Module 1 • Lesson 460 mins

How AI 'Understands' Requests

Attention Mechanism & Weighting, how word choice shifts tone, and boundaries between Parametric Knowledge and Ground Truth Verification Layers.

Attention Mechanism & token weighting calculations
Distinguishing Parametric Knowledge from Objective Ground Truth
Designing Ground Truth Verification Layers (RAG)

Model "Understanding" via Attention Weights and the Boundaries of Ground Truth

The biggest pitfall for PMs is assuming AI "reads and understands" like a human. In reality, AI "understanding" is pure mathematics: the Attention Mechanism calculating correlation weights between words.

Example: In EcoCart, understanding Attention clarifies why a small typo can derail entire outputs and why models cannot self-verify facts.

1. The Nature of AI "Understanding": Attention Weighting Mechanics

Humans read sequentially and connect text to real-world experience. LLMs process all tokens simultaneously: before generating each token, Attention calculates weights to identify which prompt words are most relevant.

"Understanding" is not human comprehension, but probability matrix computation matched against training patterns to select the next token.

Humans read in order; a model weighs everything at once

That difference is why one swapped word can pull an entire output off course.

Human reader

Sequential, left to right

Ties each word's meaning to real-world knowledge and context

Fills in typos and unstated intent automatically

Language Model

Parallel, all tokens at once

Assigns an Attention Weight to every token in the prompt

Processes exactly what you typed - nothing more, nothing less

Swap one word below to see where attention shifts - and how that changes the drafted email.

2. How Input Phrasing Shifts Tone and Direction

Models do not infer unstated intent; they weigh the literal tokens you type. A small phrasing change shifts downstream output entirely.

Example: A user intends to type "formal" ("trang trọng") but typos it as "luxurious" ("sang trọng"). A human colleague easily infers a sincere apology. A model instead assigns high attention weight to "luxurious", shifting the apology into opulent, high-society language ("elite caliber", "high society") - completely unfit for late delivery support.

One word, redirected attention, a different email entirely

The model doesn't ask “did you mean formal?” - it weighs the literal token you typed and lets that weight steer everything downstream.

Prompt tokens

Line thickness = attention weight on that token

Resulting draft

“Dear valued customer, we sincerely apologize for the inconvenience and appreciate your patience.”

Swap one word and attention shifts entirely - the model never checks if “luxurious” was a typo for “formal”.

3. Limits of Parametric Knowledge and Hallucination Risks

A model's memory is Parametric Knowledge - mathematical weights compressed from training. It cannot distinguish objective ground truth from popular linguistic patterns.

  • Common Fact: "How tall is Mount Fuji?" → Correctly returns 3,776m due to high frequency across training data.
  • Internal / Specific Fact: "What year was the founder of a new 10-person EcoCart startup born?" → Confidently invents a year (e.g. 1988) based on tech founder age patterns, lacking actual records.

Product Exercise: A seller enters "Write a description for new running shoes" with no technical specs:

  • The model generates copy using generic industry patterns ("cushioned rubber sole", "breathable air mesh").
  • PM Insight: These claims are hallucinations based on general patterns, completely untrustworthy as verified selling specifications.

4. The Boundary Between Language Prediction and Ground Truth Verification

To protect product safety, system architecture must decouple into 2 distinct layers:

ComponentGenerative Layer (LLM Model)Ground Truth Verification Layer (DB / RAG / APIs)
ObjectiveFluid phrasing and brand tone synthesis100% accurate factual data retrieval
NatureProbabilistic pattern inferenceDeterministic structured data lookup
EcoCart ExampleGenerating: "Your order #12345 is being shipped by..."Logistics API lookup: "ViettelPost - Tracking ID VP987"

PM Rule: Never allow models to recall or guess factual data (pricing, balances, inventory, policy terms) - retrieve them from deterministic databases and inject them into Prompt Context (Context Grounding).

Hands-on Workshop90 minutes

Design the Full EcoCart Refund AI Copilot - From Rule/Model Boundary to Ground Truth Layer

Module 1 Capstone: Design the complete EcoCart Refund AI Copilot for support agents:

  1. Rule/Model Boundary (Lesson 1): Separate deterministic business rules (return windows, order status) from generative model tasks (reply drafting).
  2. Automation Level (Lesson 2): Select Level 1 (Human-in-the-Loop), designing UI that clearly distinguishes AI suggestions from agent edits.
  3. System Prompt (Lesson 3): Author all 4 components (Task, Context with variables {customer_name}, {policy_text}, Constraints, Response Examples).
  4. Ground Truth Layer (Lesson 4): Inventory all factual data points (order ID, balance, policy rules) required from Database/APIs.
  5. Adversarial Testing: Run 3 edge-case inputs (incomplete info, ambiguous request, adversarial prompt) to verify system resilience.