第一篇 · PART 1

大模型 Large Language Model (LLM)

在亿万文本上训练出来的"文字接龙大师",也是所有 AI 助手、AI Agent 的"大脑"。 Trained on oceans of text, an LLM is the "brain" behind every AI assistant and agent — and at heart it just predicts the next token.

一句话定义 One-line Definition

大语言模型(LLM)是一种在海量文本数据上训练的神经网络模型,它的核心任务只有一件事:预测下一个 Token(一个文本片段,约等于半个到一个词)。

A Large Language Model is a neural network trained on enormous text corpora. Its core task is deceptively simple: predict the next token (a chunk of text, roughly half a word to a word).

想象一位读遍了全人类书籍的读者:他看似在做"文字接龙",但在接龙的过程中,他内化了语言、知识、逻辑与模式——于是当他接龙时,写出的往往不是随机的词,而是符合上下文的、有知识含量的回答

Imagine someone who has read nearly every book ever written. The task looks like word-chaining, but along the way the model internalizes language, world knowledge, logic and patterns — so its "next word" is usually a context-appropriate, knowledge-rich continuation, not a random one.

它是怎么工作的 How It Works

整个过程可以分为三步:

The pipeline has three steps:

✂️

① Token 化 Tokenization

把文本切成小片段(token),每个 token 映射为一个高维向量。中文一个字常被切成 1–2 个 token。

Text is split into tokens; each token becomes a high-dimensional vector. A Chinese character is often 1–2 tokens.

🧠

② Transformer 理解 Transformer

自注意力机制让每个 token 与上下文中的其他 token 相互"对视",模型由此理解句子的含义与结构。

Self-attention lets every token "look at" all others in the context, so the model grasps meaning and structure.

🔁

③ 自回归生成 Autoregressive Decoding

模型输出下一个 token 的概率分布并采样,把新 token 拼回输入,再预测下一个……循环往复,直到生成完整回答。

The model outputs a probability distribution over the next token, samples one, appends it, and repeats until the answer is complete.

亲手试一试:预测下一个 Token Try It: Predict the Next Token

Only the prompt is shown first; each click reveals the model's next token, conditioned on everything before it. (演示为固定脚本,真实模型每次采样略有随机)

关键点:模型每次只生成一个 token,但每一步都"看过"前面全部内容——这就是它能保持连贯的原因。它不是查数据库,而是根据学到的统计规律在"创作"。

Key idea: the model emits one token at a time, but each step conditions on everything before it — that is how coherence emerges. It is not querying a database; it is generating from learned statistical patterns.

为什么"大"这么重要 Why "Large" Matters

大模型的"大"体现在三处:参数规模(从十亿到万亿级)、训练数据(数万亿 token)、算力投入(成千上万块 GPU)。规模超过某个临界点后,模型会表现出涌现能力(emergent abilities)——翻译、写代码、多步推理等能力在小模型上几乎不存在,却在大模型上突然出现。

"Large" refers to parameter count (billions to trillions), training data (trillions of tokens), and compute (thousands of GPUs). Past a certain scale, emergent abilities appear — translation, coding, multi-step reasoning — that barely exist in small models.

🚀 例子:为什么 2025 年 AI 突然"能干活"了 Why AI suddenly "got things done" 一方面模型本身变强(GPT 系列、Claude、DeepSeek-R1 的推理能力);另一方面,模型开始被接上工具(文件、代码执行、浏览器)——后一件事,正是本网站接下来要讲的 HarnessAgentModels got stronger (GPT series, Claude, DeepSeek-R1's reasoning), and they also started being hooked up to tools — files, code execution, browsers. The latter is exactly what Harness and Agent are about on this site.

能力与局限 Capabilities & Limits

✅ 它能做的 What it can do

  • 总结、改写、问答、翻译 Summarize, rewrite, Q&A, translate
  • 生成与调试代码 Write and debug code
  • 逻辑推理与规划(推理模型更强) Reasoning & planning (stronger in reasoning models)
  • 遵循复杂指令、角色扮演 Follow complex instructions
  • 多语言、多模态(看图/听音) Multilingual, multimodal (image/audio)

⚠️ 它的局限 What it cannot do

  • 幻觉:会一本正经地编造事实 Hallucination: confidently fabricates facts
  • 知识有截止时间 Knowledge cutoff
  • 上下文窗口有限,太长会"忘" Finite context window
  • 数学/精确计算仍需小心 Exact arithmetic still error-prone
  • 无法自己行动:不能读文件、跑代码、上网——除非有人(或 Harness)替它接上工具 Cannot act on its own: no file access, no code execution, no web — unless a harness wires tools to it

