零基础概念地图A beginner-first concept map

别背术语,
动手看因果。
Don't memorize.
Make cause visible.

从一句文字如何被切开,到 Agent 为什么必须停下来校验。每页只改变一个变量,让抽象机制在你手里发生。From how a sentence gets split to why an Agent must stop and verify. Each page changes one variable so the mechanism happens in your hands.

8 个系列Eight series

当前开放 160 个可交互概念页One hundred sixty interactive lessons are now open

S1

表示与数据Representations & Data

看见文字、图片与数据怎样变成模型可处理的表示。See how text, images, and data become representations a model can process.

S2

学习机理Learning Mechanics

从标签、目标与概率,到完整训练、数据切分与泛化边界。From labels, targets, and probabilities to complete training, data splits, and generalization boundaries.

S3

模型架构Model Architecture

从神经元到 Transformer,拆开信息在结构中流动的路径。From neurons to Transformers, trace how information moves through structure.

S4

生成与解码Generation & Decoding

概率怎样一步步变成下一枚 token,以及随机性从哪里来。How probabilities become the next token, and where variation comes from.

S5

训练与适配Training & Adaptation

预训练、微调、LoRA 与对齐各自改变什么。What pretraining, fine-tuning, LoRA, and alignment each change.

S6

提示、上下文与检索Prompts, Context & Retrieval

看清提示怎样控制任务,以及有限窗口里该选择哪些材料。See how prompts shape a task and how to choose material for a limited window.

S7

Agent、工具与编排Agents, Tools & Orchestration

让目标、工具、环境、校验与停止条件形成可控闭环。Turn goals, tools, environments, checks, and stop conditions into a controlled loop.

S8

评测、安全与生产运行Evaluation, Safety & Production

用评测、权限、trace、预算和监控把系统带进现实。Bring systems into production with evals, permissions, traces, budgets, and monitoring.