ml-feasibility

Source Metadata

Original cangjie-skill frontmatter from the distillation run:

name: ml-feasibility
description: |
  当用户想系统评估一个AI项目是否值得投入时使用，结合概念简单性和数据充足性两个维度。
  用户在问"这个想法可行吗"、"我们需要多少数据"、"技术上能做到吗"时激活。
  不适用于: 纯数据科学洞察项目、探索性分析、已明确不需要AI的场景。
source_book: 《给所有人的AI入门课》 吴恩达 (Andrew Ng)
source_chapter: p06, p07
tags: [feasibility, task-evaluation, data-requirements, prioritization]
related_skills:
  - slug: ab-mapping
    relation: depends-on
  - slug: one-second-rule
    relation: composes-with
  - slug: triple-due-diligence
    relation: depends-on

ML Feasibility Two Factors — Simple Concept + Sufficient Data

R — Original Text (Reading)

"Learning simple concepts is more likely to succeed... requiring less than a second of thought... Second, sufficient data makes it easier to achieve."

— Andrew Ng, p06

I — Methodological Skeleton (Interpretation)

The success of a machine learning project depends on two independent but both necessary factors: First, concept simplicity — can a human make the required judgment in less than one second? Simple concepts correspond to tasks with clear patterns and well-defined boundaries, which AI can learn easily. Second, data sufficiency — do you have enough labeled (A, B) samples? Data is the fuel for AI learning; no matter how simple the task, it cannot be trained without data. The two factors form a 2×2 matrix:

Simple + Has Data → Proceed immediately
Simple + Lacks Data → Can try but need to solve the data problem
Complex + Has Data → Maybe feasible, but needs more data and more time
Complex + Lacks Data → Abandon or redefine the problem This framework breaks down the vague question of "Can AI do it?" into two specific questions that can be answered independently.

A1 — Application in the Book (Past Application)

Case 1: Car detection (Simple + Large amount of data → Success)

Problem: Detecting vehicles ahead in autonomous driving
Use of the methodology: Concept is simple (a human can judge in one second) + data is sufficient (massive labeled images)
Conclusion: Both factors are satisfied, should proceed
Result: Became a core capability of autonomous driving

Case 2: Gesture recognition (Complex + Limited data → Failure)

Problem: Recognizing the meaning of human gestures through a camera
Use of the methodology: Concept is complex (gesture meaning depends on context) + data is limited
Conclusion: Neither factor is satisfied, should not proceed
Result: It was indeed difficult to do well at the time

Case 3: Empathetic replies (Complex + 1000 samples → Failure)

Problem: Automatically generating empathetic customer service replies
Use of the methodology: Concept is complex (requires emotional understanding) + data volume is also insufficient
Conclusion: Concept complexity is the main obstacle
Result: Even with data, it is difficult to succeed

Case 4: Coffee cup quality inspection (Simple + 100 images → Success)

Problem: Detecting appearance defects in coffee cups
Use of the methodology: Concept is simple (you can tell good from bad at a glance) + data volume is not large but sufficient
Conclusion: Simple tasks have lower data volume requirements
Result: A small dataset can also train a usable model

A2 — Trigger Scenarios (Future Trigger) ★

In what situations will users need this skill?

The team proposes a specific AI project idea and needs to systematically evaluate whether it is worth investing resources in.
A product manager is prioritizing multiple AI features and needs an evaluation framework.
A technical lead is reviewing an AI project proposal and needs to judge its feasibility.
An entrepreneur is validating an AI startup direction and needs to quickly rule out infeasible options.

Language signals (activate when these appear in the user's words)

"Is this idea feasible?"
"How much data do we need?"
"Can this be done technically?"
"Is this project worth doing?"
"What accuracy can AI achieve?"

Distinction from adjacent skills

Difference from one-second-rule : The one-second rule evaluates only the single factor of "conceptual simplicity"; the two-factor framework adds the "data sufficiency" dimension and is a more complete evaluation.
Difference from ab-mapping : A→B mapping formalizes the problem; the two-factor framework evaluates feasibility after formalization.
Difference from triple-due-diligence : Triple due diligence is a deeper, comprehensive evaluation (technical + business + ethical); the two-factor framework is a quick feasibility judgment.

E — Executable Steps (Execution)

When the skill is activated, the agent should execute the following steps:

Evaluate conceptual simplicity
- Ask the user: "What judgment does this task require a human to make? Can an experienced person make it within one second?"
- Completion criterion: The user can clearly answer "yes/no", or give the time required for human judgment.
- Stopping condition: If the user answers "it takes more than a few minutes", mark it as a "complex concept".
Evaluate data sufficiency
- Ask the user: "How many labeled (input, output) samples do you have on hand? Can you obtain more?"
- Completion criterion: The user can give an order of magnitude for the data (e.g., "1000 records", "none", "can purchase").
- Stopping condition: If the user has no data at all and cannot obtain any, mark it as "data unavailable".
Give the 2×2 matrix positioning and action recommendations
- Simple + has data → "Recommend starting the project immediately"
- Simple + lacks data → "Recommend solving the data problem first (labeling, purchasing, synthesis)"
- Complex + has data → "Recommend proceeding cautiously; more resources and time may be needed"
- Complex + lacks data → "Recommend abandoning or redefining the problem"
- Completion criterion: The user gets a clear matrix positioning and next-step action recommendations.

B — Boundary ★

Do not use this skill in the following situations

Data science projects (the goal is insight rather than prediction, and labeled data is not needed)
Exploratory analysis (A and B have not yet been defined)
Pure technical research (no clear business application scenario)
The project has already been decided to move forward, and what is needed is an execution plan rather than a feasibility evaluation.

Failure modes the author warns about in the book

Overestimating data sufficiency—"we have data" does not equal "we have enough high-quality labeled data"
Underestimating conceptual complexity—some tasks feel simple to humans because they rely on common sense that AI does not have
Ignoring data quality issues—having data but inconsistent labeling is equivalent to insufficient data

The author's blind spots / era limitations

The definition of "simple" is subjective—the criteria for "simple" differ across domains, and the book does not provide an objective method for judgment.
This framework is based on the supervised learning paradigm and does not sufficiently discuss the data requirements of self-supervised learning, transfer learning, or pretrained large models.
The book does not discuss "data acquisition cost"—some data is theoretically obtainable but economically infeasible.
In the LLM era, many "complex" tasks become feasible through prompt engineering, and the two-factor framework needs to be updated.

Easily Confused Adjacent Methodologies

Difference from "technology maturity assessment": technology maturity assessment is more macro-level, while the two-factor framework focuses on two key variables of ML projects.
Difference from "data audit": data audit only focuses on data quality, while the two-factor framework considers both concepts and data.

Related skills

ab-mapping (depends-on): The prerequisite of the two-factor framework is that clear A and B have already been defined.
one-second-rule (composes-with): The one-second rule provides a quick assessment method for the "conceptual simplicity" dimension.
triple-due-diligence (depends-on): After the two-factor check passes, proceed to triple due diligence for in-depth assessment.

Audit Information

Verification passed: V1 ✓ / V2 ✓ / V3 ✓
Test pass rate: Pending test
Distillation time: 2026-06-29