Artificial intelligence (AI) is a branch of computer science that builds systems capable of performing tasks that normally require human intelligence, such as reasoning, perception, language understanding, and learning from data. This post traces the evolution of AI from its origins in the 1950s to the present, covering its key milestones: rule-based systems, neural networks, the deep learning revolution, and modern applications.
The Origins of AI: Rule-Based Programming and Symbolic Reasoning (1950s)
It started, as these things often do, with a small group of researchers asking an audacious question: could a machine think? In the 1950s, they set out to build systems that mirrored human intelligence, with no real idea how far the idea would eventually go.
Midjourney prompt: A computer and scientists in the year of 1955
I use the following image style for this post prompts: realistic, pastel, pink and metallic tones, stunning, — stylize 1000
| Interested in Midjouney image generation? - refer to my post Mastering Midjourney Prompts for Stunning Images. You can also check the Guide Midjourney --stylize Explained. |
 |
Elena: Were computers in 1955?
chatGPT: Yes, computers were in existence in 1955. Computers were large, room-sized machines used for scientific calculations, data processing, and other computational tasks during that time. They were not as advanced or widely accessible as modern computers but played a crucial role in early AI research and development.
.jpg)
Living Large -- Argonne's First Computer, Wikimedia Commons image
The tools they had were basic by today’s standards: rule-based programming and symbolic reasoning, hand-coding human thought processes step by step. It was slow going, but every small success suggested there might be a lot more to find.
Midjourney prompt: Rule-based programming and symbolic reasoning
Those early years set the pattern for everything that followed: bursts of genuine progress, followed by long stretches where nobody was quite sure the whole idea would work.
For those who prefer their history with dates attached, the decade delivered a remarkable run of firsts. In 1950, Alan Turing published “Computing Machinery and Intelligence”, proposing his famous imitation game — now known as the Turing test — as a practical substitute for the slippery question “can machines think?”. In the summer of 1956, John McCarthy, Marvin Minsky, Claude Shannon and Nathaniel Rochester convened the Dartmouth Summer Research Project, the workshop at which McCarthy coined the very term “artificial intelligence”. That same year, Allen Newell, Herbert Simon and Cliff Shaw demonstrated the Logic Theorist, widely regarded as the first AI program, which proved 38 of the first 52 theorems in Whitehead and Russell’s Principia Mathematica — and even found a more elegant proof for one of them.
The dominant paradigm that emerged — later nicknamed “Good Old-Fashioned AI” (GOFAI) — represented knowledge as explicit symbols and manipulated them with hand-written inference rules, typically IF-THEN productions. To support this style of programming, McCarthy created Lisp in 1958, which remained the lingua franca of AI research for decades. In the same year, Frank Rosenblatt took the opposing, data-driven route: his perceptron computed a weighted sum of inputs, applied a threshold, and adjusted its weights from labelled examples — the first trainable neural network, implemented in hardware as the Mark I Perceptron. The two rival traditions, symbolic reasoning and learning from data, were thus both born within the same decade. Symbolic AI’s ambitions culminated in projects such as Shakey the Robot (SRI International, 1966–1972), pictured below, which combined computer vision, the STRIPS automated planner, and the A* search algorithm — an algorithm still taught in every computer science curriculum today.
_-_Computer_History_Museum_(2007-11-10_23.16.01_by_Carlo_Nardone).jpg)
Shakey the Robot (developed between 1966-1972 at SRI International) - Computer History Museum, Wikimedia Commons image
The 1980s AI Renaissance: Neural Networks and Machine Learning
Progress after the 1950s was not a straight line. The initial excitement gave way to a long stagnant patch, before AI research picked back up again in the 1980s.
Growing computational power helped, but the real story of the decade is neural networks re-entering the picture. A neural network is a machine learning model composed of layers of interconnected nodes that learn patterns directly from data, rather than from hand-written rules. That idea — learning from data instead of coding rules by hand — is the one that eventually won out and shaped everything AI became afterwards.
The “stagnation” mentioned above has a concrete history, and it is worth naming. In 1969, Minsky and Papert’s book Perceptrons proved that a single-layer perceptron cannot learn functions that are not linearly separable — the XOR function being the canonical example — and neural network funding dried up almost overnight. The 1973 Lighthill Report then persuaded the British government to slash AI research funding, ushering in what we now call the first AI winter. Progress had not stopped so much as it had been defunded.
