从图灵试想到代理时代 —— 七十五年的思想简史。
From Turing's imitation game to the agentic era — a brief intellectual history, seventy-five years in the making.
人工智能不是 ChatGPT 时代才有的东西。早在 1950 年，图灵就把"机器能否思考"写进了论文的第一行。此后七十余年，这个领域至少经历了三次大起大落 —— 从符号主义的乐观，到专家系统的泡沫，到连接主义的蛰伏，再到深度学习的爆发。
Artificial intelligence did not begin with ChatGPT. Its arc — from Turing's 1950 imitation game through symbolic logic, expert-system booms, and two AI winters — sets the stage for everything that came after.
图灵在 Mind 杂志发表《计算机器与智能》，提出"模仿游戏"作为智能的操作性判据，回避了"思考"本身的定义难题。这篇论文成为整个领域的元文本。
Alan Turing publishes "Computing Machinery and Intelligence" in Mind, proposing the Imitation Game as an operational test — deliberately sidestepping the definitional problem of thought itself.
McCarthy、Minsky、Shannon、Rochester 在达特茅斯组织为期两个月的研讨会，"Artificial Intelligence"一词在提案中首次正式出现。主流范式：符号主义 + 推理搜索。
McCarthy, Minsky, Shannon and Rochester convene the Dartmouth workshop — the term "Artificial Intelligence" enters the vocabulary. The initial paradigm: symbolic reasoning and search.
Minsky 与 Papert 出版《感知机》，证明单层感知机无法解决 XOR 这类线性不可分问题。连接主义因此沉寂近二十年，神经网络研究资金几乎被切断。
Minsky and Papert's Perceptrons proves single-layer networks cannot represent XOR. Connectionism is effectively frozen out of mainstream research for nearly two decades.
Lighthill 报告指出 AI 承诺远超交付，英国政府大幅削减资金；DARPA 亦削减机器翻译与语音理解预算。领域进入第一个寒冬。
The Lighthill Report's critique of overpromising triggers UK funding cuts; DARPA follows suit. The first AI winter begins.
Rumelhart、Hinton 与 Williams 在 Nature 发表反向传播算法的系统性阐述，多层网络的训练成为可能。连接主义复兴的数学基础就此奠定。
Rumelhart, Hinton and Williams publish backpropagation in Nature. Multi-layer networks become trainable — the mathematical foundation of the connectionist revival is laid.
IBM 深蓝以搜索 + 手工启发式击败国际象棋世界冠军卡斯帕罗夫。这是工程胜利，而非智能胜利 —— 但它证明了算力的可扩展性。
IBM's Deep Blue defeats Kasparov. Not a triumph of intelligence but of engineering — and a demonstration that compute scales.
Hinton 等人用逐层贪心预训练解决了深度网络的梯度消失问题，"Deep Learning"一词正式登场。连接主义换名再战。
Hinton's layer-wise pretraining sidesteps vanishing gradients. "Deep learning" enters the lexicon — connectionism returns under a new name.
AlexNet 在 ILSVRC 比赛中以 top-5 误差率领先第二名 10 个百分点，GPU 训练 + 卷积网络 + ReLU + Dropout 的组合开启了深度学习的工业化时代。
AlexNet wins ILSVRC by a ten-point margin. GPU training meets convnets, ReLU, and dropout — industrial deep learning begins.
Google 提出 Transformer 架构，用自注意力取代 RNN 的时序依赖。这个架构将在此后近十年里定义整个自然语言处理、计算机视觉、语音、乃至结构生物学的基础模型。
Google publishes the Transformer. Self-attention replaces recurrence, and one architecture comes to dominate language, vision, speech, and structural biology.
2020 年以后的人工智能领域由少数几个关键词支配：规模、对齐、多模态、工具使用、推理时计算。每一个词背后都对应着架构、数据与训练范式上的具体决策，而不是营销术语。
Post-2020 AI is governed by a small set of organising ideas: scale, alignment, multimodality, tool use, and inference-time compute. Each maps to concrete choices in architecture, data, and training recipe — not marketing.
自注意力让每个 token 与序列中任意其他 token 建立直接关联，消除了 RNN 的序列瓶颈，也天然并行化。它从 NLP 扩散到视觉（ViT）、语音、蛋白质折叠（AlphaFold 2），乃至机器人策略。
Self-attention lets any token reference any other, eliminating RNN bottlenecks and parallelising natively. It has since spread from NLP to vision (ViT), speech, protein folding (AlphaFold 2), and robotic policy.
