AI timeline

Meta Muse Wonder and Muse Wonder Full

Meta released Muse Wonder, its first frontier-class model, via its API and Muse Code, with Wonder Full for subscribers.

Source

Anthropic Claude Haiku 5.5

Anthropic released Haiku 5.5, a small low-cost model for high-volume tasks, and halved Sonnet 5.5's cache-read price.

Source

OpenAI GPT-6 Instant and Intelligent UI for everyone

OpenAI brought GPT-6 to all ChatGPT tiers with interactive 'Intelligent UI' answers, paid users first and Free and Go users a day later.

Source

OpenAI DevDay 2026 and GPT-6.1 Sol

At DevDay in San Francisco OpenAI launched GPT-6.1 Sol, near Astra-level at a fifth of the price, and always-on agents called Dots; GPT-6.1 Astra was held back.

Source

Anthropic Claude Sonnet 5.5

Anthropic released Sonnet 5.5, a faster lower-cost companion to Opus 5.5.

Source

OpenAI GPT-6 Astra

OpenAI unveiled GPT-6 Astra, a model strong at computer use, for approved testers, with a restricted version for paid ChatGPT users the next day.

Source

Anthropic Claude Opus 5

Anthropic released Opus 5, close to Fable 5 capability at about half the price.

Source

OpenAI GPT-5.6 (Luna, Terra, Sol) and ChatGPT Work

OpenAI publicly released the GPT-5.6 family in three tiers and launched ChatGPT Work, an agent combining ChatGPT and Codex for office tasks.

Source

OpenAI Atlas browser shutdown announced

OpenAI said ChatGPT Atlas would close on 9 August 2026, with its features moving into the ChatGPT desktop app and a Chrome extension.

Source

OpenAI GPT-5.6 limited preview

OpenAI gave GPT-5.6 to a limited group and delayed public rollout after a US government request for early access.

Source

Anthropic Claude Fable 5 and Mythos 5

Anthropic released Fable 5, a public Mythos-class model with safeguards, and a limited-access Mythos 5.

Source

Anthropic Claude Opus 4.8

Anthropic released Claude Opus 4.8.

Source

Google Gemini 3.5 Flash

Google launched Gemini 3.5 Flash, positioned around agents.

Source

OpenAI GPT-5.5 with trusted access for cyber

OpenAI announced a GPT-5.5 variant for vetted cybersecurity defenders.

Source

OpenAI GPT-5.5 Instant becomes default

GPT-5.5 Instant replaced GPT-5.3 Instant as ChatGPT's default model.

Source

OpenAI GPT-5.5 (codename Spud)

OpenAI released GPT-5.5 Thinking and Pro, with API access the next day.

Source

Anthropic Claude Design preview

Anthropic previewed Claude Design, which turns prompts into visual prototypes, for paid users.

Source

Anthropic Claude Opus 4.7

Anthropic released Claude Opus 4.7.

Source

Meta Muse Spark

Meta Superintelligence Labs released Muse Spark, its first model, to replace Llama in Meta's assistant.

Source

Anthropic Claude Mythos Preview (Project Glasswing)

Anthropic gave a limited Mythos Preview model to about a dozen organizations for defensive cybersecurity work.

Source

OpenAI OpenAI announces end of Sora app and API

OpenAI said it would discontinue the Sora consumer app and API to free compute for other products.

Source

OpenAI GPT-5.4 mini and nano

OpenAI released GPT-5.4 mini, available to free users, and GPT-5.4 nano for the API only.

Source

OpenAI GPT-5.4 Thinking and Pro

OpenAI released GPT-5.4 in Thinking and Pro versions for paid ChatGPT users and the API.

Source

OpenAI GPT-5.3 Instant

OpenAI made GPT-5.3 Instant the ChatGPT default, aiming for fewer needless refusals and better web answers.

Source

Google Gemini 3.1 Pro

Google released Gemini 3.1 Pro.

Source

Anthropic Claude Sonnet 4.6

Anthropic released Sonnet 4.6, a faster and cheaper model.

Source

Alibaba (Qwen) Qwen3.5

Alibaba released Qwen3.5.

Source

OpenAI GPT-4o retired from ChatGPT

GPT-4o and several older models were removed from ChatGPT; business plans kept GPT-4o in custom GPTs until 3 April 2026.

Source

Anthropic Claude Opus 4.6

Anthropic released Opus 4.6 with agent teams and Claude in PowerPoint.

Source

OpenAI GPT-5.3-Codex

OpenAI released GPT-5.3-Codex, a faster coding model pitched for broader computer work beyond code.

Source

OpenAI Codex app for macOS

OpenAI released a desktop Codex app for managing several coding agents in parallel, with automations; Windows followed on 4 March.

Source

OpenAI Retirement of GPT-4o in ChatGPT announced

OpenAI said GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini would leave ChatGPT on 13 February 2026, with the API unchanged.

Source

OpenAI ChatGPT Go worldwide; ad tests announced

OpenAI made the $8 ChatGPT Go plan available everywhere and said it would test labeled ads for US Free and Go users.

Source

Anthropic Claude Cowork research preview

Anthropic previewed Cowork, an agent in the Claude desktop app that works on local files for non-programmers, starting with Max subscribers.

Source

Apple Apple and Google Gemini partnership

Apple and Google said the next Apple Foundation Models would be built on Gemini models and cloud technology.

Source

OpenAI GPT-5.2-Codex

OpenAI released a GPT-5.2 version optimized for agentic coding in Codex, with API access added in January 2026.

Source

Google Gemini 3 Flash

Google released Gemini 3 Flash, a faster lower-cost Gemini 3 model.

Source

OpenAI New ChatGPT Images (GPT Image 1.5)

OpenAI rolled out a new image model to all ChatGPT users and the API, with more precise edits and faster generation.

Source

OpenAI GPT-5.2

OpenAI released GPT-5.2 in Instant, Thinking and Pro versions, focused on professional work and long tasks.

