# LLMilestones

> A milestone timeline of AI, from the first artificial neuron (1943) to the frontier models of 2026: what happened, why it mattered, and what it cost.

LLMilestones is a curated, source-checked timeline of 78 milestones in artificial intelligence, focused on large language models. Every milestone has a date, the organization behind it, a short explanation of why it mattered and what it cost, and a link to its primary source. Milestones marked "Official page not opened" or "No official page found" have weaker sourcing. A Chinese version of every page is available (same URL with .zh.html).

- [All milestones](https://llmilestones.com/m/index.html): text index of every milestone, grouped by era
- [Full text for LLMs](https://llmilestones.com/llms-full.txt): every milestone in one plain-text file
- [Interactive timeline](https://llmilestones.com/): the 3D timeline (requires JavaScript)

## Prologue: Before the Transformer (1943 — 2016)

From a mathematical model of one neuron to beating the world Go champion.

- [The artificial neuron: the McCulloch-Pitts model](https://llmilestones.com/m/mcculloch.html): 1943-12, McCulloch & Pitts. Described a neuron in math and logic: sum the weighted inputs, and fire if the total passes a threshold.
- [The perceptron: a machine that learns](https://llmilestones.com/m/perceptron.html): 1958-11, Rosenblatt. Proposed the perceptron, which adjusts its own weights from examples, and built it in hardware as the Mark I.
- [Receptive fields in the cat's visual cortex](https://llmilestones.com/m/hubel.html): 1959-10, Hubel & Wiesel. Found that different neurons in the visual cortex respond to edges at specific orientations, combined layer by layer.
- [Neocognitron](https://llmilestones.com/m/neocognitron.html): 1980-04, Fukushima. Copied the visual cortex's hierarchy, recognizing shapes by combining local receptive fields layer by layer.
- [Training neural networks with backpropagation](https://llmilestones.com/m/backprop.html): 1986-10-09, Rumelhart, Hinton & Williams. Used the chain rule to pass errors back from the output layer, layer by layer, to train multi-layer networks.
- [LeNet: reading handwritten digits with a CNN](https://llmilestones.com/m/lenet.html): 1989-12, LeCun et al. · Bell Labs. Trained a convolutional neural network with backpropagation to read handwritten zip codes.
- [AlexNet](https://llmilestones.com/m/alexnet.html): 2012-09-30, Krizhevsky, Sutskever & Hinton. A deep convolutional network trained on two GPUs that won the ImageNet image-recognition challenge by a huge margin.
- [Playing Atari with deep reinforcement learning](https://llmilestones.com/m/dqn.html): 2013-12-19, DeepMind. Combined a deep neural network with reinforcement learning and learned several Atari games from screen pixels alone.
- [Generative adversarial networks (GANs)](https://llmilestones.com/m/gan.html): 2014-06-10, Goodfellow et al.. A generator and a discriminator compete, and the generator gradually learns to produce convincing fakes.
- [AlphaGo beats the world Go champion](https://llmilestones.com/m/alphago.html): 2016-03-15, DeepMind. Defeated world Go champion Lee Sedol 4 to 1.

## Act I: Foundations (2017 — 2019)

One paper dropped recurrence, and the pre-training paradigm was born.

- [Attention Is All You Need](https://llmilestones.com/m/transformer.html): 2017-06-12, Google. Introduced the Transformer, an architecture built entirely on self-attention, dropping RNNs and convolutions.
- [GPT-1: Generative Pre-Training](https://llmilestones.com/m/gpt1.html): 2018-06-11, OpenAI. Pre-train on huge amounts of text by predicting the next word, then fine-tune for specific tasks.
- [BERT](https://llmilestones.com/m/bert.html): 2018-10-11, Google. A bidirectional Transformer encoder that learns language through fill-in-the-blank pre-training.
- [AlphaFold wins CASP13](https://llmilestones.com/m/alphafold.html): 2018-12, DeepMind. Ranked first in the 13th protein-structure prediction contest, CASP13.
- [GPT-2: Too dangerous to release?](https://llmilestones.com/m/gpt2.html): 2019-02-14, OpenAI. More than ten times larger, and able to write coherent long text without fine-tuning.
- [Bengio, Hinton and LeCun win the 2018 Turing Award](https://llmilestones.com/m/turing.html): 2019-03-27, ACM. ACM awarded the 2018 Turing Award to Yoshua Bengio, Geoffrey Hinton and Yann LeCun.
- [T5: Text-to-text for everything](https://llmilestones.com/m/t5.html): 2019-10-23, Google. Recast translation, classification, summarization and every other task as text in, text out.

