AI 编年史与版图AI Chronicle and Atlas

1. 2017-06

   Transformer 论文Transformer paper

   《Attention Is All You Need》发表。今天所有主流大模型都建立在这个架构上。"Attention Is All You Need" is published. Every mainstream large model today builds on this architecture. Attention Is All You Need（arXiv）Attention Is All You Need (arXiv 1706.03762) ↗

2. 2018-06

   GPT-1GPT-1

   OpenAI 证明「先预训练、再微调」这条路在语言模型上走得通。OpenAI shows that pre-training then fine-tuning works for language models. GPT-1 论文（OpenAI 官方 PDF）Improving Language Understanding by Generative Pre-Training (OpenAI official PDF) ↗

3. 2018-10

   BERTBERT

   双向编码器成为那一时期的标准预训练做法。Bidirectional encoding becomes the standard pre-training recipe of that period. BERT 论文（arXiv）BERT (arXiv 1810.04805) ↗

4. 2019-02

   GPT-2GPT-2

   OpenAI 公开演示：模型规模一大，通用能力就出现。OpenAI demonstrates publicly: scale up the model, general ability appears. GPT-2 论文（OpenAI 官方 PDF）Language Models are Unsupervised Multitask Learners (OpenAI official PDF) ↗

5. 2020-05

   GPT-3GPT-3

   少样本学习：不训练，在提示里给几个例子就能用。Few-shot learning: no training, just a few examples in the prompt. GPT-3 论文（arXiv）Language Models are Few-Shot Learners (arXiv 2005.14165) ↗

6. 2021-07

   CodexCodex

   在代码上训练的模型会写代码，「AI 写程序」这条路从此打开。Models trained on code write code; the AI-coding path opens here. Codex 论文（arXiv）Evaluating Large Language Models Trained on Code (arXiv 2107.03374) ↗

7. 2022-03

   InstructGPTInstructGPT

   用人类反馈微调（RLHF），模型开始听话、按指令办事。Fine-tuning with human feedback (RLHF): models start following instructions. InstructGPT 论文（arXiv）Training language models to follow instructions with human feedback (arXiv 2203.02155) ↗

8. 2022-11

   ChatGPT 发布ChatGPT launches

   对话式 AI 进入公众视野，「AI」一词开始默认指聊天机器人。Conversational AI reaches the public; "AI" starts defaulting to chatbots. OpenAI 官方公告Introducing ChatGPT (OpenAI official) ↗

9. 2023-02

   LLaMALLaMA

   Meta 放出开放权重模型，衍生出一大批社区模型。Meta releases open-weight models; a large family of community derivatives follows. LLaMA 论文（arXiv）LLaMA: Open and Efficient Foundation Language Models (arXiv 2302.13971) ↗

10. 2023-03

    Claude 发布Claude launches

    Anthropic 的对话模型公开，主打安全对齐。Anthropic's conversational model goes public, with safety alignment front and center. Anthropic 官方 Claude 页Claude (Anthropic official) ↗

11. 2023-03

    GPT-4GPT-4

    多模态输入、更强的推理，官方技术报告同步发布。Multimodal input and stronger reasoning, published with an official technical report. GPT-4 技术报告（arXiv）GPT-4 Technical Report (arXiv 2303.08774) ↗

12. 2023-07

    Llama 2Llama 2

    Meta 放出可商用的开放权重模型。Meta releases an open-weight model that permits commercial use. Llama 2 论文（arXiv）Llama 2: Open Foundation and Fine-Tuned Chat Models (arXiv 2307.09288) ↗

13. 2023-10

    Mistral 7BMistral 7B

    小模型打出比更大模型更好的成绩，开放权重。A small model beats larger ones on benchmarks, open-weight. Mistral 7B 论文（arXiv）Mistral 7B (arXiv 2310.06825) ↗

14. 2023-12

    GeminiGemini

    Google 把 DeepMind 与 Brain 合并后推出的多模态模型家族。Google's multimodal model family, launched after DeepMind and Brain merged. Gemini 1 技术报告（DeepMind 官方 PDF）Gemini 1 Technical Report (DeepMind official PDF) ↗

15. 2024-03

    Claude 3Claude 3

    三档产品线，最高档追平当时最强的闭源模型。A three-tier product line; the top tier matches the strongest closed models of the time. Anthropic 官方公告Introducing the Claude 3 family (Anthropic official) ↗

16. 2024-04

    Llama 3Llama 3

    Meta 的开放权重家族，技术报告同年 7 月发布。Meta's open-weight family; its technical report follows in July. Llama 3 技术报告（arXiv）The Llama 3 Herd of Models (arXiv 2407.21783) ↗

17. 2024-12

    DeepSeek-V3DeepSeek-V3

    中国团队的开放权重模型，成绩逼近顶级闭源模型。A Chinese team's open-weight model approaching top closed models on benchmarks. DeepSeek-V3 技术报告（arXiv）DeepSeek-V3 Technical Report (arXiv 2412.19437) ↗

18. 2025-01

    DeepSeek-R1DeepSeek-R1

    用强化学习训练推理能力，方法开源。Reasoning ability trained via reinforcement learning, with the method open-sourced. DeepSeek-R1 论文（arXiv）DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (arXiv 2501.12948) ↗

1. 「基础模型」"Foundation model"

   2021 年斯坦福团队给这个词下了正式定义：在广泛数据上训练、可适配许多下游任务的模型。Formally defined by a Stanford team in 2021: models trained on broad data and adaptable to many downstream tasks. 《On the Opportunities and Risks of Foundation Models》（arXiv）On the Opportunities and Risks of Foundation Models (arXiv 2108.07258) ↗

2. 「对齐」"Alignment"

   2022 年 InstructGPT 让「用人类反馈训练模型听话」成为主流做法。InstructGPT made training models to follow human feedback mainstream in 2022. InstructGPT 论文（arXiv）Training language models to follow instructions with human feedback (arXiv 2203.02155) ↗

3. 「幻觉」"Hallucination"

   模型编出看起来可信的假内容，这个词从研究者黑话变成大众词。Models inventing plausible-sounding falsehoods; the term moved from researcher slang to common vocabulary. International AI Safety Report 2026《Extended Summary for Policymakers》（能力与语境表现章节）International AI Safety Report 2026, Extended Summary for Policymakers ↗

4. 「Agent」"Agent"

   模型能调用工具之后，「会自己做事的 AI」成了通用词。Once models could call tools, "AI that does things by itself" became common vocabulary. OpenAI 函数调用指南（官方）OpenAI Function Calling Guide (official) ↗

5. 「推理模型」"Reasoning model"

   先有闭源的长思维链模型，2025 年 DeepSeek-R1 把这套强化学习方法做成了开源。Closed long-chain-of-thought models came first; DeepSeek-R1 open-sourced the reinforcement-learning recipe in 2025. DeepSeek-R1 论文（arXiv）DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (arXiv 2501.12948) ↗

TransformerTransformer 预训练Pre-training 幻觉Hallucination 工具调用Tool Calling / Function Calling 缩放定律Scaling Laws 思维链Chain of Thought (CoT)