Source

Mistral Mistral 3 and Mistral Large 3

Mistral released the Mistral 3 family, including Mistral Large 3 and small Ministral 3 models, under open licences.

Source

DeepSeek DeepSeek-V3.2

DeepSeek released V3.2 with sparse attention for cheaper long-context use.

Source

Anthropic Claude Opus 4.5

Anthropic released Opus 4.5 with gains in coding and office work at lower prices.

Source

OpenAI Shopping research in ChatGPT

ChatGPT gained a shopping research mode that builds buyer's guides using a model based on GPT-5 thinking mini.

Source

Google Nano Banana Pro (Gemini 3 Pro Image)

Google released an image model built on Gemini 3 Pro with better text rendering.

Source

OpenAI GPT-5.1-Codex-Max

OpenAI released a coding model for long-running work that compacts its own context to keep going across sessions.

Source

Google Gemini 3 Pro

Google released Gemini 3 Pro in the Gemini app, Search and developer tools.

Source

OpenAI Group chats pilot

OpenAI piloted group chats where up to 20 people can share one ChatGPT conversation, starting in a few countries.

Source

OpenAI GPT-5.1 Instant and Thinking

OpenAI updated ChatGPT to GPT-5.1, aiming for warmer conversation, adaptive reasoning and new personality presets.

Source

OpenAI ChatGPT Atlas browser

OpenAI launched ChatGPT Atlas, a macOS web browser with ChatGPT built in and an agent mode for paid users.

Source

Anthropic Claude Haiku 4.5

Anthropic released Haiku 4.5, a small fast model.

Source

OpenAI OpenAI DevDay 2025

OpenAI announced apps inside ChatGPT with the Apps SDK, AgentKit, general availability of Codex, and Sora 2 and GPT-5 Pro in the API.

Source

OpenAI Apps in ChatGPT and the Apps SDK

OpenAI let third-party apps run inside ChatGPT conversations, built on the Model Context Protocol.

Source

OpenAI AgentKit and Agent Builder

OpenAI released AgentKit, including the visual Agent Builder, a connector registry and ChatKit for embedding agents.

Source

OpenAI Sora 2 and the Sora app

OpenAI released Sora 2, a video model with synchronized audio, alongside an invite-only social iOS app with cameos.

Source

Anthropic Claude Sonnet 4.5

Anthropic released Sonnet 4.5, focused on coding and long-running agents, with the Claude Agent SDK.

Source

OpenAI GPT-5-Codex

OpenAI released GPT-5-Codex, a GPT-5 version tuned for agentic coding, as the default for Codex cloud tasks and code review.

Source

OpenAI gpt-realtime and Realtime API GA

OpenAI made the Realtime API generally available and released gpt-realtime, a speech-to-speech model with SIP calling and MCP support.

Source

Google Gemini 2.5 Flash Image (Nano Banana)

Google released an image editing and generation model, nicknamed Nano Banana, in the Gemini app.

Source

DeepSeek DeepSeek-V3.1

DeepSeek released V3.1 with hybrid thinking and non-thinking modes.

Source

OpenAI GPT-5 released

OpenAI launched GPT-5, a unified system that routes between fast and thinking models, to all ChatGPT users including free.

Source

Anthropic Claude Opus 4.1

Anthropic released Opus 4.1 with improved coding and agentic performance.

Source

OpenAI gpt-oss open-weight models

OpenAI released gpt-oss-120b and gpt-oss-20b, its first open-weight language models since GPT-2, under Apache 2.0.

Source

OpenAI ChatGPT agent

OpenAI merged Operator and deep research into ChatGPT agent, which uses its own virtual computer to complete tasks.

Source

xAI Grok 4 and Grok 4 Heavy

xAI released Grok 4 and a multi-agent Grok 4 Heavy tier.

Source

Midjourney Midjourney V1 video model

Midjourney released its first video model, animating images into short clips.

Source

Mistral Magistral reasoning models

Mistral released Magistral, its first reasoning models.

Source

Anthropic Claude Opus 4 and Sonnet 4

Anthropic released Claude Opus 4 and Sonnet 4 and made Claude Code generally available.

Source

OpenAI Codex cloud coding agent

OpenAI previewed Codex, a cloud software engineering agent powered by codex-1, a version of o3, for Pro, Team and Enterprise users.

Source

Mistral Mistral Medium 3

Mistral released Mistral Medium 3, a mid-size model aimed at enterprises.

Source

OpenAI GPT-4o update rolled back for sycophancy

OpenAI reverted a GPT-4o update that made ChatGPT overly flattering and explained what went wrong.

Source

Alibaba (Qwen) Qwen3

Alibaba released the Qwen3 family of open-weight models with hybrid thinking.

Source

OpenAI OpenAI o3 and o4-mini

OpenAI released o3 and o4-mini, reasoning models that can use all ChatGPT tools and reason with images.

Source

OpenAI Codex CLI

OpenAI released Codex CLI, an open-source coding agent that runs in the terminal.

Source

OpenAI GPT-4.1 prompting guide

OpenAI published a cookbook guide explaining how to prompt GPT-4.1, which follows instructions more literally.

Source

OpenAI GPT-4.1, 4.1 mini and 4.1 nano in the API

OpenAI released three API models with up to one million tokens of context and better coding and instruction following.

Source

Meta Llama 4 Scout and Maverick

Meta released Llama 4 Scout and Maverick, mixture-of-experts multimodal models, and previewed Behemoth.

Source

Midjourney Midjourney V7

Midjourney released V7 in alpha.

Source

Google Gemini 2.5 Pro

Google released Gemini 2.5 Pro, a thinking model that topped several leaderboards.

Source

OpenAI Native image generation in GPT-4o

ChatGPT replaced DALL-E with GPT-4o's built-in image generation, which handles text in images and edits well.

Source

OpenAI Responses API and Agents SDK

OpenAI launched the Responses API, built-in web search, file search and computer use tools, and the open-source Agents SDK.