## Act II: Scaling up (2020 — 2022.10)

Bigger models and more data meant more capability, and it turned out to be predictable.

- [Scaling Laws for Neural Language Models](https://llmilestones.com/m/scaling.html): 2020-01-23, OpenAI. Found that loss falls smoothly, as a power law, with parameters, data and compute.
- [GPT-3: Language Models are Few-Shot Learners](https://llmilestones.com/m/gpt3.html): 2020-05-28, OpenAI. Give it a few examples in the prompt and it can do a new task without retraining.
- [LoRA: Low-Rank Adaptation](https://llmilestones.com/m/lora.html): 2021-06-17, Microsoft. Freeze the big model and train only small low-rank matrices, cutting fine-tuning cost dramatically.
- [GitHub Copilot preview](https://llmilestones.com/m/copilot.html): 2021-06-29, GitHub / OpenAI. Built on the Codex code model, it autocompletes whole blocks of code in the editor.
- [Chain-of-Thought prompting](https://llmilestones.com/m/cot.html): 2022-01-28, Google. Have the model write out its reasoning before answering, and accuracy on math and logic jumps.
- [InstructGPT: alignment with RLHF](https://llmilestones.com/m/instructgpt.html): 2022-03-04, OpenAI. Used reinforcement learning from human feedback (RLHF) to teach the model to follow instructions and give the answers people want.
- [Chinchilla: data matters too](https://llmilestones.com/m/chinchilla.html): 2022-03-29, DeepMind. For the same compute, a smaller model trained on more data beats a bigger one.
- [PaLM](https://llmilestones.com/m/palm.html): 2022-04-04, Google. Google's very large language model, showing abilities such as reasoning and explaining jokes.
- [FlashAttention](https://llmilestones.com/m/flashattention.html): 2022-05-27, Stanford. Reorders the attention computation to minimize GPU memory reads and writes; exact results, much faster.
- [ReAct: reasoning and acting together](https://llmilestones.com/m/react.html): 2022-10-06, Princeton / Google. Interleaves reasoning steps with actions such as searching or calling tools, then keeps thinking based on the results.

## Act III: The ChatGPT moment (2022.11 — 2023)

A chat box showed the world what large models could do, and the race began.

- [ChatGPT launches](https://llmilestones.com/m/chatgpt.html): 2022-11-30, OpenAI. GPT-3.5 plus RLHF, packaged as a free chat box and launched as a "research preview".
- [Constitutional AI: alignment from a set of principles](https://llmilestones.com/m/constitutional.html): 2022-12-15, Anthropic. Instead of labeling every harmful answer by hand, the model critiques and revises itself against a written set of principles.
- [LLaMA: the spark of open-source models](https://llmilestones.com/m/llama.html): 2023-02-24, Meta. A family of 7B to 65B models, released to researchers on application.
- [GPT-4](https://llmilestones.com/m/gpt4.html): 2023-03-14, OpenAI. A big jump in capability; it demonstrated image input and reached near-top human scores on many professional exams.
- [Claude launches](https://llmilestones.com/m/claude1.html): 2023-03-14, Anthropic. Anthropic, founded by former OpenAI staff, released the first Claude on the same day as GPT-4.
- [ERNIE Bot: the "hundred-model war" begins](https://llmilestones.com/m/ernie.html): 2023-03-16, Baidu. The earliest chat-model product from a major Chinese tech company.
- [QLoRA](https://llmilestones.com/m/qlora.html): 2023-05-23, University of Washington. Combines 4-bit quantization with LoRA to fine-tune huge models on a single GPU.
- [DPO: Direct Preference Optimization](https://llmilestones.com/m/dpo.html): 2023-05-29, Stanford. Learns directly from pairs of human preferences, skipping the separate reward model and reinforcement learning of RLHF.
- [Llama 2: now commercially usable](https://llmilestones.com/m/llama2.html): 2023-07-18, Meta. Open weights, free for most companies to use commercially.
- [Qwen goes open source](https://llmilestones.com/m/qwen.html): 2023-08-03, Alibaba. Alibaba open-sourced Qwen-7B, then kept releasing a full range of models from small to large.
- [Mamba: selective state-space models](https://llmilestones.com/m/mamba.html): 2023-12-01, CMU / Princeton. A sequence model whose cost grows linearly with sequence length, matching Transformers of the same size.
- [Gemini 1.0](https://llmilestones.com/m/gemini1.html): 2023-12-06, Google. A natively multimodal model, launched after Google merged Brain and DeepMind.
- [Mixtral 8×7B: MoE takes off in open source](https://llmilestones.com/m/mixtral.html): 2023-12-11, Mistral AI. A mixture-of-experts (MoE) model: many parameters in total, only some of them active each time.

