Let's talk

Guide · History of AI

The evolution of AI: a timeline from 1950 to 2026

AI moved from logic puzzles to learning from data, then to generating language, and now to agents that act. Here is the dated timeline, the eras and the hardware that made each jump possible.

By Divyaprakash JhaUpdated 12 min read

The history of AI in one paragraph

The evolution of artificial intelligence runs in four big moves. From 1950 to the 1980s, researchers tried to write intelligence down as rules and logic. From the late 1980s to the 2000s, machine learning let computers learn patterns from data. From 2012, deep learning on GPUs made neural networks practical, and the Transformer in 2017 made them scale. From 2022, generative AI reached the public through ChatGPT, and by 2025 and 2026 the focus moved from chatbots to agents that plan, use tools and finish tasks.

This AI timeline lists the dated milestones that mattered, explains why each one mattered, and ends with the hardware story and India's place in it. If you want to understand how today's models work, read the LLM guide alongside it.

The eras of AI at a glance

EraYearsCore ideaTypical systems
Symbolic AI1950s–1970sIntelligence as logic and search over symbolsLogic Theorist, ELIZA, early game programs
First AI wintermid-1970sFunding cut after promises outran resultsResearch shrinks
Expert systems1980sEncode specialists' rules for narrow domainsXCON, medical and industrial rule engines
Second AI winterlate 1980s–early 1990sRule systems prove brittle and costlySpecialised AI hardware market collapses
Statistical machine learning1990s–2000sLearn patterns from dataSpam filters, search ranking, Deep Blue-era search
Deep learning2006–2017Many-layer neural networks on GPUsAlexNet, word2vec, GANs, AlphaGo
Transformers and foundation models2017–2022One large pretrained model, adapted to many tasksBERT, GPT-2, GPT-3, diffusion models
Generative AI goes public2022–2024Chat interfaces and multimodal modelsChatGPT, GPT-4, Llama, Claude, Gemini
Reasoning and agents2024–2026Models that think longer and act through toolso1, DeepSeek-R1, MCP, coding and computer-use agents

1950–1990: rules, logic and two AI winters

  1. Turing asks whether machines can think In "Computing Machinery and Intelligence", Alan Turing proposes the imitation game, later called the Turing test.
  2. The Dartmouth workshop names the field John McCarthy, Marvin Minsky, Claude Shannon and Nathaniel Rochester's 1955 proposal coins "artificial intelligence"; the summer workshop in 1956 is treated as AI's founding.
  3. Rosenblatt's perceptron Frank Rosenblatt describes the perceptron, a simple learning neuron and the ancestor of today's neural networks.
  4. ELIZA Joseph Weizenbaum at MIT publishes ELIZA, a pattern-matching chatbot whose "therapist" script surprised users with how human it felt.
  5. The Lighthill report A critical review for the UK government leads to deep cuts in British AI funding, part of the first AI winter.
  6. Expert systems go commercial Systems such as XCON, used to configure computer orders, show that narrow rule-based AI can save real money.
  7. Backpropagation popularised Rumelhart, Hinton and Williams publish in Nature how backpropagation trains multi-layer networks, reviving neural network research.
  8. The second AI winter begins The market for specialised AI machines collapses, and expert systems prove hard to maintain.

The lesson of this era: hand-written rules work in narrow, stable domains but break when the world is messy. Every later era tried to learn the rules from data instead.

1990–2016: machine learning and the deep learning revival

  1. Deep Blue beats Kasparov IBM's chess computer defeats world champion Garry Kasparov in a six-game match, mostly through massive search rather than learning.
  2. Deep belief networks Geoffrey Hinton and colleagues show a practical way to train deep networks layer by layer, starting the "deep learning" label.
  3. ImageNet Fei-Fei Li's team presents ImageNet, a dataset of millions of labelled images that becomes the benchmark for computer vision.
  4. AlexNet Krizhevsky, Sutskever and Hinton win the ImageNet challenge by a wide margin with a deep convolutional network trained on two GPUs. Deep learning becomes the default.
  5. word2vec Tomas Mikolov and colleagues at Google show that words can be represented as vectors whose distances capture meaning.
  6. Generative adversarial networks Ian Goodfellow and colleagues introduce GANs, two networks competing to generate realistic images.
  7. AlphaGo beats Lee Sedol DeepMind's AlphaGo wins 4–1 against one of the world's best Go players, combining deep learning with search and self-play.