Source

OpenAI GPT-4.5 research preview

OpenAI released GPT-4.5, its largest non-reasoning chat model, to Pro users and developers.

Source

Anthropic Claude 3.7 Sonnet and Claude Code preview

Anthropic released Claude 3.7 Sonnet with optional extended thinking and previewed Claude Code, a terminal coding agent.

Source

xAI Grok 3

xAI released Grok 3 with a Think reasoning mode.

Source

OpenAI Deep research in ChatGPT

OpenAI launched deep research, an agent powered by a version of o3 that compiles cited reports from many web sources, for Pro users first.

Source

OpenAI OpenAI o3-mini

OpenAI released o3-mini, a low-cost reasoning model in ChatGPT (including free users) and the API.

Source

Alibaba (Qwen) Qwen2.5-Max

Alibaba released Qwen2.5-Max, a large mixture-of-experts model.

Source

DeepSeek DeepSeek app tops charts; Nvidia shares fall

DeepSeek's app reached the top of the US App Store and AI chip stocks dropped sharply.

Source

OpenAI Operator research preview

OpenAI released Operator, an agent that uses its own browser to complete web tasks, for US Pro users.

Source

OpenAI Stargate Project announced

OpenAI, SoftBank, Oracle and MGX announced a joint venture to build AI data centers in the US.

Source

DeepSeek DeepSeek-R1

DeepSeek released R1, an open-weight reasoning model rivaling o1, under the MIT licence.

Source

DeepSeek DeepSeek-V3

DeepSeek released V3, a 671B mixture-of-experts open model reportedly trained at low cost.

Source

OpenAI o3 and o3-mini previewed

On the last day of its '12 days' event OpenAI previewed o3 and o3-mini for safety testing, not public release.

Source

OpenAI o1 in the API

OpenAI brought o1 to the API with function calling and structured outputs, and updated the Realtime API.

Source

Google Gemini 2.0 Flash experimental

Google introduced Gemini 2.0, starting with an experimental Flash model aimed at agentic uses.

Source

OpenAI Sora released to the public

OpenAI launched Sora Turbo video generation for ChatGPT Plus and Pro users at sora.com.

Source

OpenAI Full o1 and ChatGPT Pro

OpenAI released the full o1 model and a $200 monthly Pro plan with o1 pro mode.

Source

Anthropic Model Context Protocol (MCP)

Anthropic open-sourced MCP, a standard for connecting AI assistants to data sources and tools.

Source

OpenAI ChatGPT search

OpenAI built web search with source links into ChatGPT, starting with paid users.

Source

Apple First Apple Intelligence features ship

Apple released iOS 18.1 with the first Apple Intelligence features in US English.

Source

Anthropic Computer use beta, new 3.5 Sonnet and 3.5 Haiku

Anthropic let Claude control a computer through screenshots, mouse and keyboard in public beta, alongside upgraded models.

Source

OpenAI Canvas in ChatGPT

OpenAI introduced canvas, a side-by-side workspace for writing and coding with ChatGPT.

Source

OpenAI DevDay 2024: Realtime API

OpenAI announced the Realtime API for speech-to-speech apps, plus prompt caching, vision fine-tuning and model distillation.

Source

Meta Llama 3.2 vision and edge models

Meta released Llama 3.2 with vision models and small models for phones.

Source

OpenAI OpenAI o1-preview and o1-mini

OpenAI released o1-preview and o1-mini, reasoning models that think before answering.

Source

OpenAI Structured Outputs in the API

OpenAI added Structured Outputs so model responses reliably match a developer-supplied JSON schema.

Source

Midjourney Midjourney V6.1

Midjourney released V6.1, available on Discord and on its website.

Source

OpenAI SearchGPT prototype

OpenAI tested SearchGPT, a search prototype with sourced answers, with a small group of users.

Source

Meta Llama 3.1 405B

Meta released Llama 3.1, including a 405-billion-parameter open-weight model.

Source

OpenAI GPT-4o mini

OpenAI released GPT-4o mini, a small low-cost model that replaced GPT-3.5 Turbo in ChatGPT.

Source

Anthropic Claude 3.5 Sonnet and Artifacts

Anthropic released Claude 3.5 Sonnet and Artifacts, a side panel for generated code and documents.

Source

Apple Apple Intelligence announced

At WWDC Apple announced Apple Intelligence, on-device and private-cloud AI features, plus optional ChatGPT integration.

Source

Microsoft Copilot+ PCs

Microsoft announced Copilot+ PCs with on-device AI hardware.

Source

Google Google I/O 2024: Veo, Imagen 3, AI Overviews

Google announced the Veo video model, Imagen 3, Gemini 1.5 Flash and AI Overviews in US Search.

Source

OpenAI GPT-4o released

OpenAI introduced GPT-4o, a natively multimodal model handling text, audio and images, and made it available to free users.

Source

OpenAI Model Spec first published

OpenAI published a draft Model Spec describing how its models should behave and resolve conflicts.

Source

Meta Llama 3 and Meta AI assistant

Meta released Llama 3 in 8B and 70B sizes and upgraded its Meta AI assistant.

Source

OpenAI ChatGPT without an account

OpenAI began letting people use ChatGPT without signing up.

Source

xAI Grok-1 weights released

xAI released the 314B-parameter Grok-1 base model weights under Apache 2.0.

Source

Anthropic Claude 3 family

Anthropic released Claude 3 Opus, Sonnet and Haiku, adding image input.

Source

Mistral Mistral Large and Le Chat

Mistral launched Mistral Large and its Le Chat assistant, plus a Microsoft partnership.

Source

Google Gemma open models

Google released Gemma, lightweight open-weight models built from Gemini research.

Source

Google Gemini 1.5 Pro

Google announced Gemini 1.5 Pro with a context window of up to one million tokens.

Source

OpenAI Sora text-to-video preview

OpenAI previewed Sora, a model that makes up to minute-long videos from text, to red teamers and selected artists.

Source

OpenAI ChatGPT memory testing begins

OpenAI started testing memory, letting ChatGPT remember details across conversations, with user controls.