## Act IV: Multimodal, open source catches up (2024)

Models learned to see, hear and speak; open models closed in on the frontier.

- [Gemini 1.5: a million-token context](https://llmilestones.com/m/gemini15.html): 2024-02-15, Google. Reads about 700,000 words, about an hour of video, or over 30,000 lines of code at once.
- [Claude 3 family](https://llmilestones.com/m/claude3.html): 2024-03-04, Anthropic. Three tiers, Haiku, Sonnet and Opus; Opus beat GPT-4 on many benchmarks.
- [Llama 3 / 3.1 405B](https://llmilestones.com/m/llama3.html): 2024-04-18, Meta. 8B and 70B models in April; the 405B Llama 3.1 in July.
- [GPT-4o: natively multimodal](https://llmilestones.com/m/gpt4o.html): 2024-05-13, OpenAI. One model handles text, images and voice, with real-time spoken conversation, interruptions and emotion.
- [Scaling Monosemanticity](https://llmilestones.com/m/monosemanticity.html): 2024-05-21, Anthropic. Extracted millions of interpretable features from a production model, Claude 3 Sonnet, such as one for the Golden Gate Bridge.
- [Scaling test-time compute optimally](https://llmilestones.com/m/ttc.html): 2024-08-06, Google DeepMind / UC Berkeley. Showed that spending compute wisely at inference time can beat a much larger model.
- [o1: the year of reasoning models](https://llmilestones.com/m/o1.html): 2024-09-12, OpenAI. Trained with reinforcement learning to think at length before answering.
- [Hopfield and Hinton win the Nobel Prize in Physics](https://llmilestones.com/m/nobel.html): 2024-10-08, Nobel Committee. The 2024 Nobel Prize in Physics went to John Hopfield and Geoffrey Hinton for foundational work on machine learning with artificial neural networks.
- [Computer Use: AI operates a computer](https://llmilestones.com/m/computeruse.html): 2024-10-22, Anthropic. The model uses a computer by looking at screenshots, moving the mouse and typing.
- [MCP: Model Context Protocol](https://llmilestones.com/m/mcp.html): 2024-11-25, Anthropic. An open protocol that lets models connect to tools and data sources in a uniform way.
- [DeepSeek-V3](https://llmilestones.com/m/dsv3.html): 2024-12-26, DeepSeek. A 671B-parameter MoE model close to GPT-4o, with fully open weights.

## Act V: Reasoning and agents (2025)

First think, then answer, then go and do the work.

