12 · The Matter Stack物质的层级

Artificial Intelligence人工智能

Level · Neuron层级 · 神经元

"We built minds before we understood our own.""我们在弄懂自己之前，先造出了别的心智。"

Plane位面

Intelligence智能

Value价值观

Liberate Humanity解放人类

Cosmic name其名

Autonomous Mind自主思维

The frontier前沿概述

Artificial intelligence is the project of building substrates that do the thing brains do: take in noisy signals, find structure, predict, decide, act. The defining shift of the 2010s was the discovery that scale is a feature, not just a constant: bigger models trained on more data with more compute behave qualitatively differently. The defining shift of the 2020s is that these systems have begun to act in the world — agents that browse, code, transact, negotiate. We do not yet know where this curve ends, only that humanity is going to find out within a generation.人工智能，是为“大脑所做之事”打造衬底的工程：吸收嘈杂信号、找出结构、预测、决策、行动。2010 年代的关键转变是：人们发现“规模”是一种特性，而不仅是常量——更大模型、更多数据、更多算力，会带来质变。2020 年代的关键转变是：这些系统开始在世界上行动——浏览、写代码、交易、谈判的智能体。曲线终点在哪里我们尚不可知，唯一可知的是：人类会在一代人之内，亲眼看到。

Historical evolution历史演化

1956
Dartmouth Workshop达特茅斯会议
McCarthy coins 'artificial intelligence.'麦卡锡造出“人工智能”一词。
1986
Backpropagation反向传播
Rumelhart, Hinton, Williams make multilayer networks learnable.Rumelhart、Hinton、Williams 让多层网络变得可学。
2012
AlexNetAlexNet
ImageNet falls to a deep convolutional network on two consumer GPUs.ImageNet 被运行在两块消费级 GPU 上的深度卷积网络攻破。
2017
TransformerTransformer
'Attention is all you need' — the architecture that powers every modern LLM.《Attention Is All You Need》——支撑所有现代大模型的架构。
2022–23
ChatGPT momentChatGPT 时刻
GPT-4 and Claude bring frontier capability into the consumer mainstream.GPT-4 与 Claude 把前沿能力带入大众消费。
2024–26
Agent era智能体时代
Long-horizon reasoning and tool use push AI from chatbox to coworker.长程推理与工具使用，把 AI 从聊天框推到协作者。

State of the art今日状态

Frontier models are now multimodal, agentic, and approaching expert-level performance on coding, mathematics, and scientific writing. Cost per intelligent token has fallen ~10× per year for three straight years. The hard problems are alignment, interpretability, and political economy — not raw capability.前沿模型已多模态、智能体化，并在编码、数学与科学写作上接近专家水准。每一份「智能 token」的成本，连续三年每年下降约十倍。难题在对齐、可解释性与政治经济学——而不在原始能力本身。

Where it goes next未来走向

→
Agentic AI as a default interface to most software within 3 years.三年内，智能体成为绝大多数软件的默认接口。
→
AI-driven scientific discovery in materials, biology, math.材料、生物、数学领域的 AI 驱动科学发现。
→
AGI debate moves from 'if' to 'how should it be governed.'关于 AGI 的辩论，从“是否会发生”转向“该如何治理”。

Applications today现今应用

Software engineering and knowledge work copilots.软件工程与知识工作的副驾驶。
Scientific research — protein folding, materials.科学研究——蛋白折叠、新材料。
Education — universal tutoring at near-zero marginal cost.教育——边际成本近零的全球家教。
Public health — diagnostic triage and drug discovery.公共健康——诊断分诊与药物发现。