Originally published here.

At a time when there are small AI updates every week and seemingly major AI updates every month, it is easy to forget where we come from and how much AI has progressed over the years. That’s why I have put together a brief timeline of AI, starting from the 1950s to today. With a strong sense that a even more incredible updates are just around the corner, so now feels like a good time to ground ourselves in the story so far.

1930: Natural language processing (NLP) traces its origins to the 1930s, beginning with early efforts in machine translation. The field gained real momentum after Warren Weaver’s influential 1949 memorandum, Machine Translation of Languages: Fourteen Essays, which sparked significant progress and excitement.

1950: Computer Scientist Alan Turing proposes a test for Machine Intelligence. The machine is intelligent if it can trick a human into believing it is also human.

1954: In the early 1950s, machine translation gained momentum with institutional support and media hype, culminating in the 1954 Georgetown–IBM demo, an experiment jointly set up by the two organisations. Though widely celebrated, the system translated just 49 pre-selected Russian sentences to English using a 250-word vocabulary - far from real-world fluency.

1955: Computer Scientist John McCarthy coins the term Artificial Intelligence to describe “the science and engineering of making intelligent machines”.

1960: Marvin Minsky publishes “Steps Towards Artificial Intelligence”, championing a rule-based approach, proposing that intelligence can be modelled through symbols and rules. Mental processes are broken down into discrete symbolic operations (e.g., IF-THEN rules), much like logical reasoning. He went on to found the MIT AI Laboratory.

1964: A chatbot called ELIZA founded by Joseph Weizenbaum at MIT has first conversations with humans.

1970-1980: The first AI Winter: Many failed approaches lead to a drastic decline in enthusiasm and funding.

1980: Focus shifts to more narrow expert systems, applied to medical and financial analysis. By 1985, global companies spend over $1 billion on in-house expert systems, fuelling a supporting industry of software and hardware firms. Companies like Symbolics and Lisp Machines are building tools and specialised machines tailored for AI development in the LISP programming language and software.

1986: Hinton, Rumelhart and Williams publish “Learning Representations by Backpropagating Errors”, enabling the training of deeper neural networks. This approach allowed for efficient training of multi-layer neural networks by propagating error gradients backwards through the layers. Their work reignited interest in neural networks and enabled them to solve more complex tasks.

1987–2000: The second AI Winter: AI interest drastically decreases once again, following the crash of the LISP machine market and disillusionment with expert systems, which proved brittle, costly, and hard to maintain. Simultaneously, large-scale initiatives like Japan’s Fifth Generation Project (a $850 million government-led effort to develop advanced AI systems capable of human-like reasoning, natural language understanding, and logic programming) and the Strategic Computing Initiative (a US Department of Defense programme aimed at advancing AI for military applications) failed to deliver on their ambitious goals, prompting further funding cuts and a broader scepticism toward AI.

1997: Built by IBM, Deep Blue is the first computer to defeat a reigning world chess champion, Garry Kasparov, in a match under standard tournament conditions in 1997.

LSTM (Long Short-Term Memory): Hochreiter & Schmidhuber introduce a form of recurrent neural network that can learn long-term dependencies.

2002: Roomba becomes the first robot vacuum cleaner developed by iRobot. It was one of the first widely adopted household robots, designed to autonomously clean floors using sensors, simple AI algorithms, and robotic mobility

2009: Fei-Fei Li and her research team create ImageNet, a large-scale visual database designed for use in image recognition research. From 2010 to 2017, ImageNet hosted an annual challenge where teams competed to build models that could classify and detect objects in images.

2011: Apple launches virtual assistant Siri.

IBM’s question-answer computer Watson wins first prize in the television quiz show Jeopardy.

2012: In 2012, a deep learning model by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton called AlexNet won the ImageNet challenge created by Fei-Fei Li years earlier with a huge margin. This success sparked the deep learning revolution in computer vision. Unlike previous models, AlexNet was trained on GPUs (NVIDIA GTX 580), enabling much faster computation.

2013: Word2Vec is a neural network-based model developed by Tomas Mikolov and colleagues at Google in 2013 that learns vector representations of words (known as word embeddings) in such a way that semantic relationships between words are captured in a continuous vector space. It’s an unsupervised learning algorithm trained on large text corpora.

2014: Eugene Goostman, a fictional AI chatbot developed by Vladimir Veselov, Eugene Demchenko, and Sergey Ulasen becomes widely known after it claims to have "passed the Turing Test" in 2014.

Amazon launches voice assistant Alexa.

Ian Goodfellow introduces GANs (Generative Adversarial Networks): two neural nets compete, producing realistic images and data.

2015: AlphaGo, an artificial intelligence program developed by DeepMind becomes the first AI to defeat a world champion in the board game Go, a feat long considered a grand challenge in AI.

2017: The "Attention is All You Need" paper, published by Vaswani et al. in 2017, introduces the Transformer architecture. This is a foundational breakthrough in AI that replaces the older use of recurrence (like LSTMs) with self-attention mechanisms. It most notably enables the parallel processing of sentences or images rather than the sequential processing. This change radically improved the scalability, speed, and performance of language models and later generative AI systems.

2018: GPT by OpenAI is the first Generative Pretrained Transformer that shows promise of pretraining and fine-tuning. GPT-2 and GPT-3 scale up parameters - text generation becomes strong.

2021: CLIP (Contrastive Language–Image Pretraining) learns to connect text and images by training on hundreds of millions of image–caption pairs from the internet. DALL·E uses similar principles, but in reverse: it generates images from text instead of understanding them. Both are created by OpenAI.

2022: Stable Diffusion: Open-source diffusion models make text-to-image generation widely accessible. A diffusion model generates data by learning to reverse a noisy process, gradually turning random noise into structured, high-quality outputs.

OpenAI launches ChatGPT that combines GPT-3.5 with Reinforcement Learning from Human Feedback (RLHF) to align dialogue-based outputs. ChatGPT becomes one of the fastest adopted technologies, reaching 100m users 5x faster than Instagram or Tiktok.

2024: GPT-4, Claude, Gemini: Multimodal, more aligned, and more capable LLMs mark the beginning of truly general-purpose AI agents.

2025: OpenAI and Gemini Deep Think achieve gold-medal standard at the International Mathematical Olympiad.

2027: To be continued.