Kaplan (2020) 与 Chinchilla (2022) 证明，语言模型的测试损失随参数、数据、计算呈可预测的幂律下降。这把"要不要继续加算力"从哲学问题变成了工程问题。
Kaplan and Chinchilla show that test loss falls as a power law in parameters, tokens, and compute. Scaling stops being a philosophical bet and becomes an engineering curve.
InstructGPT (2022) 将基于人类反馈的强化学习引入大语言模型的后训练阶段，使模型从"能生成"变为"按意图生成"。此后 DPO、RLAIF、宪法 AI 等变体层出不穷，对齐研究成为显学。
InstructGPT adapts reinforcement learning from human feedback for LLM post-training, shifting models from "capable" to "directable". DPO, RLAIF, and constitutional AI follow.
GPT-4V、Gemini、Claude 等将视觉、音频、视频与文本置于共享的 token 空间。模型不再是"文本模型附带图像理解"，而是"多模态原生"。视频生成（Sora、Veo）进一步把时序也拉入同一表示。
GPT-4V, Gemini, and Claude place vision, audio, and video in a shared token space. Models are no longer "text plus bolt-on vision" but multimodal-native — video generation pulls time into the same representation.
工具调用（function calling）、计算机操作（computer use）、代码执行与 MCP 协议把 LLM 从"对话框内的回答者"变成"能在系统里动手的执行者"。Agent 不是新架构，而是推理循环 + 工具 + 记忆的编排。
Function calling, computer use, code execution, and MCP promote LLMs from chat respondents to in-system actors. Agents are not a new architecture but an orchestration: reasoning loop + tools + memory.
OpenAI o1 / o3、DeepSeek R1 证明：在固定参数量下，让模型"多想几步"（长链思考、树搜索、自我验证）可以线性换取正确率。训练时 scaling law 之后，推理时 scaling law 成为新前沿。
o1 / o3 and DeepSeek R1 show that at fixed parameters, letting a model "think longer" — chain-of-thought, tree search, self-verification — buys accuracy. After training-time scaling, inference-time scaling is the new frontier.
关于未来，最有价值的不是预测而是把开放问题列清楚。下列八个问题，是 2026 年这个时间点上学界与工业界真正在分歧的地方。答案落在哪一侧，会决定接下来十年的形态。
The useful forecast is not a prediction but an enumeration of open questions. Eight of them — where the field genuinely disagrees in 2026 — will shape the decade ahead.
继续堆算力与数据，还是需要结构性突破（新架构、世界模型、神经符号混合）？推理模型的出现让"训练规模一条路"的共识开始松动。
Continue stacking compute and data, or wait for structural breakthroughs — new architectures, world models, neuro-symbolic hybrids? Reasoning models have cracked the pure-scale consensus.
公开网页文本已接近被"吃完"。合成数据、自博弈、模型蒸馏能否持续提供高质量监督信号，还是会触发不可逆的模式崩溃？
Public web text is approaching exhaustion. Can synthetic data, self-play, and distillation sustain high-quality supervision — or will mode collapse eventually bite?
基础模型 + 机器人策略学习 + 仿真，能否复现 LLM 在语言上的"涌现"轨迹？VLA（Vision-Language-Action）模型是目前最值得押注的方向之一。
Can foundation models + robot policy learning + simulation reproduce the "emergence" curve we saw with language? VLA (vision-language-action) models are among the most bet-on directions.
AlphaFold 已经把蛋白质折叠从研究问题变为数据库问题。下一个候选：材料设计、核聚变等离子体控制、形式化数学证明。
AlphaFold turned protein folding from problem to database. Next candidates: materials design, fusion plasma control, and formal mathematics.
能力增长快于人类理解它的速度。机制可解释性（mech interp）、可扩展监督、行为审计是目前最硬核的三条研究线，尚未合流。
Capability is outrunning understanding. Mechanistic interpretability, scalable oversight, and behavioural auditing form three live threads that have yet to converge.
训练集群集中于少数国家与公司。芯片出口管制、能源分配、数据中心选址已成国家战略议题。算力不再只是工程问题。
Training clusters concentrate in a few nations and firms. Export controls, energy budgets, and data-centre siting are now instruments of statecraft.
Llama、DeepSeek、Qwen 证明开源模型可以紧咬前沿。开源会扩大创新半径，也会加速滥用风险扩散 —— 这是个真正的政策两难。
Llama, DeepSeek, and Qwen prove open-weight models can track the frontier. Openness widens innovation and diffuses misuse risk in the same breath — a live policy dilemma.
Agent 能可靠执行的工作流越来越长。哪些岗位会消失、哪些会被增强、哪些会被创造，决定于接下来五年的产品形态而非模型能力上限。
Agents handle longer reliable workflows each quarter. Which roles disappear, which get augmented, and which appear will be determined by product design, not the model capability ceiling.