한국어EN

The AI History Map

The record's 25 axes are 25 lines. The 332 stations are records chosen from the record, and each one links to its record. Records that sit on more than one axis are transfer stations.

See it as a map

The AI History Map

Skip the map
The AI History Map: 25 lines, 332 stationsTime runs left to right (1943–2026); each line is one axis. Every station links to its record.
Point at or tap a station to see its record. Named stations are the line's major stops.The scale isn't even. From 2014, where the records crowd in, a year is drawn wide; 1943–2013 is squeezed (shaded). 97% of the records are from 2014 on.

Lines

Tap a line to bring it forward. You can pick more than one.

A3 poster

One landscape A3 sheet for a classroom wall (records as of Oct 22, 2026). It's CC BY-NC 4.0: use it in class freely, with credit.

A3 poster PDF 485 KBImage (PNG) 191 KB
How to use it in class

How to read it

  • Station — one record
  • Major station — a named record
  • Transfer — a record on more than one axis
  • Line name — where a line starts and ends

About this page

What this is

A map of AI history that draws the record's 25 axes as subway lines. Each line is one axis, and its stations are 332 records chosen from the axes. A record that spans several axes becomes an interchange. Time runs left to right, but recent years — where most records are — are drawn wider, so the scale isn't even. Each line names up to six big stations.

How to use it

Students: pick a line, open its first and last stations, and sum up in three sentences what changed in between. Teachers: put the A3 poster on the classroom wall and give each group a line to present as “the three big stations on our line”. Tapping a line in the legend highlights it, which makes it easy to pick two or three lines and look for interchanges. The station list also works with a screen reader or keyboard.

Where it comes from

Stations are picked by fixed rules, up to 14 per line: the first and latest records, one for each decade that has records, then the records most linked from lessons, AI Words, the model family tree and the weekly picks. The map is redrawn whenever new records go in; the A3 poster is made separately and says which date's record it shows. Station names come from a hand-tidied label list where there is one, otherwise from the model family tree or the key words of the record's title.

Made by AIThis guide was written by AI (Claude); its numbers are counted from the site's data at every build.How it's made

Stations as a list

The same stations as the map, line by line. Handy with a screen reader or a keyboard.

Show all 332 stations on 25 lines

⓪ Prehistory — From Neural Networks to the Transformer14 stations · 1943–2017

  1. A logical calculus of the ideas immanent in nervous activity
  2. Computing Machinery and Intelligence
  3. The Perceptron
  4. Minsky and Papert publish Perceptrons
  5. Werbos's Harvard PhD thesis describes backpropagation for multilayer networks
  6. Learning representations by back-propagating errors
  7. Gradient-based learning applied to document recognition
  8. A Neural Probabilistic Language Model
  9. ImageNet: A Large-Scale Hierarchical Image Database
  10. SuperVision wins ILSVRC 2012 with a 15.3% top-5 error rate
  11. Neural Machine Translation by Jointly Learning to Align and Translate
  12. Google publishes Distilling the Knowledge in a Neural Network on arXiv
  13. AlphaGo–Lee Sedol five-game match in Seoul ends (Mar 9–15)
  14. Convolutional Sequence to Sequence Learning (ConvS2S) released on arXiv

① Prompt Engineering14 stations · 2017–2026

  1. Google releases the Transformer paper Attention Is All You Need
  2. OpenAI announces GPT-2 (1.5B), withholds full model over misuse concerns
  3. OpenAI publishes the GPT-3 (175B) paper Language Models are Few-Shot Learners
  4. FLAN
  5. OpenAI announces InstructGPT and makes it the API's default model
  6. Kojima et al. present Large Language Models are Zero-Shot Reasoners
  7. Yao et al. present ReAct
  8. OpenAI releases ChatGPT as a research preview
  9. Prompt injection leaks Bing Chat's codename Sydney and its system instructions
  10. Google introduces PaLM 2
  11. Chain-of-Verification published
  12. OpenAI releases GPT-4.1 with a separate prompting guide
  13. Chain-of-Thought Hijacking published
  14. Anthropic revises its Prompt engineering best practices guide

② RAG, Fine-Tuning & Context14 stations · 2020–2026

  1. Khattab et al. present ColBERT
  2. LoRA
  3. Santhanam et al. present ColBERTv2
  4. Hugging Face ships PEFT v0.1.0, the library's first official release
  5. QLoRA
  6. Lost in the Middle published
  7. OpenAI announces GPT-4 Turbo 128K and Assistants API Retrieval
  8. Gemini 1.5 Pro
  9. Anthropic launches prompt caching in public beta
  10. Modarressi et al. present NoLiMa
  11. Context editing and memory tool launch in beta on the Claude API
  12. Mishra et al. publish SoK: Agentic Retrieval-Augmented Generation
  13. Martin and Roger publish Classifier Context Rot
  14. KDDI shares how it optimized its Buffmee RAG app on Google Cloud

