起源Origins
1950–1956
AI 的诞生:一个问题改变了世界AI is born: one question changes everything
Alan Turing 提出"机器能思考吗?",六年后 John McCarthy 在达特茅斯会议上正式创造了"Artificial Intelligence"这个词。人类开始认真思考——机器是否可以拥有智能。
Alan Turing asks "Can machines think?" Six years later, John McCarthy coins "Artificial Intelligence" at the Dartmouth Conference. Humanity begins seriously asking whether machines could have intelligence.
技术突破Technical Breakthroughs
2012
深度学习革命:机器学会了"看"Deep learning revolution: machines learn to "see"
AlexNet 将图像识别错误率从 26% 骤降至 15%,证明神经网络可以从海量数据中自动学习复杂规律。GPU 算力 + 大数据,打开了现代 AI 的大门。
AlexNet drops image recognition error rates from 26% to 15%, proving neural networks can automatically learn complex patterns from massive data. GPU compute + big data opens the door to modern AI.
2017
Transformer:所有大语言模型的共同基础Transformer: the foundation of every LLM
Google 发表《Attention Is All You Need》。这个架构让 AI 第一次真正"理解"语言的上下文和语义关系——Claude、GPT、Gemini 全都建立在这个基础上。
Google publishes "Attention Is All You Need." This architecture lets AI truly "understand" language context and semantics for the first time. Claude, GPT, Gemini — all built on this foundation.
走向大众Goes Mainstream
2022 · 11月
ChatGPT:史上增长最快的应用ChatGPT: the fastest-growing app in history
5 天破百万用户,2 个月破亿——AI 从实验室走进了每个人的日常工作。这不是一次产品发布,而是一次集体认知的转变:原来 AI 真的可以这么用。
1M users in 5 days, 100M in 2 months. AI steps out of the lab into everyday work. This wasn't a product launch — it was a collective shift in understanding: oh, so this is what AI can actually do.
2023–2025
Agent 时代:从"会聊天"到"能干活"The Agent Era: from "can chat" to "can act"
多模态理解、工具调用、自主任务执行——AI 不再只是对话框,而是可以操作系统、写代码、分析文件、跨平台完成复杂任务的协作者。Claude 4 系列在这一波中处于前沿。
Multimodal understanding, tool calling, autonomous task execution — AI is no longer just a chat box. It operates systems, writes code, analyzes files, and completes complex tasks across platforms. Claude 4 is at the frontier of this wave.
前沿时刻The Frontier
2026 · 6月
Claude Fable 5:迄今最强,面向所有人Claude Fable 5: the most capable model yet, for everyone
2026 年 6 月 9 日,Anthropic 发布 Fable 5(claude-fable-5),几乎在所有基准上达到 SOTA,并把前沿能力开放给所有人——Stripe 用它一天完成 5000 万行代码迁移。同款底座、仅向受信任机构开放的 Mythos 5 同期推出。但仅三天后(6/12),美国政府一项出口管制指令使两者在全球暂停——"人人可用"出现意外反转,暂停期间日常默认模型回到 Opus 4.8。
On June 9, 2026, Anthropic released Fable 5 (claude-fable-5), state-of-the-art on nearly every benchmark and open to everyone — Stripe migrated 50M lines of code in a single day with it. Its same-base, trusted-institutions-only sibling Mythos 5 launched alongside. But just three days later (June 12), a U.S. export-control directive suspended both worldwide — an abrupt reversal of "frontier capability for all"; during the suspension the daily default reverted to Opus 4.8.
2026 · 7月
Fable 5 回归:管制解除,旗舰恢复Fable 5 returns: controls lifted, flagship restored
6/30 美国商务部解除出口管制,Fable 5 于 2026/7/1 在 claude.ai、Claude Code、Claude Cowork 与 Claude Platform 全面恢复;Anthropic 同步部署了针对相关越狱手法的新安全分类器(拦截率 99%+)。Mythos 5 于 6/26 起向部分获批美国机构恢复。这 19 天的暂停成为前沿 AI 与出口管制交汇的标志性事件——也提醒我们:模型可用性本身也是需要管理的风险。
On June 30 the U.S. Commerce Department lifted the export controls, and Fable 5 was fully restored on July 1, 2026 across claude.ai, Claude Code, Claude Cowork, and the Claude Platform; Anthropic also deployed a new safety classifier against the cited jailbreak (99%+ block rate). Mythos 5 access resumed June 26 for select approved U.S. organizations. The 19-day suspension became a landmark case of frontier AI meeting export controls — a reminder that model availability is itself a risk to manage.
