Timeline

1950s

1950Paper

Turing Test: Can Machines Think?

Alan Turing publishes 'Computing Machinery and Intelligence', proposing the famous Turing Test — a way to judge whether a machine can exhibit intelligent behavior indistinguishable from a human. This paper laid the philosophical foundation for the entire field of AI.

Computing Machinery and Intelligence (1950)

1950Paper

Turing Test: Can Machines Think?

Computing Machinery and Intelligence (1950)

1956Milestone

The Dartmouth Conference: AI Gets Its Name

John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon organize a summer workshop at Dartmouth College. This is where the term 'Artificial Intelligence' was officially coined, marking the birth of AI as an academic discipline.

A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence (1955)

1956Milestone

The Dartmouth Conference: AI Gets Its Name

A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence (1955)

1957Breakthrough

The Perceptron: First Neural Network

Frank Rosenblatt invents the Perceptron, a simple neural network that can learn to classify patterns. Implemented on custom hardware (the Mark I Perceptron), it was the first machine that could 'learn' from experience, inspiring decades of neural network research.

The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain (1958)

1957Breakthrough

The Perceptron: First Neural Network

The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain (1958)

1960s

1966Product

ELIZA: The First Chatbot

Joseph Weizenbaum at MIT creates ELIZA, a program that could carry on conversations by pattern matching. Its most famous script, DOCTOR, simulated a psychotherapist so convincingly that some users believed they were talking to a real person.

ELIZA — A Computer Program For the Study of Natural Language Communication Between Man And Machine (1966)

1966Product

ELIZA: The First Chatbot

ELIZA — A Computer Program For the Study of Natural Language Communication Between Man And Machine (1966)

1966Milestone

Shakey: The First Intelligent Robot

SRI International begins developing Shakey, the first robot that could reason about its own actions. It combined computer vision, natural language understanding, and planning — a landmark in robotics and AI integration.

Shakey the Robot (1966)

1966Milestone

Shakey: The First Intelligent Robot

Shakey the Robot (1966)

1969Paper

Perceptrons: The Book That Froze Neural Networks

Marvin Minsky and Seymour Papert publish 'Perceptrons', mathematically proving the limitations of single-layer neural networks. The book's influence was so great that funding for neural network research dried up, contributing to the first 'AI Winter'.

Perceptrons: An Introduction to Computational Geometry (1969)

1969Paper

Perceptrons: The Book That Froze Neural Networks

Perceptrons: An Introduction to Computational Geometry (1969)

1970s

1974Milestone

The First AI Winter

After years of overpromising and underdelivering, AI research faces a severe funding crisis. The UK's Lighthill Report criticizes AI's lack of progress, and DARPA cuts funding. Many researchers leave the field, and the first 'AI Winter' sets in.

Artificial Intelligence: A General Survey (Lighthill Report) (1973)

1974Milestone

The First AI Winter

Artificial Intelligence: A General Survey (Lighthill Report) (1973)

1976Product

Expert Systems: AI Goes to Work

MYCIN, developed at Stanford, demonstrates that AI systems encoding human expert knowledge can diagnose blood infections as well as doctors. Expert systems become the first commercially successful form of AI, reviving interest in the field.

Computer-Based Medical Consultations: MYCIN (1976)

1976Product

Expert Systems: AI Goes to Work

Computer-Based Medical Consultations: MYCIN (1976)

1980s

1986Paper

Backpropagation: Teaching Neural Networks to Learn

David Rumelhart, Geoffrey Hinton, and Ronald Williams publish their landmark paper on backpropagation, showing how multi-layer neural networks can be trained effectively. This breakthrough revived interest in neural networks and laid the groundwork for modern deep learning.

Learning representations by back-propagating errors (1986)

1986Paper

Backpropagation: Teaching Neural Networks to Learn

Learning representations by back-propagating errors (1986)

1987Milestone

The Second AI Winter

The expert systems market collapses as businesses find them expensive to maintain and brittle. The Lisp machine market crashes, and once again AI research funding dries up. The field enters its second major winter.

AI: The Tumultuous History of the Search for Artificial Intelligence (1993)

1987Milestone

The Second AI Winter

AI: The Tumultuous History of the Search for Artificial Intelligence (1993)

1990s

1997Milestone

Deep Blue Defeats World Chess Champion

IBM's Deep Blue defeats world chess champion Garry Kasparov in a six-game match. This was the first time a computer beat a reigning world champion under standard tournament conditions, capturing global attention for AI's potential.