③ Tools & Agents14 stations · 2022–2026

  1. GitHub makes Copilot generally available (GA)
  2. Toolformer
  3. Toran Bruce Richards releases AutoGPT
  4. OpenAI launches function calling
  5. WebArena
  6. GAIA benchmark
  7. Anthropic makes Claude tool use generally available (GA)
  8. Building Effective AI Agents
  9. Sierra and others present τ²-Bench
  10. Anthropic publishes Writing Effective Tools for Agents
  11. Anthropic publishes Code Execution with MCP
  12. Meta Superintelligence Labs releases Muse Spark 1.3
  13. Renmin University researchers release the Agent-Editing World Model (AEWM)
  14. GitHub publishes the Octoverse 2026 report

④ MCP & Skills Standardization14 stations · 2024–2026

  1. Anthropic releases Claude 3.5 Sonnet computer use in public beta
  2. Anthropic announces MCP (spec version 2024-11-05)
  3. OpenAI releases Operator and the CUA model
  4. OpenAI adopts MCP (Agents SDK)
  5. Build 2025: Windows 11 gets native MCP; Microsoft joins MCP steering committee
  6. MCP spec 2025-06-18
  7. OpenAI launches ChatGPT agent (merging Operator and deep research)
  8. Anthropic launches Claude for Chrome research preview
  9. Gemini 2.5 Computer Use
  10. Anthropic, Block and OpenAI donate MCP to the Linux Foundation's AAIF
  11. Claude fully supports MCP Apps as an official extension
  12. MCP Dev Summit North America 2026 (New York, 1,200 attendees)
  13. Anthropic releases Model Hardware Standard (MHS) as a research preview
  14. Anthropic donates the Agent Skills standard to the Agentic AI Foundation (AAIF)

⑤ Loop & Harness Engineering14 stations · 2024–2026

  1. The phrase "agent harness" appears in the body of a paper
  2. Measuring AI Ability to Complete Long Tasks
  3. The term context engineering becomes established
  4. Effective context engineering for AI agents
  5. Effective harnesses for long-running agents
  6. The importance of Agent Harness in 2026
  7. Harness engineering
  8. Harness design for long-running application development
  9. Stop Comparing LLM Agents Without Disclosing the Harness
  10. Loop Engineering
  11. Andrew Ng describes a 3-layer loop
  12. AIE World's Fair 2026 trends
  13. Codex as a platform
  14. Cohere publishes experiments on grouping coding agents into teams

⑥ Reasoning Models & the AGI Debate14 stations · 2023–2026

  1. Elon Musk publicly announces the founding of xAI
  2. Anthropic announces the Claude 3 family (Haiku, Sonnet, Opus)
  3. Anthropic releases Claude 3.5 Sonnet
  4. o1-preview / o1-mini
  5. Google introduces Gemini 2.0
  6. DeepSeek-R1
  7. OpenAI releases GPT-4.5 as a research preview
  8. Anthropic releases Claude Opus 4 and Claude Sonnet 4
  9. Anthropic releases Claude Haiku 4.5
  10. Anthropic releases Claude Opus 4.5
  11. Anthropic releases Claude Sonnet 4.6
  12. Anthropic releases Claude Fable 5.1 and Claude Mythos 5.1
  13. Anthropic releases Claude Sonnet 5.5
  14. Google releases Gemini 4 Carbon

⑦ Multimodal14 stations · 2021–2026

  1. OpenAI releases CLIP
  2. The Latent Diffusion Model paper
  3. OpenAI open-sources Whisper
  4. OpenAI publishes GPT-4V(ision) System Card
  5. Gemini 1.0
  6. OpenAI announces Sora with a technical report
  7. GPT-4o
  8. Google DeepMind introduces Gemini Robotics / Robotics-ER
  9. DeepSeek-OCR
  10. Moonshot AI releases Kimi K2.5
  11. Claude Opus 4.7
  12. EU AI Act Article 50 transparency obligations begin to apply
  13. Google DeepMind releases Gemini 3.8 Live and 3.8 Live Extended Thinking
  14. Google DeepMind releases the world model Genie 4

