一周路线The Full Week
一周one week
六站各选代表主题，共二十五步。把能力、成本、许可与评测放在一起看，最后用具体任务检查自己的选择。Follow twenty-five representative steps across six stations. Connect capability, cost, licensing and evaluation, then examine a choice for a concrete task.
- 人工智能Artificial Intelligence, AI
- 大模型 / 大语言模型Large Language Model, LLM
- 提示词Prompt
- 能力边界Capability Boundary
- 幻觉Hallucination
- 温度Temperature
- 知识截止Knowledge Cutoff
- 联网 / 检索增强Retrieval-Augmented Generation, RAG
- 智能体 / AgentAgent
- 工具调用Tool Calling / Function Calling
- MCPModel Context Protocol
- 工作流编排Orchestration / Pipeline
- 评估Evaluation, Evals
- TransformerTransformer
- 注意力机制Attention
- 训练三阶段SFT → Reward Model → RL
- 上下文窗口的代价Context Cost
- 显存VRAM
- 量化Quantization
- GGUFGGUF
- llama.cppllama.cpp
- 开源模型Open-weight Model
- 许可证License
- 基准测试Benchmark
- 模型评测的局限Evaluation Limits
- 拿到一个新模型的介绍页，能列出该问的问题：能力、边界、成本、许可、锁定
- 判断一件事该用 API 还是本地跑，并说出判断依据
- 看懂一份基准榜单，并说出它至少一处代表不了你的实际效果
- Given a new model's page, list the questions to ask: capability, limits, cost, licence, lock-in
- Decide API or local for a given job, and state the reasoning
- Read a benchmark table and name at least one thing it cannot say about your actual use
先试着回答，再展开参考答案。答不上来的地方，可从上方步骤返回对应词条。参考回答是核对要点，不要求逐字复述。Try answering before opening the reference. Revisit the relevant step above if needed. These answers highlight key ideas; you do not need to repeat the wording.
第 1 题 · 参考回答Question 1 · Reference answer
先写任务和验收条件，再查能力限制、数据去向、总成本、许可证、导出与迁移方式。不能用一个榜单名次替代这些问题。Define the task and acceptance criteria, then check limits, data handling, total cost, licensing, export and migration. A leaderboard position cannot answer these questions.
第 2 题 · 参考回答Question 2 · Reference answer
示例：低频且无部署经验的任务可先评估 API；不能把数据交给外部服务时，应评估符合约束的本地或私有方案。两边都要核算成本、质量、维护和硬件限制。For example, assess an API for occasional work without deployment expertise; assess a suitable local/private option when external data transfer is disallowed. Compare cost, quality, maintenance and hardware limits in both cases.
第 3 题 · 参考回答Question 3 · Reference answer
核对基准版本、输入、工具与评分口径；它未必覆盖你的中文材料、长文档、响应时间或失败成本。用自己的代表性任务补足缺口。Check benchmark version, inputs, tools and scoring. It may omit your Chinese documents, long inputs, latency requirements or cost of failure. Add representative tasks of your own.
用在一个具体任务里Apply this to a concrete task
半小时入门The Half-Hour Primer · 一个下午看懂原理One Afternoon Under the Hood · 内容工作者路线The Writer's Route · 动手开发者路线The Builder's Route · 选型决策者路线The Decision-Maker's Route · 研究学习者路线The Researcher's Route · 运维上线路线The Operator's Route · 普通消费者路线The Everyday User's Route · 图像 / 视频 / 音频创作者路线The Image / Video / Audio Creator's Route