Three things came together in this period: large datasets, better training methods and cheap parallel compute from gaming GPUs. Remove any one and the 2012 breakthrough does not happen.

2017–2022: the Transformer and foundation models

  1. "Attention Is All You Need" Google researchers introduce the Transformer, which processes whole sequences in parallel using attention. It becomes the base of almost every modern language model (paper).
  2. GPT-1 and BERT OpenAI's GPT shows the value of generative pre-training on unlabelled text; Google's BERT sets new records on language understanding tasks.
  3. GPT-2 OpenAI's 1.5B-parameter model writes coherent paragraphs and is released in stages over concerns about misuse.
  4. GPT-3 At 175B parameters, GPT-3 shows "few-shot" learning: give a few examples in the prompt and it performs new tasks without retraining.
  5. DALL-E and CLIP OpenAI links text and images, generating pictures from written descriptions.
  6. Diffusion image models go mainstream DALL-E 2, Midjourney and the openly released Stable Diffusion make high-quality image generation available to millions.

The key idea of this era is the foundation model: train one very large model on broad data, then adapt it with prompts or light fine-tuning. Scale became a strategy, and that pushed demand for GPUs sharply upward.

2022–2024: generative AI reaches everyone

  1. ChatGPT On 30 November, OpenAI releases ChatGPT as a free research preview. A chat interface plus training from human feedback turns a language model into a mass-market product.
  2. Llama and the open-weight wave In February Meta releases LLaMA to researchers, followed by Llama 2 in July with a licence allowing commercial use. Open-weight models from many labs follow.
  3. GPT-4 Released on 14 March, GPT-4 accepts images as well as text and performs at a much higher level on professional and academic exams.
  4. Assistants multiply Anthropic's Claude and Google's Bard, later Gemini, give businesses a choice of frontier models.
  5. Natively multimodal models Models such as GPT-4o and Gemini handle text, images and audio in one system, enabling real-time voice conversation.
  6. Reasoning at test time In September OpenAI releases o1, trained to think through problems step by step before answering. Spending more compute at answer time improves maths, science and coding results.
  7. Computer use and the Model Context Protocol In October Anthropic shows Claude operating a computer screen; in November it releases MCP, an open standard for connecting models to tools and data.

2025–2026: from chat to agents

  1. DeepSeek-R1 Released on 20 January with open weights, DeepSeek-R1 shows reasoning ability comparable to leading closed models at much lower reported cost, and resets expectations about who can build frontier AI.
  2. Agent2Agent protocol In April Google announces A2A, an open protocol for agents built by different vendors to discover each other and work together. It later moves to the Linux Foundation.
  3. Coding agents Agentic coding tools that read a codebase, run commands and make changes become a daily tool for software teams.
  4. A shared home for agent standards In December the Linux Foundation launches the Agentic AI Foundation, with Anthropic's MCP among its founding projects.
  5. India hosts the AI Impact Summit In February, New Delhi hosts the AI Impact Summit, which closes with the New Delhi Declaration endorsed by dozens of countries and organisations.
  6. Capable open MoE models In April Google releases Gemma 4, including a mixture-of-experts model with about 26B total and 4B active parameters, small enough to run on a single workstation GPU.

The broad trend of 2025 and 2026 is clear even without naming every launch. Models are measured less by how well they chat and more by whether they can complete a multi-step job: research a question, fix a bug, process a batch of documents, or run a business workflow with a human approving the important steps. Protocols such as MCP and A2A are the plumbing that lets those agents reach tools and each other. The agentic architecture guide explains how such systems are designed.

Generative AI history: the quieter ideas behind ChatGPT

The famous launches sit on top of a few less visible ideas. Knowing them explains why progress felt so sudden in 2022.

Scaling laws (2020)

In January 2020 OpenAI researchers published "Scaling Laws for Neural Language Models", showing that a model's error falls smoothly and predictably as you add parameters, data and compute. This turned model building into something closer to engineering: if you could pay for more compute, you could forecast a better model. In 2022 DeepMind's Chinchilla work refined the recipe, showing that many large models had been trained on too little data for their size.

Learning from human feedback (2022)

A model trained only to predict the next word is not a helpful assistant. It may ramble, refuse oddly or follow harmful requests. OpenAI's InstructGPT paper in 2022 described fine-tuning with human ratings, known as reinforcement learning from human feedback (RLHF). People compare answers, a reward model learns their preferences, and the language model is tuned toward answers people prefer. ChatGPT applied this approach to a conversational format.