Deep Blue (2002)

1997Milestone

Deep Blue Defeats World Chess Champion

Deep Blue (2002)

1997Paper

LSTM: Memory for Neural Networks

Sepp Hochreiter and Jürgen Schmidhuber introduce Long Short-Term Memory networks, solving the vanishing gradient problem that prevented neural networks from learning long sequences. LSTMs would later power speech recognition, machine translation, and more.

Long Short-Term Memory (1997)

1997Paper

LSTM: Memory for Neural Networks

Long Short-Term Memory (1997)

1998Paper

LeNet-5: Convolutional Neural Networks

Yann LeCun and team develop LeNet-5, a convolutional neural network that could read handwritten digits on checks. It became one of the first practical applications of deep learning and established the CNN architecture still used today.

Gradient-Based Learning Applied to Document Recognition (1998)

1998Paper

LeNet-5: Convolutional Neural Networks

Gradient-Based Learning Applied to Document Recognition (1998)

2000s

2001Milestone

The Rise of Statistical Machine Learning

Support Vector Machines and other statistical methods dominate AI research. The focus shifts from hand-crafted rules to learning from data. Kernel methods, random forests, and boosting become the workhorses of practical AI applications.

A Training Algorithm for Optimal Margin Classifiers (1992)

2001Milestone

The Rise of Statistical Machine Learning

A Training Algorithm for Optimal Margin Classifiers (1992)

2009Milestone

ImageNet: A Dataset That Changed Everything

Fei-Fei Li and team release ImageNet, a massive visual database of over 14 million labeled images organized by the WordNet hierarchy. The associated ImageNet Large Scale Visual Recognition Challenge (ILSVRC) would become the benchmark that sparked the deep learning revolution.

ImageNet: A large-scale hierarchical image database (2009)

2009Milestone

ImageNet: A Dataset That Changed Everything

ImageNet: A large-scale hierarchical image database (2009)

2010s

2012Breakthrough

AlexNet: The Deep Learning Big Bang

Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton's deep neural network wins the ImageNet competition by a huge margin, cutting the error rate nearly in half. This moment is widely considered the starting gun of the modern AI era, proving that deep learning works at scale.

ImageNet Classification with Deep Convolutional Neural Networks (2012)

2012Breakthrough

AlexNet: The Deep Learning Big Bang

ImageNet Classification with Deep Convolutional Neural Networks (2012)

2013Paper

Word2Vec: Words as Math

Tomas Mikolov and team at Google develop Word2Vec, which turns words into numerical vectors that capture meaning. The famous example: 'King - Man + Woman = Queen'. This technique transformed how machines understand language.

Efficient Estimation of Word Representations in Vector Space (2013)

2013Paper

Word2Vec: Words as Math

Efficient Estimation of Word Representations in Vector Space (2013)

2014Paper

GANs: AI Learns to Create

Ian Goodfellow introduces Generative Adversarial Networks, where two neural networks compete — one generates fake data, the other tries to detect it. This 'adversarial' approach revolutionized AI's ability to create realistic images, videos, and more.

Generative Adversarial Nets (2014)

2014Paper

GANs: AI Learns to Create

Generative Adversarial Nets (2014)

2014Paper

Attention Mechanism: Teaching AI to Focus

Bahdanau, Cho, and Bengio introduce the attention mechanism for machine translation, allowing neural networks to 'focus' on relevant parts of input when generating output. This concept would later become the foundation of the Transformer architecture.

Neural Machine Translation by Jointly Learning to Align and Translate (2015)

2014Paper

Attention Mechanism: Teaching AI to Focus

Neural Machine Translation by Jointly Learning to Align and Translate (2015)

2015Paper

ResNet: Going Deeper Than Ever

Kaiming He and team at Microsoft introduce Residual Networks with 'skip connections' that allow training of networks with 152+ layers. ResNet won the ImageNet challenge and showed that deeper networks could keep improving, a key insight for scaling AI.

Deep Residual Learning for Image Recognition (2016)

2015Paper

ResNet: Going Deeper Than Ever

Deep Residual Learning for Image Recognition (2016)

2016Breakthrough

AlphaGo Defeats World Go Champion

Google DeepMind's AlphaGo defeats Lee Sedol, one of the world's best Go players, 4-1 in a historic match. Go was considered far too complex for AI to master, with more possible positions than atoms in the universe. This victory stunned the world.