What actually revived the field commercially in the early 1980s was the expert system: a symbolic AI program encoding a specialist’s knowledge as hundreds or thousands of IF-THEN rules. MYCIN, developed at Stanford in the 1970s, used roughly 600 rules and certainty factors to recommend antibiotic treatments for blood infections, performing comparably to human specialists in evaluations. XCON (also known as R1), deployed at Digital Equipment Corporation from 1980 to configure VAX computer orders, was reportedly saving the company around $25 million a year by 1986 — the first proof that AI could pay its own way.
On the neural network side, the theoretical breakthroughs of the decade were just as consequential. John Hopfield’s 1982 recurrent network showed how a neural system could store and retrieve memories as stable states, re-legitimising the connectionist approach. Then, in 1986, Rumelhart, Hinton and Williams published their Nature paper popularising backpropagation: by applying the chain rule of calculus to propagate error gradients backwards through hidden layers, multi-layer networks could finally learn the non-linear functions that had defeated the single-layer perceptron. This solved the very limitation Minsky and Papert had identified seventeen years earlier — although, in a twist of historical irony, the collapse of the specialised Lisp machine market in the late 1980s promptly triggered a second AI winter before the idea could fully bloom.
Key AI Milestones: A Timeline from the 1950s to Today
Here are key milestones in AI development, along with the year in which each milestone occurred:
Midjourney prompt: Deep Neural Networks and GPU
The Deep Learning Revolution: GPUs and Big Data (21st Century)
Deep learning is a subfield of machine learning that uses multi-layer neural networks to learn hierarchical representations directly from raw data, and it is the primary driver of the 21st-century AI breakthroughs in image recognition and natural language processing. This is where AI stopped being a research curiosity and started being something you could actually use.
The CSIRO GPU cluster at the data centre, ikimedia Commons image
The pivotal moment has a precise date: September 2012, when AlexNet — an eight-layer convolutional network with about 60 million parameters, designed by Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton — won the ImageNet Large Scale Visual Recognition Challenge with a top-5 error rate of 15.3%, against 26.2% for the nearest competitor. That ten-point margin, unheard of in a benchmark competition, sent the entire computer vision community scrambling to retrain on neural networks. AlexNet’s recipe combined ReLU activations, dropout regularisation, and — crucially — training on a pair of consumer NVIDIA GTX 580 GPUs. GPUs matter because deep learning is, computationally, mostly matrix multiplication, and a graphics card’s thousands of cores perform those multiplications in parallel far more cheaply than a CPU ever could. The other ingredient was data: ImageNet supplied 1.2 million labelled training images across 1,000 categories, a scale of supervision that simply had not existed before.
The years that followed read like a highlight reel. Word2vec (2013) showed that word meanings could be captured as arithmetic on vectors; sequence-to-sequence models (2014) reinvented machine translation; ResNet (2015) stacked 152 layers using skip connections and pushed ImageNet top-5 error below the estimated human level of roughly 5%. In March 2016, DeepMind’s AlphaGo — combining deep policy and value networks with Monte Carlo tree search — defeated Lee Sedol 4–1 at Go, a game whose branching factor had long made it the standard example of what computers could not do. Then, in 2017, Google researchers published “Attention Is All You Need”, introducing the transformer architecture: by replacing recurrence with self-attention, it allowed every token in a sequence to be processed in parallel, making it dramatically more efficient to train at scale. Essentially every large language model since — including the one I have been chatting with throughout this post — is a transformer.
Since my keywords promised you GPT parameter counts, here they are — the scaling story in one small table:
| Model | Release | Parameters | Notable detail |
|---|---|---|---|
| GPT-1 | June 2018 | 117 million | Introduced generative pre-training on unlabelled text |
| GPT-2 | February 2019 | 1.5 billion | Initially withheld over misuse concerns |
| GPT-3 | June 2020 | 175 billion | 96 layers; demonstrated few-shot learning |
| GPT-4 | March 2023 | undisclosed | OpenAI stopped publishing parameter counts |
ChatGPT itself, launched in November 2022, is not a bigger model but a better-behaved one: a GPT-3.5-series model fine-tuned with reinforcement learning from human feedback (RLHF), the technique OpenAI introduced with InstructGPT, in which human preference rankings train a reward model that steers the network towards helpful answers. It reached an estimated 100 million monthly active users within two months — the fastest-growing consumer application in history at the time, according to UBS analysis of Similarweb data. On the image side, the pictures decorating this very post come from the same revolution: DALL·E 2 (April 2022), Stable Diffusion (August 2022) and Midjourney are all diffusion models, which learn to reverse a gradual noising process, generating images by iteratively denoising random pixels under text guidance.