Retrieval and tools

Models trained once cannot know today's prices or your company's files. Retrieval-augmented generation, which fetches relevant documents and puts them in the prompt, and function calling, which lets a model request an action from software, became standard between 2020 and 2023. They are the direct ancestors of today's agents and of protocols such as MCP.

Reinforcement learning for reasoning

Reasoning models extend the feedback idea: instead of human ratings alone, they are trained on problems with checkable answers, such as maths and code, and rewarded for reaching correct results. The model learns to write out longer chains of thought and to correct itself. DeepSeek published details of this approach with R1, which helped many other labs reproduce it.

How the hardware changed: CPU to GPU to AI accelerators

Each era of AI was shaped by the compute available to it.

  • CPUs and special machines (1950s–1990s). Early AI ran on general-purpose mainframes and, in the 1980s, on dedicated Lisp machines. When cheaper desktop computers caught up, the Lisp machine market collapsed.
  • GPUs (late 2000s onward). Graphics cards, built to shade millions of pixels in parallel, turned out to be ideal for neural network maths. NVIDIA's CUDA made them programmable for general work, and AlexNet's two-GPU training in 2012 proved the point.
  • TPUs and AI accelerators (2016 onward). Google announced its Tensor Processing Unit in 2016, a chip designed only for neural network workloads. Others followed with their own accelerators, and GPUs gained tensor cores dedicated to matrix maths.
  • Memory becomes the bottleneck (2020s). For large language models, memory capacity and bandwidth matter as much as raw compute. High-bandwidth memory, large VRAM and unified memory now define AI hardware.

For what this means when you choose a machine today, see GPU, VRAM, RAM and CPU for AI.

India's AI story

India's role has grown from a large pool of engineering talent into a deliberate national programme.

  • IndiaAI Mission (2024). On 7 March 2024 the Union Cabinet approved the IndiaAI Mission with an outlay of ₹10,371.92 crore over five years, covering shared GPU compute, datasets, homegrown foundation models, skills and startup funding (PIB release).
  • Indic language models. Under the mission and alongside it, Indian teams are building models trained for Indian languages, and public projects such as Bhashini focus on translation and speech across Indian languages. Language coverage matters because a large share of Indian users are more comfortable in their own language than in English.
  • AI Impact Summit (2026). In February 2026 India hosted the AI Impact Summit in New Delhi, ending with the New Delhi Declaration.

The practical opportunity is not only national models. It is millions of small businesses using AI in Hindi, Marathi and other languages for sales, service and accounts. That is where much of India's real AI adoption is likely to happen.

What 75 years of AI history teach

  • Hype cycles are normal. Two AI winters followed periods of overpromising. Judge systems by what they reliably do today.
  • Data and compute beat clever rules. Methods that scale with more data and more compute have won again and again.
  • Old ideas return. Neural networks date to 1958; they waited decades for hardware and data.
  • Interfaces drive adoption. GPT-3 existed in 2020, but ChatGPT's simple chat box in 2022 is what reached the public. Agents will spread the same way: when they are easy to trust and use.

Where this goes next is the subject of the future of AI: 2027 to 2036.

Questions people ask

When did artificial intelligence start?

The idea is usually dated to Alan Turing's 1950 paper on whether machines can think. The term artificial intelligence was coined for the Dartmouth summer workshop proposed in 1955 and held in 1956, which is treated as the founding of the field.

What were the AI winters?

Periods when funding and interest in AI fell sharply after promises outran results. The first came in the mid-1970s, and the second in the late 1980s and early 1990s after the expert-systems boom faded.

What started the deep learning boom?

Better training methods from around 2006, large labelled datasets such as ImageNet, and GPUs. The turning point was AlexNet winning the ImageNet challenge in 2012 by a wide margin.

What is the Transformer and why does it matter?

The Transformer is a neural network design introduced in the 2017 paper Attention Is All You Need. It processes text in parallel using attention, scaled far better than earlier designs, and underpins GPT, BERT, Llama, Claude and Gemini.

When was ChatGPT released?

OpenAI released ChatGPT as a free research preview on 30 November 2022. GPT-4 followed on 14 March 2023.

What are reasoning models?

Models trained to spend extra computation thinking step by step before answering. OpenAI's o1 in September 2024 and DeepSeek-R1 in January 2025 made this approach mainstream.

What is India doing in AI?

The Union Cabinet approved the IndiaAI Mission in March 2024 with an outlay of about 10,372 crore rupees for compute, datasets, models and skills. India also hosted the AI Impact Summit in New Delhi in February 2026.