Source

Google Bard renamed Gemini; Gemini Advanced

Google rebranded Bard as Gemini and launched Gemini Advanced with Ultra 1.0 in a paid plan.

Source

OpenAI GPT Store and ChatGPT Team

OpenAI opened the GPT Store for sharing custom GPTs and launched the ChatGPT Team plan for small groups.

Source

Midjourney Midjourney V6

Midjourney released V6 in alpha with better prompt following and text.

Source

Mistral Mixtral 8x7B

Mistral released Mixtral 8x7B, an open sparse mixture-of-experts model, after posting the weights by torrent.

Source

Google Gemini 1.0 announced

Google DeepMind introduced Gemini in Ultra, Pro and Nano sizes; Pro began powering Bard the same day.

Source

OpenAI Sam Altman agreed to return as CEO

OpenAI reached an agreement for Altman to return as CEO with a new initial board.

Source

Anthropic Claude 2.1

Anthropic released Claude 2.1 with a 200,000-token context window and beta tool use.

Source

OpenAI Board removes Sam Altman as CEO

OpenAI's board announced Sam Altman's departure as CEO; he returned days later after staff and investor pressure.

Source

OpenAI GPTs: custom versions of ChatGPT

OpenAI let users build custom ChatGPT versions with instructions, knowledge files and actions, without code.

Source

OpenAI GPT-4 Turbo preview

OpenAI previewed GPT-4 Turbo with a 128,000-token context window and cheaper pricing.

Source

OpenAI OpenAI DevDay 2023

At its first developer conference OpenAI announced GPT-4 Turbo with 128K context, the Assistants API, GPTs and lower prices.

Source

OpenAI Assistants API beta

OpenAI released a beta API for building assistants with threads, retrieval and code interpreter tools.

Source

xAI Grok announced

xAI unveiled Grok, a chatbot for X Premium+ subscribers.

Source

Mistral Mistral 7B

Mistral AI released its first model, Mistral 7B, under Apache 2.0.

Source

OpenAI ChatGPT can see, hear and speak

OpenAI announced voice conversations and image input (GPT-4V) for ChatGPT Plus and Enterprise.

Source

OpenAI DALL-E 3 announced

OpenAI announced DALL-E 3, built to follow prompts closely and to be used through ChatGPT.

Source

OpenAI ChatGPT Enterprise

OpenAI launched a business plan with enterprise security, unlimited GPT-4 and admin tools.

Source

Meta Code Llama

Meta released Code Llama, Llama 2 models specialized for programming.

Source

OpenAI Fine-tuning for GPT-3.5 Turbo

Developers could fine-tune GPT-3.5 Turbo on their own data.

Source

Alibaba (Qwen) Qwen-7B open weights

Alibaba released Qwen-7B as open weights.

Source

OpenAI Custom instructions in ChatGPT

ChatGPT gained custom instructions so users can set preferences that apply to every conversation.

Source

Meta Llama 2

Meta released Llama 2 with a licence allowing commercial use, in partnership with Microsoft.

Source

Anthropic Claude 2 and claude.ai

Anthropic released Claude 2 and opened the claude.ai chat site to the public in the US and UK.

Source

Open source Ollama first release

Ollama released a tool for downloading and running open models locally.

Source

OpenAI GPT-4 API generally available; Code Interpreter rolls out

OpenAI opened GPT-4 to all paying API users and began rolling out Code Interpreter to ChatGPT Plus.

Source

OpenAI Function calling in the API

OpenAI added function calling and new GPT-4 and GPT-3.5-turbo versions with lower prices.

Source

OpenAI ChatGPT iOS app

OpenAI released an official ChatGPT app for iPhone in the US.

Source

OpenAI Web browsing and plugins beta for Plus

OpenAI began rolling out web browsing and plugins to all ChatGPT Plus subscribers as beta features.

Source

Anthropic 100K context window

Anthropic expanded Claude's context window to 100,000 tokens.

Source

Google PaLM 2 at Google I/O

Google announced PaLM 2, powering Bard and many Google products, and opened Bard without a waitlist.

Source

OpenAI Option to turn off chat history

ChatGPT users gained a setting to stop conversations being used for training and kept in history.

Source

Alibaba (Qwen) Alibaba unveils Tongyi Qianwen

Alibaba Cloud announced its Tongyi Qianwen (Qwen) chatbot model for business testing.

Source

Open source AutoGPT released

Toran Bruce Richards released AutoGPT, an open-source agent that chains GPT-4 calls toward goals.

Source

OpenAI ChatGPT plugins announced

OpenAI introduced plugins, including web browsing and a code interpreter, to a limited set of users.

Source

Google Bard opens to US and UK users

Google began giving users in the US and UK access to Bard through a waitlist.

Source

Microsoft Microsoft 365 Copilot announced

Microsoft announced Copilot for Word, Excel, PowerPoint, Outlook and Teams.

Source

Midjourney Midjourney V5

Midjourney released V5 in alpha with more photorealistic output.

Source

OpenAI GPT-4 Technical Report on arXiv

OpenAI posted the GPT-4 technical report, giving benchmark results but few architecture details.

Source

Anthropic Claude launched

Anthropic released Claude and Claude Instant to businesses through its API and partners.

Source

OpenAI GPT-4 released

OpenAI announced GPT-4, a large multimodal model, available in ChatGPT Plus and via an API waitlist.

Source

Open source llama.cpp released

Georgi Gerganov released llama.cpp, letting LLaMA models run on ordinary laptops in C/C++.

Source

OpenAI ChatGPT (gpt-3.5-turbo) and Whisper APIs

OpenAI opened the gpt-3.5-turbo chat model and Whisper to developers at much lower prices than earlier models.

Source

Meta LLaMA released to researchers

Meta released LLaMA, a family of 7B to 65B models, under a research-only licence; the weights soon leaked.

Source

Microsoft New Bing with chat

Microsoft unveiled an AI-powered Bing and Edge with a chat assistant built on an OpenAI model.

