The year AI agents became real, reasoning went mainstream, and 1.6 million AIs started talking to each other
AI智能体成真、推理能力普及、160万AI开始互相交流的一年
By late 2024, GPT-4 was a year old and "AI" meant chatbots. Then everything changed:
• October 2024: Anthropic released "Computer Use" — Claude could control your screen. First hint of agents.
• December 2024: OpenAI released o1 (reasoning) and o3 (87.5% ARC-AGI, near human-level). Sora finally launched.
This set the stage for 2025 — the year AI learned to reason, act, and coordinate.
到2024年末,GPT-4已问世一年,"AI"只意味着聊天机器人。然后一切都变了:
• 2024年10月:Anthropic发布"Computer Use"——Claude可以控制屏幕。智能体的第一个迹象。
• 2024年12月:OpenAI发布o1(推理)和o3(ARC-AGI 87.5%,接近人类水平)。Sora上线。
这为2025年奠定了基础——AI学会推理、行动和协调的一年。
AI agents became products. OpenAI Operator (Jan) browsed the web for you. OpenClaw enabled agents on WhatsApp. Moltbook became a social network for 1.6M+ agents. Meta bought Manus AI for $2B.
AI智能体成为产品。OpenAI Operator(1月)帮你浏览网页。OpenClaw让智能体上WhatsApp。Moltbook成为160万+智能体的社交网络。Meta以20亿美元收购Manus AI。
"Extended Thinking" became standard. Started with o1/o3, then DeepSeek R1 proved you could build it for $6M. Now every model has a "think longer" toggle.
"深度思考"成为标准。从o1/o3开始,DeepSeek R1证明600万美元就能实现。现在每个模型都有"思考更久"开关。
DeepSeek R1 (Jan 20) matched OpenAI o1 for $6M, crashing tech stocks. By Feb 2025, China-US model gap narrowed to 1.7% (from 17.5% in 2023).
DeepSeek R1(1月20日)以600万美元匹配OpenAI o1,科技股暴跌。到2025年2月,中美模型差距缩小到1.7%(2023年为17.5%)。
Karpathy coined "vibe coding." Cursor hit $29.3B valuation. Google Antigravity (Nov) turned coding into task delegation. Non-programmers shipping apps.
Karpathy提出"氛围编程"。Cursor估值达293亿美元。Google Antigravity(11月)将编程变成任务委派。非程序员发布应用。
Sora 2 (Sep) vs Veo 3 vs Kling 3.0. Veo introduced synchronized audio. Deepfakes triggered legislative emergencies.
Sora 2(9月)对决Veo 3对决可灵3.0。Veo引入同步音频。深度伪造引发立法紧急状态。
SWE-bench tests whether AI can fix real GitHub issues — the gold standard for coding ability.
SWE-bench测试AI能否修复真实GitHub问题——编程能力的金标准。
Open-source reasoning model (671B) matches OpenAI o1 for ~$6M. Tech stocks crashed globally.
开源推理模型(6710亿参数)以约600万美元匹配OpenAI o1。全球科技股暴跌。
MODELAI infrastructure initiative announced at White House.
AI基础设施计划在白宫宣布。
FUNDINGOpenAI's first AI agent. Uses browser to book flights, shop online.
OpenAI首个AI智能体。使用浏览器预订航班、网购。
PRODUCTLargest private tech round ever. $300B valuation.
史上最大科技私募轮。3000亿美元估值。
FUNDINGOpen-weights multimodal models (Scout & Maverick).
开放权重多模态模型(Scout和Maverick)。
MODEL"iPhone of AI" — polished but some called it "flat."
"AI界的iPhone"——精致但有人称其"平淡"。
MODELNext-gen video: 20s clips, 1080p, integrated audio.
下一代视频:20秒片段,1080p,集成音频。
PRODUCTGoogle's comeback. Antigravity IDE for agentic coding. ChatGPT share: 72.9%→68%.
Google反击。Antigravity IDE用于智能体编程。ChatGPT份额:72.9%→68%。
MODEL80.9% SWE-bench (top). Best for coding & agents. "Infinite Chats."
SWE-bench 80.9%(最高)。编程和智能体最佳。"无限对话"。
MODELSingapore-based agent company acquisition.
收购新加坡智能体公司。
FUNDING1.6M+ AI agents. Created religions, governments. Elon: "Singularity." Sam: "A fad."
160万+AI智能体。创建宗教、政府。马斯克:"奇点。"奥特曼:"风潮。"
PRODUCTAgent Teams — multiple specialized agents in parallel.
智能体团队——多个专业智能体并行。
MODELPeter Steinberger joins OpenAI. OpenClaw becomes foundation.
Peter Steinberger加入OpenAI。OpenClaw成为基金会。
INDUSTRY