- [DeepSeek-R1: an open reasoning model](https://llmilestones.com/m/r1.html): 2025-01-20, DeepSeek. Open-sourced under MIT, with reasoning on par with o1 and its full thinking visible.
- [Claude 3.7 Sonnet and Claude Code](https://llmilestones.com/m/claudecode.html): 2025-02-24, Anthropic. A hybrid reasoning model, plus a coding agent that reads, edits and runs code right in the terminal.
- [Manus: a general AI agent](https://llmilestones.com/m/manus.html): 2025-03-06, Butterfly Effect. Plans, browses and writes code on its own inside a cloud virtual machine, delivering a finished task rather than just an answer.
- [Gemini 2.5 Pro](https://llmilestones.com/m/gemini25.html): 2025-03-25, Google. Gemini with built-in thinking; it topped LMArena and other leaderboards at launch.
- [Llama 4](https://llmilestones.com/m/llama4.html): 2025-04-05, Meta. Meta's first MoE Llama, released as Scout and Maverick.
- [Qwen3](https://llmilestones.com/m/qwen3.html): 2025-04-29, Alibaba. One model can switch between thinking and non-thinking modes; open-sourced under Apache 2.0.
- [Claude 4 (Opus 4 / Sonnet 4)](https://llmilestones.com/m/claude4.html): 2025-05-22, Anthropic. Focused on coding and on agents that work continuously for long stretches.
- [Kimi K2](https://llmilestones.com/m/kimik2.html): 2025-07-11, Moonshot AI. A trillion-parameter open-weight MoE model (modified MIT license), built for agents and tool use.
- [gpt-oss: OpenAI opens weights again](https://llmilestones.com/m/gptoss.html): 2025-08-05, OpenAI. Two open-weight reasoning models, 120B and 20B, under Apache 2.0.
- [GPT-5](https://llmilestones.com/m/gpt5.html): 2025-08-07, OpenAI. Unifies quick answers and deep reasoning in one system that decides when to think.
- [DeepSeek-R1 published in Nature](https://llmilestones.com/m/r1nature.html): 2025-09-17, DeepSeek. The R1 paper was peer-reviewed and published in Nature.
- [Gemini 3](https://llmilestones.com/m/gemini3.html): 2025-11-18, Google. Google's next-generation flagship, with the top scores at the time on many reasoning and multimodal benchmarks.
- [Claude Opus 4.5](https://llmilestones.com/m/opus45.html): 2025-11-24, Anthropic. Further gains in coding and agentic tasks, with a big price cut.

## Act VI: The frontier and its limits (2026)

Models so capable they had to be locked up first: capability, safety and openness all pushed to the edge.

- [OpenClaw: an open-source personal agent](https://llmilestones.com/m/openclaw.html): 2026-01-30, Peter Steinberger. An open-source agent that runs on your own computer or server, takes tasks through WhatsApp, Telegram, Slack and other chat apps, and works with many models.
- [Qwen3.5](https://llmilestones.com/m/qwen35.html): 2026-02-16, Alibaba. Open-sourced a 397B-parameter MoE model that activates only 17B parameters per inference.
- [Gemini 3.1 Pro](https://llmilestones.com/m/gemini31.html): 2026-02-19, Google. Google released Gemini 3.1 Pro in preview, focused on complex reasoning.
- [Claude Mythos Preview and Project Glasswing](https://llmilestones.com/m/mythospreview.html): 2026-04-07, Anthropic. A model exceptionally good at finding and exploiting software vulnerabilities, offered only to a few organizations through Project Glasswing to harden critical software.
- [Muse Spark](https://llmilestones.com/m/musespark.html): 2026-04-08, Meta. The first model from Meta Superintelligence Labs: natively multimodal, with no open weights.
- [GPT-5.5](https://llmilestones.com/m/gpt55.html): 2026-04-23, OpenAI. OpenAI released GPT-5.5 (codename Spud), improving agentic coding and research ability.
- [DeepSeek-V4](https://llmilestones.com/m/dsv4.html): 2026-04-24, DeepSeek. Released two open-weight models, V4-Pro and V4-Flash, with a 1M-token context.
- [Claude Fable 5 and Mythos 5](https://llmilestones.com/m/fable5.html): 2026-06-09, Anthropic. The same model as Mythos 5, released to the public with safeguards added; Anthropic's top-tier model.
- [Claude Opus 5](https://llmilestones.com/m/opus5.html): 2026-07-24, Anthropic. Performance close to Fable 5 at half its API price.
- [Kimi K3](https://llmilestones.com/m/kimik3.html): 2026-07-27, Moonshot AI. 2.8T parameters, native vision and a 1M-token context; weights released on July 27.
- [GPT-6 (Astra / Sol / Luna)](https://llmilestones.com/m/gpt6.html): 2026-09-03, OpenAI. On September 3 OpenAI launched its strongest model, GPT-6 Astra, to a small group of organizations, opening a restricted version to paid users the next day; the smaller Sol and Luna followed on…
- [Claude Opus 5.5](https://llmilestones.com/m/opus55.html): 2026-09-22, Anthropic. The first of the Claude 5.5 family, on par with Fable 5.1 on most tasks.
- [Gemini 4 Argon](https://llmilestones.com/m/gemini4.html): 2026-10-13, Google. The first Gemini 4 model, launched in Antigravity and the Gemini API.
- [Muse Spark Open](https://llmilestones.com/m/musesparkopen.html): 2026-10-15, Meta. Released weights for an 816B-parameter MoE model (about 42B active) that takes text, image, video and audio input.