从最新热点回看早期基础，这里按时间倒序整理 AI 发展中的关键研究、热门模型、产品拐点与行业事件，并尽量优先采用官方或论文来源。
优先保留官方公告、官方博客、arXiv 或高可信研究来源。
重点标记公众关注度高、产业影响大的关键时刻。
只保留更主流、更有代表性、可公开核验的事件。
如果你第一次看这条时间线，可以先从这些高影响事件开始。
Anthropic released Claude 3.7 Sonnet, a hybrid reasoning model designed to better balance fast responses and deeper deliberation.
DeepSeek released DeepSeek-R1, a reasoning-focused model family that gained major attention for strong performance and open availability.
DeepSeek released DeepSeek-V3, an open-weight mixture-of-experts model that drew broad attention for its capability and efficiency.
Google introduced Gemini 2.0, positioning it as a new generation of models built for the agentic era.
Alibaba's Qwen team released Qwen2.5, a highly capable open model family that became very popular among developers and local-model users.
OpenAI announced o1, a new reasoning-focused model family designed to spend more time thinking before responding.
Anthropic released Claude 3.5 Sonnet, improving reasoning and coding while introducing the Artifacts workflow in Claude.
The Council of the European Union gave final approval to the EU AI Act, the world's first broad horizontal AI regulation.
OpenAI released GPT-4o, a natively multimodal model for text, vision, and voice with much lower latency.
Meta released the first Llama 3 models, substantially improving open-weight LLM quality and usability.
xAI released Grok-1 weights, drawing broad attention because of the model's scale and xAI's public profile.
Cognition introduced Devin as an AI software engineer product focused on multi-step coding workflows, drawing widespread attention across the tech industry.
Anthropic released the Claude 3 family, including Haiku, Sonnet, and Opus, with strong reasoning and vision capabilities.
Google introduced Gemini 1.5 Pro with a breakthrough long-context window reaching up to 1 million tokens.
OpenAI unveiled Sora, a text-to-video model capable of generating strikingly realistic and cinematic video clips.
Mistral AI released Mixtral 8x7B, a sparse mixture-of-experts model that became one of the hottest open models of its time.
Google opened NotebookLM more broadly, turning its source-grounded AI notebook into one of the most talked-about AI productivity products.
Google launched Gemini 1.0, its flagship multimodal model family spanning Nano, Pro, and Ultra.
OpenAI's board abruptly removed Sam Altman as CEO, triggering one of the most dramatic corporate crises in tech before he returned days later.
At its first DevDay, OpenAI launched GPTs, the Assistants API, and major ChatGPT platform upgrades for developers and consumers.
Mistral AI released Mistral 7B, a compact open-weight model that quickly gained popularity for strong efficiency and performance.
OpenAI announced DALL·E 3 with major prompt-following improvements and deep ChatGPT integration.
Meta released Llama 2 for research and commercial use, making a high-profile open-weight LLM widely available.
Anthropic released Claude 2 with stronger reasoning and a much larger context window for long documents.
OpenAI released GPT-4, a more capable multimodal model with major gains in reasoning, coding, and reliability.
Anthropic announced Claude, introducing its Constitutional AI approach to helpful and safer chat systems.
Meta introduced LLaMA, a family of research language models that strongly influenced the open LLM ecosystem.
OpenAI released ChatGPT as a research preview, bringing a GPT-3.5-based conversational assistant to the public.
Stability AI and collaborators released Stable Diffusion, a widely accessible open image generation model.
OpenAI announced DALL·E 2, a major leap in text-to-image generation with more realistic and coherent visual outputs.
GitHub and OpenAI launched GitHub Copilot, bringing AI code completion to mainstream developers.
DeepMind's AlphaFold 2 achieved a major breakthrough in protein structure prediction at CASP14.
OpenAI introduced GPT-3, a 175B-parameter language model with strong few-shot and zero-shot performance across many tasks.
OpenAI announced GPT-2, a much larger language model with strong zero-shot generation ability.
Google introduced BERT, a bidirectional language representation model that quickly became a standard benchmark and production NLP backbone.
OpenAI introduced GPT, showing that generative pretraining plus task-specific fine-tuning could produce strong language understanding results.
Google researchers introduced the Transformer architecture, replacing recurrence with attention mechanisms for sequence modeling.