Source

Google Bard announced

Google announced Bard, an experimental conversational AI service based on LaMDA, for trusted testers.

Source

Quora (Poe) Poe opens to the public

Quora opened Poe, an app for chatting with several AI bots, to everyone on iOS.

Source

OpenAI ChatGPT Plus subscription

OpenAI introduced a $20 monthly plan with priority access and faster responses.

Source

OpenAI Microsoft extends OpenAI partnership

Microsoft announced a multiyear, multibillion-dollar extension of its investment in OpenAI.

Source

Anthropic Constitutional AI paper

Anthropic described training a harmless assistant using a written set of principles and AI feedback.

Source

OpenAI ChatGPT launched as a research preview

OpenAI released ChatGPT, a free conversational interface built on a GPT-3.5 model fine-tuned for dialogue.

Source

Midjourney Midjourney V4 alpha

Midjourney released its V4 model in alpha.

Source

OpenAI DALL-E 2 waitlist removed

OpenAI opened DALL-E 2 to everyone without a waitlist.

Source

OpenAI Whisper speech recognition open-sourced

OpenAI released Whisper, an open-source multilingual speech recognition model.

Source

Character.AI Character.AI public beta

Character.AI opened a public beta for chatting with user-created characters.

Source

Stability AI Stable Diffusion public release

Stability AI and partners released Stable Diffusion with open weights.

Source

OpenAI DALL-E 2 enters beta

OpenAI began inviting a million waitlisted users to DALL-E 2 with a credit-based pricing model.

Source

Midjourney Midjourney open beta

Midjourney opened its Discord-based image generator to the public.

Source

Google Imagen text-to-image research

Google Research presented Imagen, a text-to-image diffusion model, without public release.

Source

OpenAI DALL-E 2 announced

OpenAI introduced DALL-E 2, producing more realistic, higher-resolution images and image edits from text.

Source

Google PaLM announced

Google described PaLM, a 540-billion-parameter language model trained with its Pathways system.

Source

OpenAI InstructGPT becomes API default

OpenAI made instruction-following models trained with human feedback (RLHF) the default in its API.

Source

OpenAI OpenAI API waitlist removed

OpenAI made its API available to developers in supported countries without a waitlist.

Source

OpenAI OpenAI Codex in the API

OpenAI released an improved Codex model to API beta users, able to turn natural language into code.

Source

OpenAI Codex paper: Evaluating LLMs Trained on Code

OpenAI published research on Codex, a GPT model fine-tuned on public code, and introduced the HumanEval benchmark.

Source

Google LaMDA unveiled

Google showed LaMDA, a conversational language model, at Google I/O.

Source

OpenAI DALL-E announced

OpenAI showed a GPT-3-style model that creates images from text descriptions.

Source

OpenAI CLIP released

OpenAI released CLIP, a model that learns to link images and text from internet data.

Source

OpenAI OpenAI API launched (GPT-3 private beta)

OpenAI opened a commercial API giving waitlisted developers access to GPT-3 models.

Source

OpenAI GPT-3 paper: Language Models are Few-Shot Learners

OpenAI described the 175-billion-parameter GPT-3 and showed it could do tasks from a few examples in the prompt.

Source

OpenAI Jukebox music generation model

OpenAI released a model that generates raw audio songs, including rudimentary singing.

Source

OpenAI Full GPT-2 (1.5B) released

OpenAI completed the staged release by publishing the full 1.5-billion-parameter GPT-2 weights.

Source

Google BERT used in Google Search

Google said it was using BERT to better understand search queries.

Source

OpenAI GPT-2 774M model released

As part of the staged release, OpenAI published the 774-million-parameter GPT-2 model with a six-month report.

Source

OpenAI Microsoft invests $1 billion in OpenAI

Microsoft agreed to invest one billion dollars and become OpenAI's exclusive cloud provider.

Source

OpenAI MuseNet music model

OpenAI showed a transformer that composes multi-instrument music in several styles.

Source

OpenAI OpenAI LP capped-profit structure

OpenAI created a capped-profit company under its non-profit to raise outside capital.

Source

OpenAI GPT-2 announced with staged release

OpenAI unveiled the 1.5-billion-parameter GPT-2 but released only a small version, citing misuse concerns.

Source

Google BERT paper

Google introduced BERT, a bidirectional transformer pre-trained on text, which set records on language understanding tasks.

Source

OpenAI GPT-1: generative pre-training paper

OpenAI described a transformer language model pre-trained on unlabeled text and then fine-tuned for specific tasks, the first model later called GPT.

Source

Google Attention Is All You Need

Google researchers introduced the Transformer architecture, which underlies modern large language models.

Source

Research Google Neural Machine Translation

Google announced a neural machine translation system, first used for Chinese to English in Google Translate, replacing its phrase-based approach.

Source

Paper DeepMind's WaveNet

DeepMind introduced WaveNet, a neural network that generates raw audio waveforms and produced more natural synthetic speech.

Source

Program Microsoft's Tay chatbot

Microsoft launched Tay, a Twitter chatbot that learned from users; it was taken offline within a day after being manipulated into posting offensive messages.

Source

Event AlphaGo defeats Lee Sedol

DeepMind's AlphaGo beat Go champion Lee Sedol four games to one in Seoul, a milestone many experts had expected to be a decade away.

Source

Paper AlphaGo paper and win over Fan Hui

DeepMind published AlphaGo in Nature, revealing it had beaten European champion Fan Hui 5–0 in October 2015, the first win by a program over a professional Go player on a full board without handicap.

Source

OpenAI OpenAI founded

OpenAI was announced as a non-profit AI research lab with backing from several tech investors.

Source

Paper Residual networks (ResNet)

Kaiming He and colleagues at Microsoft Research introduced residual connections that made very deep networks trainable, winning ImageNet 2015.

Source

Invention Google open-sources TensorFlow

Google released TensorFlow, its machine learning software library, as open source.

