课堂-AI 能力与边界入门Lesson - AI Capabilities and Limits

能把一个模糊需求改写成有输入、有完成标准的任务。

能区分输出通顺与输出有来源支持，并各举一个课堂例子。

能说出三类要人工复核的内容：数字与日期、外部事实、影响他人权益的判断。

能按场景与多个指标描述一次结果，不只报一个总分。

Rewrite a vague need into a task with inputs and completion criteria.

Distinguish fluent output from source-supported output and give one classroom example of each.

Name three kinds of content that need human review: numbers and dates, external facts, and decisions affecting people’s rights.

Describe a result by scenario and multiple metrics rather than reporting only one total score.

微软 Learn《Introduction to generative AI and agents》模块页：课前读 LLM 与提示两节。

HELM 论文摘要页（arXiv 2211.09110）：看它为什么分场景、分指标。

每组一张记录表，五列：任务、输入材料、模型输出、判断依据、待复核项。

学员自带一个对话式 AI 入口；课堂不统一采购，也不收集账号信息。

Microsoft Learn module Introduction to generative AI and agents: read the LLM and prompt sections before class.

HELM abstract page (arXiv 2211.09110): see why it evaluates by scenario and metric.

One recording sheet per group with five columns: task, input material, model output, basis for judgment, items to recheck.

Learners bring their own chat AI entry point; the class does not procure accounts or collect account information.

1. 2 分钟导入：各组写下一个真实需求，教师把它改写成「输入＋完成标准」。
2. 4 分钟演示：同一份材料问两次。一次只说「介绍一下」，一次给字段与限制。
3. 4 分钟练习：各组填记录表，把断言分成「材料支持」与「需外部核对」两类。
4. 2 分钟小结：每组报一条边界与一条兜底动作，教师按场景与指标口径归拢。

1. 2-minute opener: each group writes a real need, and the teacher rewrites it as input plus completion criteria.
2. 4-minute demo: ask about the same material twice, once with only introduce this and once with fields and limits.
3. 4-minute practice: groups fill the sheet and split claims into supported by the material and needing external checks.
4. 2-minute wrap-up: each group reports one limit and one fallback action; the teacher groups them by scenario and metric.

把语气笃定、结构完整当成内容正确；流畅度不是证据。

把一个总分或榜单名次当成所有任务都能用；HELM 分场景测七个指标。

把「模型说不知道」当成能力差；缺输入或超范围时承认未知是可用的行为。

Treating confident tone and complete structure as correctness; fluency is not evidence.

Treating one total score or leaderboard rank as fit for every task; HELM measures seven metrics across scenarios.

Treating the model saying it does not know as weakness; admitting unknown when input is missing or out of scope is usable behaviour.

任务写全「输入＋完成标准」，缺一项就回到导入环节重写。

至少标出一处需外部核对的断言，并说明准备用哪个来源核对。

能说清这节课没做什么：只核对来源与条件，不做性能实测，也不给模型排名。

The task includes input and completion criteria; if either is missing, return to the opener and rewrite.

At least one claim is marked as needing external checking, with the source to be used named.

Learners can state what this lesson did not do: only source and condition checks, no performance benchmark and no model ranking.