human team, detect anomalies characteristic of attacks (see our article on the most common cyberattacks), and automate the first response inside SOC centers.
• On the attack side — those same capabilities make it possible to generate more convincing phishing emails without the usual language mistakes, create voice and video deepfakes to bypass identity verification, or automate the search for vulnerabilities in code.
• A brand-new class of vulnerabilities — systems built on large language models have also introduced entirely new security problems that essentially didn't exist before 2022 — for example prompt injection, where an attacker hides instructions for an AI system inside a web page, document, or email, aiming to make it act against the user's interests.

Summary

From Turing's question "can machines think?" in 1950, through two harsh AI winters, the quiet revolution of statistical machine learning in the 1990s, to the deep learning explosion after 2012 and large language models from 2018 onward — the path to today's AI was neither straight nor inevitable. It was a series of mathematical breakthroughs, fortunate coincidences (the availability of cheap GPUs, a massive increase in data from the internet), and repeated cycles of inflated expectations followed by sobering up. Understanding this history is the best safeguard against taking today's progress — and its risks — uncritically.