Source

Program Amazon Echo and Alexa

Amazon introduced the Echo speaker with its voice assistant Alexa, initially by invitation.

Source

Paper Sequence-to-sequence learning

Ilya Sutskever, Oriol Vinyals and Quoc Le showed that LSTM networks could map one sequence to another, achieving strong machine translation results.

Source

Paper Attention mechanism for neural translation

Dzmitry Bahdanau, Kyunghyun Cho and Yoshua Bengio added an attention mechanism letting a translation network focus on relevant source words, a precursor to transformers.

Source

Paper Generative adversarial networks

Ian Goodfellow and colleagues introduced GANs, in which two networks compete, one generating samples and one judging them, to produce realistic data.

Source

Event Eugene Goostman Turing test claim

Organisers at the University of Reading claimed the chatbot Eugene Goostman passed a Turing test by convincing a third of judges; many researchers disputed the claim.

Source

Event Google acquires DeepMind

Google agreed to buy DeepMind, reportedly for around 400 to 650 million U.S. dollars.

Source

Paper DeepMind's deep Q-network plays Atari

DeepMind showed a single neural network learning to play several Atari games directly from screen pixels using reinforcement learning.

Source

Paper word2vec

Tomas Mikolov and colleagues at Google published word2vec, an efficient way to represent words as vectors that capture meaning from context.

Source

Research Google Brain learns to recognise cats

A Google and Stanford team trained a large neural network on unlabelled YouTube frames; it learned features responding to faces and cats.

Source

Research AlexNet wins ImageNet

Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton's convolutional network, trained on GPUs, won the ImageNet challenge by a wide margin, starting the deep learning boom.

Source

Program Siri launches on the iPhone 4S

Apple announced Siri, a voice assistant built on technology from a startup it had acquired, as a feature of the iPhone 4S.

Source

Event IBM Watson wins Jeopardy!

IBM's Watson question-answering system beat champions Ken Jennings and Brad Rutter on the television quiz show Jeopardy!.

Source

Event DeepMind founded

Demis Hassabis, Shane Legg and Mustafa Suleyman founded DeepMind in London to research general-purpose learning systems.

Source

Research ImageNet dataset

Fei-Fei Li and colleagues presented ImageNet, a database of millions of labelled images that became the main benchmark for computer vision.

Source

Paper Hinton's deep belief networks

Geoffrey Hinton, Simon Osindero and Yee-Whye Teh showed how to train deep neural networks one layer at a time, a turning point for what became known as deep learning.

Source

Event Stanford's Stanley wins the DARPA Grand Challenge

Five autonomous vehicles completed the 212 km desert course, with Stanford's Stanley finishing first.

Source

Event First DARPA Grand Challenge

DARPA held a race for driverless vehicles in the Mojave Desert; no vehicle finished, and the best travelled about 12 km.

Source

Invention iRobot's Roomba

iRobot launched the Roomba vacuum cleaner, which became the first widely sold home robot.

Source

Invention Honda's ASIMO

Honda unveiled ASIMO, a humanoid robot able to walk smoothly and climb stairs.

Source

Paper LeNet-5 and the MNIST dataset

LeCun, Bottou, Bengio and Haffner described gradient-based learning for document recognition, including the LeNet-5 network used to read cheques.

Source

Paper Long short-term memory (LSTM)

Sepp Hochreiter and Jürgen Schmidhuber introduced LSTM, a recurrent network design that can retain information over long sequences.

Source

Event Deep Blue defeats Garry Kasparov

IBM's Deep Blue won a six-game match against world chess champion Garry Kasparov, the first match loss by a reigning champion to a computer under tournament conditions.

Source

Program Richard Wallace's A.L.I.C.E.

Richard Wallace released A.L.I.C.E., a pattern-matching chatbot written in AIML that later won the Loebner Prize three times.

Source

Paper Support-vector networks

Corinna Cortes and Vladimir Vapnik published support-vector machines, a classification method that dominated much of machine learning for the next decade.

Source

Program Tesauro's TD-Gammon

Gerald Tesauro of IBM built TD-Gammon, a neural network that learned backgammon by playing itself and reached near top human level.

Source

Event First Loebner Prize contest

The first Loebner Prize, an annual Turing-test style competition for chatbots, was held at the Computer Museum in Boston.

Source

Paper LeCun applies backpropagation to handwritten digits

Yann LeCun and colleagues at Bell Labs trained a convolutional neural network to read handwritten ZIP codes.

Source

Program Jabberwacky chatbot project begins

Rollo Carpenter began Jabberwacky, a chatbot meant to learn conversation from its users; it went online in 1997.

Source

Event The second AI winter

The market for specialised Lisp machines collapsed in 1987 and expert systems proved costly to maintain, leading to another drop in AI funding.

Source

Paper Rumelhart, Hinton and Williams popularise backpropagation

A paper in Nature showed how backpropagation of errors could train multi-layer neural networks to learn useful internal representations.

Source

Film The Terminator

James Cameron's film about Skynet, a defence AI that turns on humanity, became a common cultural reference for AI risk.

Source

Literature William Gibson's Neuromancer

Gibson's novel, featuring AIs named Wintermute and Neuromancer, popularised the idea of cyberspace and shaped cyberpunk fiction.

Source

Research The Cyc project begins

Douglas Lenat started Cyc, a long-running effort to hand-encode millions of common-sense facts and rules.

Source

Film Blade Runner

Ridley Scott's film, adapted from Philip K. Dick's novel, brought artificial humans called replicants to a wide audience.

Source

Paper Hopfield networks

Physicist John Hopfield described a recurrent neural network that stores memories as stable states, helping revive interest in neural networks.

Source

Event Japan's Fifth Generation Computer project

Japan's trade ministry launched a ten-year national programme to build logic-based computers for AI, prompting rival programmes in the U.S. and Europe.

Source

Paper Searle's Chinese room argument

Philosopher John Searle argued that a program manipulating symbols by rules would not understand language, even if its answers looked fluent.