Mastering the game of Go with deep neural networks and tree search (2016)

2016Breakthrough

AlphaGo Defeats World Go Champion

Mastering the game of Go with deep neural networks and tree search (2016)

2017Paper

Attention Is All You Need: The Transformer

Google researchers introduce the Transformer architecture, replacing recurrent networks with pure attention mechanisms. This paper is arguably the most influential in modern AI — it's the foundation of GPT, BERT, and virtually every large language model today.

Attention Is All You Need (2017)

2017Paper

Attention Is All You Need: The Transformer

Attention Is All You Need (2017)

2018Paper

BERT: Understanding Language Both Ways

Google releases BERT (Bidirectional Encoder Representations from Transformers), which reads text in both directions to understand context better. BERT dramatically improved Google Search and set new records on virtually every NLP benchmark.

BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2019)

2018Paper

BERT: Understanding Language Both Ways

BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2019)

2019Product

GPT-2: Too Dangerous to Release?

OpenAI trains GPT-2, a 1.5 billion parameter language model that generates remarkably coherent text. Initially, OpenAI withheld the full model citing concerns about misuse — sparking the first major debate about AI safety and responsible release.

Language Models are Unsupervised Multitask Learners (2019)

2019Product

GPT-2: Too Dangerous to Release?

Language Models are Unsupervised Multitask Learners (2019)

2020s

2020Breakthrough

GPT-3: The 175-Billion Parameter Giant

OpenAI releases GPT-3 with 175 billion parameters — 100x larger than GPT-2. It can write essays, code, poetry, and even carry on conversations with minimal prompting. GPT-3 demonstrated that scaling up models dramatically improves capabilities.

Language Models are Few-Shot Learners (2020)

2020Breakthrough

GPT-3: The 175-Billion Parameter Giant

Language Models are Few-Shot Learners (2020)

2020Breakthrough

AlphaFold 2: Solving Biology's Grand Challenge

DeepMind's AlphaFold 2 solves the 50-year-old protein folding problem, predicting 3D protein structures with near-experimental accuracy. This breakthrough has massive implications for drug discovery, disease understanding, and biology as a whole.

Highly accurate protein structure prediction with AlphaFold (2021)

2020Breakthrough

AlphaFold 2: Solving Biology's Grand Challenge

Highly accurate protein structure prediction with AlphaFold (2021)

2021Product

DALL-E: AI Paints from Words

OpenAI introduces DALL-E, which generates images from text descriptions. Type 'an armchair in the shape of an avocado' and the AI creates it. This marks the beginning of the generative AI revolution in visual arts.

Zero-Shot Text-to-Image Generation (2021)

2021Product

DALL-E: AI Paints from Words

Zero-Shot Text-to-Image Generation (2021)

2021Product

GitHub Copilot: AI Writes Code

GitHub launches Copilot, an AI pair programmer powered by OpenAI Codex. It autocompletes code, writes functions from comments, and helps developers work faster. This is one of the first widely-adopted AI tools in professional software development.

Evaluating Large Language Models Trained on Code (2021)

2021Product

GitHub Copilot: AI Writes Code

Evaluating Large Language Models Trained on Code (2021)

2022Product

Stable Diffusion: Open-Source Image Generation

Stability AI releases Stable Diffusion as open source, democratizing AI image generation. Unlike DALL-E, anyone could run it on their own computer. This sparked an explosion of AI art tools, plugins, and communities worldwide.

High-Resolution Image Synthesis with Latent Diffusion Models (2022)

2022Product

Stable Diffusion: Open-Source Image Generation

High-Resolution Image Synthesis with Latent Diffusion Models (2022)

2022Product

ChatGPT: AI Goes Mainstream

OpenAI launches ChatGPT, a conversational AI that reaches 100 million users in just two months — the fastest-growing consumer application in history. It makes AI accessible to everyone and sparks a global conversation about AI's role in society.

Training language models to follow instructions with human feedback (2022)

2022Product

ChatGPT: AI Goes Mainstream

Training language models to follow instructions with human feedback (2022)

2023Product

LLaMA: Open-Source Large Language Models

Meta releases LLaMA, a family of open-source large language models. This democratizes access to powerful AI, allowing researchers and companies worldwide to build on top of state-of-the-art models without depending on proprietary APIs.

