A guide from LMTimeline

The History of AI: A Timeline from 2006 to Today

Artificial intelligence didn't arrive overnight. It compounded — one breakthrough enabling the next — from a quiet deep-learning revival in the mid-2000s to today's frontier models and the governments now weighing in on their release. Here is the short version, in order. For the living, day-by-day record, see the full AI timeline.

2006–2011: The deep-learning groundwork

Modern AI traces to a 2006 revival of neural networks, when Geoffrey Hinton and colleagues showed deep networks could be trained effectively layer by layer. The ideas weren't new, but two things were about to change everything: data and compute. In 2009, the ImageNet dataset gave researchers millions of labeled images to train on, and GPUs — originally built for video games — turned out to be ideal for the matrix math that neural networks rely on.

2012–2016: Deep learning breaks through

The dam broke in 2012, when AlexNet won the ImageNet competition by a stunning margin, proving deep convolutional networks could see. A cascade followed: word embeddings (word2vec, 2013) gave machines a sense of meaning, generative adversarial networks (GANs, 2014) let them create images, and sequence-to-sequence models reshaped translation. In 2016, DeepMind's AlphaGo beat a world champion at Go — a game long thought decades away from machine mastery.

2017–2019: The Transformer era begins

In 2017, a Google paper titled "Attention Is All You Need" introduced the Transformer — the architecture underneath nearly every major model since. It scaled better than anything before it. BERT (2018) showed Transformers could deeply understand language; OpenAI's GPT-2 (2019) showed they could generate it convincingly enough that its full release was initially withheld over misuse concerns — an early preview of the safety debates to come.

2020–2022: Scale and the generative boom

GPT-3 (2020) demonstrated that simply making models bigger unlocked surprising new abilities, and "scaling laws" became the field's organizing principle. Image generation went mainstream with DALL·E, Midjourney and the open-source Stable Diffusion. Then, in November 2022, ChatGPT put a conversational model in front of everyone — and reached a hundred million users faster than any consumer product in history, turning AI from a research story into a cultural one.

2023–2024: The arms race and open weights

What followed was a sprint. OpenAI's GPT-4, Anthropic's Claude and Google's Gemini pushed the frontier of reasoning and multimodality, while Meta's Llama and a wave of open-weight models from Mistral and others made capable AI freely downloadable. Context windows grew from thousands of tokens to millions; models learned to see, hear and code. AI moved from a chat box into search, office software, phones and developer tools.

2025–today: Agents, reasoning and governance

The newest era is defined by models that reason step by step and act as agents — using tools, writing and running code, and coordinating subagents on long tasks. It's also the era when governments started shaping releases directly: in mid-2026, the US administration asked OpenAI to stagger its GPT-5.6 rollout and temporarily restricted Anthropic's most powerful models to vetted partners over national-security concerns. AI policy and AI capability are now the same story.

Keep reading

• A Timeline of Large Language Models (LLMs) — from word2vec to today's frontier models.
• The live AI timeline — filter by importance, type or year, and subscribe for new entries.