Source

Program XCON expert system enters commercial use

Digital Equipment Corporation began using XCON (R1), a rule-based system that configured computer orders, a sign of the commercial expert systems boom.

Source

Event The first AI winter

After the Lighthill report and U.S. funding changes, money and interest in AI research fell sharply for several years.

Source

Event The Lighthill report

A review for the UK Science Research Council by James Lighthill judged that AI had failed to deliver on its promises, leading to major funding cuts in Britain.

Source

Program MYCIN expert system

Stanford's MYCIN used several hundred rules to recommend antibiotics for blood infections, performing comparably to specialists in tests though it was never used clinically.

Source

Program Colby's PARRY

Psychiatrist Kenneth Colby built PARRY, a program simulating a person with paranoid schizophrenia; it was later connected to ELIZA for a recorded machine-to-machine conversation.

Source

Invention Prolog programming language

Alain Colmerauer and Philippe Roussel created Prolog, a logic programming language widely used in AI and later chosen for Japan's Fifth Generation project.

Source

Program Winograd's SHRDLU

Terry Winograd's SHRDLU understood typed English commands about a simulated world of blocks and could answer questions about it.

Source

Paper McCarthy and Hayes describe the frame problem

John McCarthy and Patrick Hayes identified the frame problem: how a reasoning system can represent what stays unchanged when an action occurs.

Source

Paper Minsky and Papert's Perceptrons

Marvin Minsky and Seymour Papert analysed the limits of single-layer perceptrons, and the book is often linked to a decline in neural network research.

Source

Film 2001: A Space Odyssey and HAL 9000

Stanley Kubrick's film, written with Arthur C. Clarke, premiered with HAL 9000, a conversational computer that turns against the crew.

Source

Literature Philip K. Dick's Do Androids Dream of Electric Sheep?

Dick's novel follows a bounty hunter tracking androids nearly indistinguishable from humans, raising questions about empathy and identity.

Source

Event ALPAC report on machine translation

A U.S. advisory committee concluded that machine translation had not met expectations, leading to sharp cuts in funding for the work.

Source

Research Shakey the robot at SRI

SRI's Shakey was the first mobile robot to reason about its own actions, combining vision, planning and movement; the project produced the A* search algorithm.

Source

Program Weizenbaum's ELIZA

Joseph Weizenbaum published ELIZA, a program that held text conversations by matching patterns; its 'DOCTOR' script imitated a psychotherapist and many users attributed understanding to it.

Source

Program DENDRAL project begins

Edward Feigenbaum, Joshua Lederberg and colleagues at Stanford started DENDRAL, a program that inferred molecular structures from mass-spectrometry data and became the first expert system.

Source

Invention Unimate, the first industrial robot, goes to work

The Unimate robotic arm, developed by George Devol and Joseph Engelberger, began handling hot die-cast parts at a General Motors plant in New Jersey.

Source

Paper Arthur Samuel's checkers program and 'machine learning'

Arthur Samuel of IBM published his work on a checkers program that improved by playing itself, helping establish the term machine learning.

Source

Event McCarthy and Minsky start the MIT AI Project

McCarthy and Minsky founded the Artificial Intelligence Project at MIT, which grew into one of the field's leading laboratories.

Source

Invention John McCarthy creates Lisp

McCarthy designed Lisp, a programming language based on symbolic lists that became the main language of AI research for decades.

Source

Research Rosenblatt's Perceptron

Frank Rosenblatt published the perceptron, a trainable single-layer neural network, and demonstrated it for the U.S. Navy; the Mark I Perceptron hardware followed.

Source

Program General Problem Solver

Newell, Shaw and Simon began the General Problem Solver, a program meant to solve many kinds of formal problems using means-ends analysis.

Source

Event The Dartmouth Summer Research Project

About a dozen researchers met at Dartmouth College over roughly eight weeks; the workshop is widely treated as the founding event of AI as a field.

Source

Program Newell, Shaw and Simon's Logic Theorist

Allen Newell, Cliff Shaw and Herbert Simon presented the Logic Theorist, a program that proved theorems from Principia Mathematica, often called the first AI program.

Source

Paper Dartmouth proposal coins 'artificial intelligence'

John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon proposed a summer research project on 'artificial intelligence', the first known use of the term.

Source

Event Georgetown–IBM machine translation demonstration

Georgetown University and IBM publicly demonstrated a computer translating more than sixty Russian sentences into English, the first public machine-translation demo.

Source

Program Strachey's draughts program

Christopher Strachey wrote one of the first game-playing programs, a draughts (checkers) player that ran on the Ferranti Mark 1.

Source

Invention Minsky and Edmonds build SNARC

Marvin Minsky and Dean Edmonds built SNARC, a machine of about 40 artificial neurons using vacuum tubes that simulated a rat learning a maze.

Source

Literature Asimov's I, Robot

Asimov collected his robot stories, built around the Three Laws, into the book I, Robot.

Source

Paper Turing's 'Computing Machinery and Intelligence'

Turing replaced the question 'Can machines think?' with the imitation game, now called the Turing test, and answered common objections to machine intelligence.

Source

Paper Shannon's 'Programming a Computer for Playing Chess'

Claude Shannon set out how a computer could play chess by evaluating positions and searching ahead, a template for game-playing programs.

Source

Invention Grey Walter's robot tortoises

Neurophysiologist William Grey Walter built small autonomous robots, Elmer and Elsie, that used light sensors to steer and find their charging station.

Source

Paper Hebb's The Organization of Behavior

Psychologist Donald Hebb proposed that connections between neurons strengthen when they are active together, a learning rule that influenced neural network research.

Source

Paper Turing's report 'Intelligent Machinery'

In an unpublished report for the National Physical Laboratory, Turing discussed machines that learn and early network-like 'unorganised machines'.

Source

Paper Norbert Wiener's Cybernetics

Norbert Wiener's book introduced cybernetics, the study of control and communication through feedback in animals and machines.