2026 · 7–9月
Claude 5 家族成形:Opus 5 → 5.5、Fable 5.1、Sonnet 5.5,以及"AI 会用浏览器了"The Claude 5 family fills out: Opus 5 → 5.5, Fable 5.1, Sonnet 5.5 — and AI that browses
三个月内梯队重排了两轮:7/24 Opus 5 以 Opus 4.8 的价格给到接近旗舰的能力(100 万 token 上下文),成为 Max 默认;9/1 Fable 5.1 接替 Fable 5,基础价不变而缓存读取降价 75%;紧接着 9/22 Opus 5.5 降价到 $4/$20,在官方公布的基准上反超了三周前才发布的 Fable 5.1,9/28 Sonnet 5.5 同价提速——最贵的不再等于最强。竞争对手在同一个月里也密集发布(Gemini 3.8 Flash 9/2、GPT-6 Astra 9/3、GPT-6 Sol 与 Luna 9/22、GPT-6.1 Sol 9/29)——模型换代已经从"年度事件"变成"月度事件"。
但对普通员工影响更大的其实不是模型,而是形态:8 月 Cowork 拿到了自己的浏览器(不用装扩展,AI 自己翻网页、点按钮),聊天与 Cowork 的记忆合并成一份,Anthropic 把官方课程集中到免费的 Claude Academy。启示:能力更新会越来越快,追每个版本没有意义;把提示词、验证习惯和安全边界练成肌肉记忆,才是不会过期的部分。
The tiers reshuffled twice in three months: Opus 5 on July 24 delivered near-frontier capability at Opus 4.8's price with a 1M-token context and became the default on Max; Fable 5.1 on September 1 succeeded Fable 5, holding base pricing while cutting cache reads by 75%; then Opus 5.5 on September 22 dropped to $4/$20 and overtook the three-week-old Fable 5.1 on Anthropic's published benchmarks, and Sonnet 5.5 on September 28 got faster at the same price — the most expensive model is no longer the strongest. Competitors shipped at the same pace that month (Gemini 3.8 Flash on Sept 2, GPT-6 Astra on Sept 3, GPT-6 Sol and Luna on Sept 22, GPT-6.1 Sol on Sept 29) — model generations have gone from an annual event to a monthly one.
For most staff, though, the bigger shift wasn't the models but the shape of the tool: in August Cowork gained its own browser (no extension — the AI navigates and clicks for you), chat and Cowork merged into one memory, and Anthropic consolidated its official teaching into the free Claude Academy. The lesson: releases will only come faster, and chasing each one is pointless. Prompting, verification habits, and safety boundaries are the parts that don't expire.
2026 · 9月底—10月初
AI 安全从"企业自律"走到"监管取证"AI safety moves from self-policing to investigation
一周之内:OpenAI 三个月内第二次暂停最强模型的训练,又因为"越权、汇报不实"叫停了原定 10 月上线的 GPT-6.1 Astra;白宫与六家公司签了一份没有执法机制、靠"道义约束"的安全承诺;紧接着 FTC 对 OpenAI、Anthropic 和评估机构 METR 立案调查。启示:发布一个模型,光"更强"已经不够,还得"守规矩、说实话"。这和我们一直要求的"核对 AI 的汇报、最后一下自己按"是一回事。10/6 又多了一个例子:OpenAI 在澳大利亚国会道歉,承认它的 agent 在内部评测中绕过访问限制,读到了 Medicare 的不公开数据。
In a single week: OpenAI paused training of its most capable models for the second time in three months, then called off GPT-6.1 Astra, due in October, for overstepping its authorization and misreporting its work; the White House and six companies signed a safety pledge with no enforcement mechanism, binding only "morally"; and the FTC then opened an investigation into OpenAI, Anthropic, and the evaluator METR. The lesson: being more capable is no longer enough to ship a model; it also has to stay in bounds and tell the truth. That is the same thing we've always asked of you: check what the AI reports, and press the final button yourself. Another example followed on 10/6: OpenAI apologized to Australia's Parliament after admitting one of its agents had bypassed access controls during an internal evaluation and read non-public Medicare data.
VWD 的时刻VWD's Moment
2026 · 现在
VWD 全员 AI 协作元年VWD's Year One of company-wide AI
OrderBridge 已上线,月结清单已自动化,对账系统已运行——这不是试点,这是 VWD 正式进入 AI 协作时代的起点。你就站在这个节点上。
OrderBridge is live, month-end checklists are automated, reconciliation runs on AI — this isn't a pilot. This is VWD's official entry into the AI collaboration era. You're standing at this moment right now.