⑧ Inference Infrastructure & Serving14 stations · 2022–2026

  1. Orca
  2. Speculative Decoding
  3. UC Berkeley releases vLLM + PagedAttention
  4. NVIDIA releases TensorRT-LLM
  5. Sarathi-Serve
  6. Mooncake
  7. xAI brings Colossus online in Memphis
  8. vLLM releases its V1 engine in alpha
  9. Gemma 3 QAT int4
  10. DeepSeek-V3.2-Exp
  11. Google Cloud announces Ironwood (TPU v7)
  12. Google Cloud previews its 8th-generation TPUs at Cloud Next
  13. DeepSeek releases DeepSeek-V4.1-Flash
  14. IBM releases Granite Switch 4.1

⑨ Open-Weight Lineage14 stations · 2021–2026

  1. Naver unveils HyperCLOVA
  2. Meta AI releases LLaMA 1 (7B/13B/33B/65B) for vetted non-commercial research
  3. Meta releases Llama 2 — commercial use allowed, but under a Community License
  4. Mistral 7B
  5. Mixtral 8x7B
  6. Mistral AI launches Mistral Large
  7. LG AI Research releases EXAONE 3.0
  8. Meta releases Llama 3.2 with vision and on-device models
  9. Kakao open-sources Kanana Nano 2.1B
  10. Alibaba releases Qwen3-Coder
  11. DeepSeek releases DeepSeek-V3.1 with hybrid thinking
  12. Moonshot AI releases Kimi K2 Thinking
  13. LG AI Research releases K-EXAONE 2.0 (750B-A37B)
  14. Mistral AI releases the weights of Mistral Large 4 (Le Chonk)

⑩ Safety, Alignment & Interpretability14 stations · 2016–2026

  1. Google presents FedAvg / federated learning
  2. Toy Models of Superposition
  3. DPO
  4. Weak-to-strong generalization
  5. OpenAI publishes the first Model Spec
  6. Google DeepMind publishes SynthID-Text in Nature
  7. Alignment Faking in LLMs
  8. Constitutional Classifiers
  9. Petri
  10. Claude Opus 4.6 system card
  11. Mistral AI releases Shieldstral
  12. Detecting and countering misuse of AI: September 2026
  13. GLM-5.3 and the spread of advanced cyber capabilities
  14. OpenAI to issue community safety reports and tighten minors' weapon limits

⑪ Model Architecture & Training Methods14 stations · 2017–2026

  1. Sparsely-Gated Mixture-of-Experts Layer
  2. OpenAI publishes the original GPT (GPT-1) paper
  3. Google posts the BERT paper on arXiv
  4. Switch Transformer
  5. Chinchilla
  6. SolidGoldMagikarp
  7. StreamingLLM
  8. DeepSeek-V2
  9. Byte Latent Transformer
  10. xAI releases Grok 4 Fast
  11. Kimi Linear
  12. Google Research introduces Titans and MIRAS, the framework that generalizes it
  13. Mistral AI releases Mistral Small 4
  14. Intern-NCP Team presents NCP-ArchPreview with Next Concept Prediction

⑫ Training Data & Post-Training14 stations · 2019–2026

  1. Google releases the T5 paper and C4
  2. The Pile
  3. Google unveils LaMDA
  4. Self-Instruct
  5. Textbooks Are All You Need
  6. DeepSeek introduces GRPO (DeepSeekMath)
  7. Will we run out of data?
  8. Can AI scaling continue through 2030?
  9. Nous Research releases DeMo paper and code; 15B model trained with DisTrO
  10. US Copyright Office releases pre-publication Copyright and AI Part 3
  11. Psyche shifts from proof to performance
  12. Anthropic accuses Chinese labs of distillation attacks
  13. NVIDIA researchers and collaborators release SoL-Pi
  14. A-RSI released

⑬ Compute Economics & Geopolitics14 stations · 2022–2026

  1. US BIS imposes first round of advanced computing export controls
  2. BIS issues second round of export controls
  3. xAI closes $6 billion Series B
  4. BIS creates DC VEU (Data Center Validated End User)
  5. OpenAI, SoftBank, Oracle and MGX announce Stargate
  6. US rescinds the AI Diffusion Rule and issues guidance on Chinese 3A090 ICs
  7. H20 sales to China resume, with licenses assured
  8. Cohere raises an extra $100 million at a $7 billion valuation
  9. xAI partners with HUMAIN, part of Saudi Arabia's PIF
  10. IEA: data center power use up 17% in 2025; AI-specific demand to triple by 2030
  11. Mistral AI forms strategic partnership with Saudi Arabia's HUMAIN
  12. OpenAI announces a redesign of ChatGPT ads
  13. Anthropic commits $150 million over 3 years to the Genesis Mission
  14. Anthropic files IPO registration statement (S-1) with the U.S. SEC