Source

Invention ENIAC is unveiled

The University of Pennsylvania publicly presented ENIAC, a programmable electronic general-purpose computer built for the U.S. Army.

Source

Paper McCulloch and Pitts model the neuron

Warren McCulloch and Walter Pitts showed that networks of simple on/off model neurons could compute logical functions, the first mathematical model of a neural network.

Source

Literature Asimov's Three Laws of Robotics in 'Runaround'

Isaac Asimov's short story 'Runaround' set out the Three Laws of Robotics, rules meant to keep robots safe and obedient to humans.

Source

Paper Turing's 'On Computable Numbers'

Alan Turing described an abstract machine that manipulates symbols on a tape and showed that a universal machine could carry out any computation, a foundation of computer science.

Source

Film Fritz Lang's Metropolis

Lang's silent film, which features a robot made to impersonate a woman named Maria, premiered in Berlin and shaped later screen images of robots.

Source

Literature Karel Čapek's R.U.R. introduces the word 'robot'

Čapek's play about artificial workers who rebel against their makers premiered in Prague and gave the world the word robot.

Source

Invention Torres Quevedo's El Ajedrecista

Spanish engineer Leonardo Torres Quevedo built an electromechanical machine that could play a king-and-rook chess endgame against a human without help.

Source

Literature E. M. Forster's 'The Machine Stops'

Forster's short story imagines people living isolated lives entirely dependent on a global machine that eventually fails.

Source

Literature L'Ève future popularises the word 'android'

Auguste Villiers de l'Isle-Adam's novel features an artificial woman built by a fictional Thomas Edison and helped spread the term andréide (android).

Source

Literature Butler's Erewhon and 'The Book of the Machines'

Butler's novel Erewhon describes a society that has banned machines for fear they would develop consciousness.

Source

Literature Samuel Butler's 'Darwin among the Machines'

In a letter to a New Zealand newspaper, Samuel Butler suggested that machines might evolve and one day overtake humanity.

Source

Paper Boole's The Laws of Thought

George Boole published an algebra of logic in which reasoning is expressed with symbols and operations, the basis of the Boolean logic used in computers.

Source

Paper Ada Lovelace's notes on the Analytical Engine

Ada Lovelace translated Menabrea's article on the Analytical Engine and added notes, including a method to compute Bernoulli numbers and the remark that the machine could not originate anything on its own.

Source

Invention Babbage describes the Analytical Engine

Babbage designed a general-purpose mechanical computer with a store, a mill and punched-card input, though it was never completed.

Source

Invention Babbage proposes the Difference Engine

Charles Babbage announced plans for a machine to calculate mathematical tables automatically, the first of his mechanical computing designs.

Source

Literature Mary Shelley's Frankenstein

Mary Shelley's novel tells of a scientist who creates a living being and then abandons it, a lasting reference point for debates about creators and their creations.

Source

Literature E. T. A. Hoffmann's 'The Sandman'

Hoffmann's story follows a student who falls in love with Olimpia, a lifelike automaton, a theme later studied as the uncanny.

Source

Invention The Mechanical Turk chess automaton

Wolfgang von Kempelen presented a chess-playing automaton at the Habsburg court; it toured for decades before being revealed as a hidden human operator.

Source

Invention Jaquet-Droz automata: the Writer, the Draughtsman and the Musician

Swiss watchmakers Pierre Jaquet-Droz, his son Henri-Louis and Jean-Frédéric Leschot built three automata, including a boy who writes a programmable text with a quill.

Source

Invention Vaucanson's Digesting Duck

Jacques de Vaucanson exhibited a mechanical duck that appeared to eat and digest grain, one of the most famous automata of the 18th century.

Source

Literature Swift's 'Engine' in Gulliver's Travels

Jonathan Swift's satire describes a professor at Lagado whose engine of rotating word blocks produces books by random combination, an early parody of mechanical text generation.

Source

Idea Leibniz imagines a calculus of reasoning

Gottfried Wilhelm Leibniz's Dissertation on the Art of Combinations began his project of a universal symbolic language and a calculus ratiocinator in which disputes could be settled by calculation.

Source

Idea Hobbes: reasoning is reckoning

In Leviathan, Thomas Hobbes described reasoning as a form of calculation, adding and subtracting ideas, a view later linked to computational theories of mind.

Source

Idea Descartes asks whether machines could think

In the Discourse on the Method, René Descartes argued that a machine could imitate a body but could never use language flexibly or reason across all situations as a person does.

Source

Myth The golem of Prague legend

Jewish folklore tells of a golem, a clay figure animated through sacred words; the best-known version credits Rabbi Judah Loew of Prague with making one to protect the community.

Source

Invention Leonardo da Vinci's mechanical knight

Leonardo da Vinci sketched an armoured knight moved by pulleys and cables that could sit, stand and move its arms.

Source

Idea Ramon Llull's Ars Magna

The Majorcan philosopher Ramon Llull presented a method of combining basic concepts with rotating paper discs to generate statements mechanically, later cited as an ancestor of logical computation.

Source

Invention Al-Jazari's Book of Knowledge of Ingenious Mechanical Devices

The engineer Ismail al-Jazari completed an illustrated treatise describing water clocks and automata, including mechanical musicians and a serving figure.

Source

Invention Hero of Alexandria writes on automata

Hero of Alexandria described programmable automata and self-moving theatre devices driven by weights, ropes and pegged drums.

Source

Literature Ovid's Pygmalion: a statue brought to life

In the Metamorphoses, Ovid tells of the sculptor Pygmalion, whose ivory statue is brought to life, an early story of a made figure becoming a person.

Source

Myth Talos, the bronze guardian of Crete, in the Argonautica

Apollonius of Rhodes' Argonautica tells of Talos, a bronze giant made to guard Crete, who is defeated when the fluid in his single vein drains away.

Source

Myth Homer describes Hephaestus' golden attendants

The Iliad describes the smith god Hephaestus working with golden maidservants who can move, speak and think, one of the oldest images of artificial helpers.

Source