Source Metadata
Original cangjie-skill frontmatter from the distillation run:
id: f27
title: "ML产出软件,DS产出洞察 — 项目类型判定框架"
book: "AI for Everyone"
author: "Andrew Ng（吴恩达）"
source: p04, p11, p12
tags: [scoping, machine-learning, data-science, project-management]
related_skills:
- slug: ml-workflow
relation: contrasts-with
- slug: ab-mapping
relation: depends-on
R — Reading (original quote)
"Machine learning projects usually produce software that can run and output results... the output of data science projects is usually a presentation" — p04
I — Interpretation (restate in your own words)
Machine learning and data science are two fundamentally different modes of work:
- Machine learning (ML): the output iscontinuously running software, which can repeatedly make predictions and judgments. For example: spam filters, recommendation systems, defect detectors.
- Data science (DS): the output isinsights or recommendations, usually presented in the form of a report or PPT, used to support one-off business decisions. For example: a market analysis that persuades a travel company to buy more advertising.
The two have different outputs → different team skills required → different success metrics → different project processes. Confusing these two project types is one of the most common reasons AI projects fail.
A1 — Past Application (past application cases)
- Online advertising delivery system (ML): build software that runs 24 hours a day, predicts user click-through rate (CTR) in real time, and automatically decides which ad to show.
- Advertising strategy analysis (DS): an analyst uses data mining to discover that "travel companies are underinvesting on the platform," creates a PPT, and presents it to executives to persuade travel companies to increase their spending. This is a one-off decision recommendation and does not require deploying software.
A2 — Future Trigger (future trigger conditions)
Trigger scenarios：
- When a team discusses an AI idea and argues about "what kind of project this actually is"
- When the boss asks "what exactly is this project supposed to deliver in the end"
- When hiring and agonizing over "should I hire an ML engineer or a data analyst"
Language signals：
- "Should we do ML or DS?"
- "What should this project produce?"
- "Are we building a system or doing analysis?"
- "Are we building a system or making a report?"
- "Is this a machine learning project or a data science project?"
Key distinction: ask one key question—Is the final output continuously running software, or a decision recommendation?
E — Execution (execution steps)
- Clarify the expected result: ask stakeholders—do they ultimately want software that can run continuously, or a decision recommendation?
- If the answer is software → follow the ML workflow: collect data → train model → deploy to production → continuously iterate.
- If the answer is a recommendation → follow the DS workflow: collect data → analyze and explore → form hypotheses → create a report/PPT.
- Confirm team skill fit: ML projects need ML engineers; DS projects need data analysts/business analysts.
Completion criteria: the team reaches consensus on "whether we are doing ML or DS" and matches the correct process and talent.
B — Boundary (usage boundaries)
Not applicable scenarios：
- The project contains both DS and ML elements (a common case—first use DS to discover insights, then build an ML system based on those insights)
- Purely academic research project
- Exploratory research (no clear output goal yet)
Failure modes：
- Treating DS as ML: putting great effort into building a system, but the business only needs a report
- Treating ML as DS: finishing the PPT and stopping there, without deploying software, unable to continuously generate value
Author's blind spot：
- Andrew Ng describes the boundary between ML and DS as clearer than it is in reality. In reality, many projectsmix the two—first use DS to find opportunities, then use ML to implement and execute. Don't get stuck because you can't classify clearly; the key is to identify the project'scurrent stageand its main form of output.
Related skills
- ml-workflow (contrasts-with): The ML workflow is a "building a system" process, whereas DS produces reports rather than software; the two processes are completely different
- ab-mapping (depends-on): After determining it is an ML project, you need to further define the inputs and outputs using an A→B mapping