LLaMA: Open and Efficient Foundation Language Models (2023)

2023Product

LLaMA: Open-Source Large Language Models

LLaMA: Open and Efficient Foundation Language Models (2023)

2023Breakthrough

GPT-4: Multimodal Intelligence

OpenAI releases GPT-4, which can understand both text and images. It passes the bar exam, scores in the top percentiles on SATs, and demonstrates reasoning abilities far beyond its predecessors. The era of multimodal AI begins.

GPT-4 Technical Report (2023)

2023Breakthrough

GPT-4: Multimodal Intelligence

GPT-4 Technical Report (2023)

2023Product

Claude: AI with a Focus on Safety

Anthropic launches Claude, an AI assistant built with Constitutional AI — a method that trains models to be helpful, harmless, and honest. Anthropic, founded by former OpenAI researchers, represents the growing emphasis on AI safety in the industry.

Constitutional AI: Harmlessness from AI Feedback (2022)

2023Product

Claude: AI with a Focus on Safety

Constitutional AI: Harmlessness from AI Feedback (2022)

2023Product

Midjourney V5: Photorealistic AI Art

Midjourney V5 achieves near-photorealistic image generation, making it increasingly difficult to distinguish AI art from photographs. This sparks intense debates about copyright, authenticity, and the future of creative professions.

Midjourney V5 Release (2023)

2023Product

Midjourney V5: Photorealistic AI Art

Midjourney V5 Release (2023)

2023Product

Gemini: Google's Multimodal AI

Google DeepMind launches Gemini, a natively multimodal model that can understand and generate text, code, images, audio, and video. It represents Google's most ambitious AI effort and their answer to GPT-4.

Gemini: A Family of Highly Capable Multimodal Models (2023)

2023Product

Gemini: Google's Multimodal AI

Gemini: A Family of Highly Capable Multimodal Models (2023)

2024Product

Sora: AI Generates Realistic Video

OpenAI reveals Sora, a text-to-video model that generates stunningly realistic minute-long videos from text prompts. It represents a major leap in generative AI, raising both excitement and concern about AI-generated media.

Video generation models as world simulators (2024)

2024Product

Sora: AI Generates Realistic Video

Video generation models as world simulators (2024)

2024Product

Claude 3: Approaching Human-Level Understanding

Anthropic releases Claude 3 Opus, which matches or exceeds GPT-4 on many benchmarks. The model family (Haiku, Sonnet, Opus) offers different capability-cost tradeoffs, showing the industry moving toward tiered AI offerings.

The Claude 3 Model Family: Opus, Sonnet, Haiku (2024)

2024Product

Claude 3: Approaching Human-Level Understanding

The Claude 3 Model Family: Opus, Sonnet, Haiku (2024)

2024Breakthrough

OpenAI o1: AI That Thinks Step by Step

OpenAI releases o1, a model trained to 'think' through problems step by step before answering. It excels at math, coding, and scientific reasoning, demonstrating that teaching AI to reason explicitly is a powerful new paradigm beyond just scaling model size.

Learning to Reason with LLMs (2024)

2024Breakthrough

OpenAI o1: AI That Thinks Step by Step

Learning to Reason with LLMs (2024)

2024Milestone

The Scaling Debate: Are We Hitting a Wall?

As 2024 progresses, a major debate emerges: are scaling laws reaching diminishing returns? While models keep growing, some researchers argue that simply making models bigger isn't enough. The industry begins exploring new approaches like test-time compute and reasoning models.

Scaling Laws for Neural Language Models (2020)

2024Milestone

The Scaling Debate: Are We Hitting a Wall?

Scaling Laws for Neural Language Models (2020)

2025Breakthrough

DeepSeek R1: China's Open-Source Reasoning Model

Chinese AI lab DeepSeek releases R1, an open-source reasoning model that rivals top Western models at a fraction of the training cost. It demonstrates that frontier AI capabilities are becoming globally distributed and that cost-efficient training methods can close the gap.

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (2025)

2025Breakthrough

DeepSeek R1: China's Open-Source Reasoning Model

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (2025)

2025Milestone

The Age of AI Agents

2025 sees the rise of AI agents — systems that can autonomously browse the web, write and execute code, manage files, and complete complex multi-step tasks. From Claude's computer use to OpenAI's Operator, AI moves from answering questions to taking actions.

Anthropic Claude Computer Use (2024)

2025Milestone

The Age of AI Agents

Anthropic Claude Computer Use (2024)