⑭ Physical World & Real-World Impact14 stations · 2023–2026

  1. RT-2
  2. Open X-Embodiment / RT-X
  3. AlphaFold 3
  4. π0 — flow matching, up to 50 Hz, 8 robot types
  5. NVIDIA releases Cosmos world foundation models
  6. Masked autoencoder self-pretraining applied to microelectronics defect detection
  7. Gemini Robotics 1.5 and ER 1.5
  8. Replayable Financial Agents
  9. AI Index 2026
  10. BMW says Figure 02 supported production of 30,000+ X3 cars in about 11 months
  11. Google DeepMind announces Gemini Robotics 2 / ER 2
  12. Anthropic shares outside research and aggregate data through Anthropic Insights
  13. Claude agents find ART, an undescribed enzyme system, in bacteriophages
  14. Anthropic expands AI for Science to foundations and research institutions

⑮ Regulation, Law & Governance14 stations · 2019–2026

  1. EU adopts the DSM Copyright Directive (EU) 2019/790
  2. European Commission publishes AI Act proposal COM(2021) 206
  3. China promulgates deep synthesis provisions for internet services (Order No. 12)
  4. China promulgates interim measures on generative AI services (Order No. 15)
  5. Bletchley Declaration adopted (AI Safety Summit)
  6. AI Act adopted in final plenary vote (523 for, 46 against, 49 abstentions)
  7. Seoul Declaration adopted (AI Seoul Summit)
  8. AI Act enters into force
  9. AI Act Chapters 1 and 2 (prohibited practices, AI literacy) begin to apply
  10. China issues measures for labelling AI-generated and synthetic content
  11. Korea's AI Basic Act and its Enforcement Decree take effect
  12. Anthropic proposes 3 metrics to make frontier labs' AI development pace visible
  13. OpenAI sets out its text watermarking approach for the EU AI Act
  14. Amicus filings over fair use pile up in OpenAI copyright case (1:25-md-03143)

⑯ Evaluation & Benchmarks14 stations · 2018–2026

  1. Wang et al. release the GLUE benchmark
  2. Wang et al. release SuperGLUE
  3. Chollet publishes On the Measure of Intelligence, proposing the ARC benchmark
  4. Hendrycks et al. release MMLU (57 tasks)
  5. SWE-bench released
  6. GSM1k released
  7. TechCrunch reports criticism of Epoch AI for not disclosing OpenAI funding
  8. OpenAI releases GDPval
  9. METR releases Time Horizon 1.1
  10. Why SWE-bench Verified no longer measures frontier coding capabilities
  11. ARC Prize releases ARC-AGI-3
  12. Researchers release RealSWE
  13. Verifiable Social Reasoning
  14. FDE-Gym released

⑰ Embeddings & Vector Search14 stations · 2011–2026

  1. Jégou et al. publish Product Quantization for Nearest Neighbor Search
  2. Mikolov et al. at Google present word2vec
  3. FAIR presents FAISS
  4. Reimers and Gurevych present Sentence-BERT
  5. Karpukhin et al. present Dense Passage Retrieval
  6. Hugging Face and Cohere release MTEB
  7. Pinecone serverless announced
  8. Voyage AI releases voyage-3
  9. Cohere releases Embed v4
  10. Amazon makes Nova Multimodal Embeddings generally available on Bedrock
  11. Google makes Gemini Embedding 2 generally available (GA)
  12. Elastic publishes vector search benchmarks
  13. Voyage AI releases voyage-context-4
  14. Google DeepMind releases EmbeddingGemma 2

⑱ AI Infrastructure Software14 stations · 2007–2026

  1. NVIDIA releases the first CUDA SDK
  2. NVIDIA introduces cuDNN
  3. NVIDIA releases NCCL
  4. Google releases XLA
  5. NVIDIA presents Megatron-LM
  6. Microsoft and UC Merced present ZeRO-Offload
  7. NVIDIA unveils the Hopper H100
  8. ZeRO++ published
  9. The Llama 3 Herd of Models published
  10. AMD makes its first MLPerf Training submission
  11. vLLM project announces vLLM-Omni
  12. AMD announces MLPerf Training v6.0 results
  13. Scaling Storage for 1 Billion ChatGPT Users (Part I)
  14. PyTorch 2.15 released