Midjourney prompt: The unyielding spirit of human curiosity
Elena: Why AI cannot draw human hands properly? Something always wrong with the fingers?
chatGPT: Drawing human hands accurately and realistically is a challenging task for AI due to several reasons. One of the primary reasons is the complex and intricate nature of human hands. Hands possess a wide range of movements and configurations, and their structure involves numerous joints, bones, muscles, and tendons. Capturing the subtle nuances and intricacies of these elements accurately requires a deep understanding of human anatomy, perspective, and proportions.
Additionally, human hands exhibit a high level of variability among individuals. Factors such as hand size, finger length, and joint flexibility can vary significantly from person to person. AI models typically learn from large datasets to generalize patterns, but the vast diversity in hand appearances makes it difficult to capture the precise details of every individual hand.
Another challenge lies in the interpretation and translation of the three-dimensional structure of hands onto a two-dimensional drawing surface. AI models often struggle to perceive depth and accurately represent the spatial relationships between the various parts of a hand. This can result in distorted or unnatural-looking hand shapes, especially when attempting to capture complex poses or foreshortening.
Lastly, the lack of specific training data for hands or limited representation of diverse hand poses and gestures in training datasets can also contribute to AI's difficulty in drawing human hands realistically. Training AI models with more extensive and diverse hand-related datasets can help improve their accuracy in hand drawing tasks.
While AI has made significant progress in generating realistic images and understanding human anatomy, achieving flawless hand drawings remains an ongoing challenge due to the complexity and individuality of human hands.
The other big shift was accessibility. Big data and cloud computing meant that AI tools stopped being the preserve of well-funded labs, and became something individuals and small businesses could actually use.
State of AI Art Machine Learning Models, Wikimedia Commons image
Modern AI Applications and Future Directions: Virtual Assistants, Autonomous Vehicles, Healthcare
AI tools are now woven into daily life, mostly without us noticing: virtual assistants on our phones, recommendation systems on every shopping site, and a growing list of tasks quietly handled by models running in the background.
Behind those enchanting whispers sits some rather concrete machinery. Virtual assistants such as Siri (2011) and Alexa (2014) chain together on-device wake-word detection, cloud-based speech recognition, and natural language understanding. Recommendation systems trace their modern lineage to the Netflix Prize (2006–2009), which popularised matrix factorisation for collaborative filtering — predicting your taste from the ratings of people statistically similar to you — though by now the large platforms have largely moved on to deep learning models. Autonomous driving is graded on the SAE’s six levels (0–5); at the time of writing, Waymo operates Level 4 driverless taxis in Phoenix, while most consumer “autopilot” systems remain firmly at Level 2, with the human legally in charge. In healthcare, DeepMind’s AlphaFold 2 achieved a median accuracy of 92.4 GDT at the CASP14 assessment in 2020 — effectively solving the 50-year-old protein structure prediction problem — and in 2022 released predicted structures for over 200 million proteins, essentially every catalogued protein known to science.
The story is far from finished. Researchers keep pushing on explainable AI (understanding why a model made the decision it did), on autonomous vehicles, and on healthcare diagnostics — three areas where getting it wrong has real consequences, so progress tends to be slower and more cautious than the hype suggests.
Ethical considerations deserve real attention here, not a token paragraph: privacy, fairness, and accountability are not solved problems, and the balance between moving fast and moving responsibly is still being negotiated in real time, by regulators, companies, and the people building these systems.
I don’t know exactly where this goes next — nobody does, whatever the confident predictions on your feed might suggest. But the pattern from the last seventy years holds: real progress, interspersed with periods where it looks like nothing is happening, followed by another leap nobody quite saw coming.
Midjourney prompt: Self-driving car
p.s. I like the idea of sitting on the top of the self-driving car, with the wind in my hair and good weather conditions. Although, I would have used good sun protection :)
It is funny how AI created the image of a self-driving car with a place for a human on it. I am concerned about the safety, though.
I have linked research papers and educational material about the AI milestones inside the timeline. I hope that you do not mind. I will update these post links soon.
Did you like this post? Please let me know if you have any comments or suggestions.
Posts about AI that might be interesting for youReferences
1. Guide Midjourney –stylize Explained
2. Mastering Midjourney Prompts for Stunning Images
3. New Chat (chatGPT by OpenAI)
4. ChatGPT sets record for fastest-growing user base - analyst note (Reuters)
Enjoyed this? Get more like it.
Weekly notes on AI tools, Python, and what I'm actually building — plus two free gifts: the 15-page Fantastic AI: The 2026 Toolkit and a Git Commands & Contribution Workflow Cheatsheet.