代表性模型 Representative Models

截至 2026 年末,主流模型大致如下(能力与排名持续变动,价格以官方为准):

As of late 2026, the mainstream landscape looks roughly like this (leaderboards and pricing keep shifting):

OpenAI · GPT 系列 GPT series

ChatGPT 的缔造者。旗舰为 GPT-5.5 系列(2025 年 11 月发布,含 Instant 变体),多模态、推理强、已向免费用户开放;2026 年持续迭代。

Creator of ChatGPT. Flagship: GPT-5.5 series (Nov 2025, incl. Instant variant) — multimodal, strong reasoning, free tier included; iterating through 2026.

闭源 Closed

Anthropic · Claude

Opus 4.8(2026 年 5 月,支持算力调节、"动态工作流"、强化诚实性);Sonnet 5(2026 年 6 月)成为订阅默认模型,主打更省成本的 Agent 能力。

Opus 4.8 (May 2026: compute scaling, dynamic workflows, honesty); Sonnet 5 (Jun 2026) is now the default subscription model, tuned for cost-efficient agentic use.

闭源 Closed

Google · Gemini

Gemini 3(2025 年 11 月)→ 3.5 系列(2026 I/O 大会)→ 3.7 Flash(2026 年 8 月)。超长上下文、深度整合 Google 生态与搜索。

Gemini 3 (Nov 2025) → 3.5 series (I/O 2026) → 3.7 Flash (Aug 2026). Million-scale context, deep Google-ecosystem integration.

闭源 Closed

DeepSeek · V4

V4 预览版(2026 年 4 月)开源上线:Flash/Pro 双版本、百万 token 上下文、Agent 能力大幅升级,并首次亮相 DeepSeek Harness。前作 V3/R1 为 2025 年的里程碑。本网站正运行在 DeepSeek 系列模型之上。

V4 preview (Apr 2026) is open-sourced: Flash/Pro variants, million-token context, big agent upgrades, and the debut of DeepSeek Harness. V3/R1 were the 2025 milestones. This site runs on a DeepSeek-family model.

开源 Open

阿里 · Qwen(通义千问)

Qwen3.8 系列(2026 年 8 月)开源:旗舰 qwen3.8-2.4t-a95b(2.4T 总参数 / 95B 激活的 MoE)、27B 多模态,以及一系列小参数模型;全球下载超 30 亿次。

Qwen3.8 series (Aug 2026) open-sourced: flagship qwen3.8-2.4t-a95b (2.4T total / 95B active MoE), a 27B multimodal model, plus small-parameter variants; 3B+ downloads worldwide.

开源 Open

Kimi · 月之暗面

K3(2026 年 7 月)开源:约 2.8 万亿参数,是当前全球参数规模最大的开源模型,技术报告同步发布;前作 K2 为 2025 年开源标杆。

K3 (Jul 2026) open-sourced: ~2.8T parameters — currently the world's largest open-weight model, with a technical report; K2 was the 2025 milestone.

开源 Open
🔗 值得注意的联动:模型与 Harness 一起发布 Notable: model + harness shipped together DeepSeek V4 发布时,DeepSeek Harness 同步亮相并开源——这印证了本篇的核心观点:模型越来越强,但真正让模型"干活"的是外围的 Harness。详见 第二篇 HarnessDeepSeek open-sourced DeepSeek Harness alongside V4 — reinforcing this page's core point: models are getting stronger, but the harness around them is what makes them actually work. See Part 2: Harness.

关键认知:模型 ≠ 工具 ≠ 助手 Model ≠ Tool ≠ Assistant

🔌 裸模型只会"聊天" A bare model can only chat 大模型本身不能打开你的 FITS 文件、不能运行 Python、不能上网。它只能接收文字、输出文字。要让它真正"帮你干活",需要一层基础设施把模型与文件系统、Shell、浏览器、IDE 等工具连接起来——这层基础设施,就是下一节的主角:HarnessAn LLM alone cannot open your FITS file, run Python, or browse the web — it only takes text in and emits text out. To make it actually work for you, an infrastructure layer must connect the model to files, shell, browser, IDE — that layer is the Harness, the star of the next page.

下一步:什么是 Harness? → Next: What is a Harness?