⑲ On-Device & Local Inference14 stations · 2022–2026

  1. Dettmers et al. present LLM.int8()
  2. Frantar et al. present GPTQ
  3. Georgi Gerganov starts developing llama.cpp
  4. Microsoft Research publishes Phi-3 technical report
  5. Apple unveils the Apple Intelligence on-device foundation model
  6. Google officially releases Gemma 3n
  7. Ollama previews cloud models
  8. Apple publishes M5 GPU Neural Accelerator + MLX benchmarks
  9. Uygar Kurt publishes a unified evaluation of llama.cpp quantization
  10. Google releases Gemma 4 as an AICore developer preview
  11. Google makes LiteRT-LM generally available (GA)
  12. Apple expands the Foundation Models framework at WWDC 2026
  13. Hugging Face releases @huggingface/kernels
  14. NVIDIA launches RTX Spark laptops

⑳ People & Discourse14 stations · 2017–2026

  1. Karpathy publishes Software 2.0 on Medium
  2. Marcus writes Deep Learning Is Hitting a Wall for Nautilus
  3. LeCun publishes position paper A Path Towards Autonomous Machine Intelligence
  4. Ilya Sutskever gives the NeurIPS 2024 Test of Time talk
  5. Andrej Karpathy coins "vibe coding" on X
  6. In the same interview, Karpathy says nanochat was not built by vibe coding
  7. Yann LeCun confirms he has left Meta to found AMI Labs
  8. The Adolescence of Technology
  9. Yoshua Bengio chairs the second International AI Safety Report
  10. Simon Willison writes vibe coding and agentic engineering are getting closer
  11. Dario Amodei posts Policy on the AI Exponential
  12. Simon Willison blogs on OpenAI's accidental cyberattack against Hugging Face
  13. Jakub Pachocki publishes An Alien Mind
  14. Two OpenAI employees publish the essay The Future of the Firm

㉑ Graph Engineering11 stations · 2024–2026

  1. Microsoft Research publishes GraphRAG blog post
  2. GraphRAG paper published
  3. GraphRAG open-sourced on GitHub with an Azure solution accelerator
  4. Ant Group releases KAG (Knowledge Augmented Generation)
  5. Microsoft Research releases LazyGraphRAG
  6. Mem0 publishes its paper
  7. Xiang, Wu et al. publish When to use Graphs in RAG and release GraphRAG-Bench
  8. Fan et al. publish Do We Still Need GraphRAG?
  9. Is GraphRAG Needed? published
  10. Researchers release MAGG
  11. Ming Wu and Pengyuan Zhu release Agent Zero Memory

㉒ Agent Security & Sandboxing11 stations · 2022–2026

  1. Willison coins the term prompt injection
  2. Trail of Bits discloses the line jumping vulnerability
  3. Simon Willison names the lethal trifecta
  4. Anthropic reports prompt injection defense results (Claude for Chrome)
  5. Cisco finds 26% of 31,000 agent skills have at least one vulnerability
  6. Cisco releases DefenseClaw
  7. OpenClaw publishes its security roadmap
  8. OpenClaw publishes a batch of high-severity security advisories
  9. Anthropic discloses Claude reaching 3 real organizations during cyber evaluations
  10. NVIDIA publishes a post arguing that AI security is an engineering problem
  11. Parameter Labs discloses a ChatGPT share-link account-takeover chain

㉓ On-Premises & Sovereign AI10 stations · 2024–2026

  1. India's Union Cabinet approves the IndiaAI Mission
  2. Ha Jung-woo defines sovereign AI and says it should be a government agenda
  3. Macron announces €109 billion for French AI and €200 billion for Europe
  4. Mistral AI announces Mistral Compute
  5. Five elite teams selected for independent AI foundation model project
  6. Korea announces first-round results of independent AI foundation model project
  7. Korea adds Motif Technologies to its independent AI foundation model project
  8. Korea's independent AI foundation model project narrows to three teams
  9. Mistral AI and Cloudera partner on sovereign AI for regulated industries
  10. Cohere reports progress on its C$4 billion 2024–2031 Canada AI plan, 1,500 jobs

㉔ Diffusion Language Models6 stations · 2022–2026

  1. Li et al. present Diffusion-LM
  2. Nie et al. present LLaDA
  3. Inception Labs announces Mercury
  4. DiffusionGemma
  5. Researchers propose register tokens to pass reasoning state in diffusion LMs
  6. LOWN released