Skip to content

module-0-prerequisites

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

module-0-prerequisites

README.md

Module 0: Prerequisites - AI Verification Basics

Duration: 20 minutes

For: Developers new to LLMs or AI verification

Goal: Understand core concepts before diving into verification techniques

What You'll Learn

  • 🤖 What is an LLM and how does it work?
  • 💭 What are hallucinations and why do they happen?
  • ⚖️ Probabilistic vs. Deterministic systems
  • 🎯 Why verification is critical for production AI

1. What is a Large Language Model (LLM)?

The Simple Explanation

An LLM is a text prediction machine trained on massive amounts of internet text.

How it works:

  1. You give it a prompt: "The capital of France is"
  2. It predicts the most likely next word: "Paris"
  3. It keeps predicting: "Paris, known for the Eiffel Tower..."

Key insight: It doesn't "know" facts. It predicts patterns.

Real Example

Prompt: "2 + 2 ="
LLM predicts: "4" ✅ (saw this pattern a million times)

Prompt: "2843 + 7291 ="  
LLM predicts: "10134" ✅ or "9134" ❌ (less common, might guess wrong)

What LLMs Are Good At

✅ Pattern Recognition

  • Writing emails
  • Translating languages
  • Summarizing text
  • Code generation (common patterns)

✅ Creative Tasks

  • Brainstorming ideas
  • Writing stories
  • Conversational responses

What LLMs Are Bad At

❌ Precise Calculations

  • Math (they predict digits, not calculate)
  • Logic (they pattern-match, not reason)
  • Code execution (they generate code, can't run it)

❌ Factual Accuracy

  • Might mix up dates, numbers, names
  • Can't distinguish truth from plausible-sounding fiction

2. What Are Hallucinations?

Definition

Hallucination: When an LLM generates plausible-sounding but incorrect information.

Why They Happen

LLMs are trained to predict plausible text, not true text.

Example:

User: "Who was the first person on Mars?"
LLM: "Neil Armstrong in 1969."

✅ Plausible (Armstrong was first on Moon)
❌ Wrong (no one has been to Mars yet)

Types of Hallucinations

  1. Factual Errors

    • Wrong dates, numbers, names
    • "iPhone 15 was released in 2019" ❌

  2. Logic Errors

    • "If A > B and B > C, then C > A" ❌

  3. Calculation Errors

    • "15% of $200 is $35" ❌ (should be $30)

  4. Invented References

    • "According to study XYZ-2023..." (study doesn't exist)

Real-World Impact

Healthcare: Wrong dosage (could be fatal)
Finance: Incorrect interest calculation ($12,889 error in production)
Legal: Fake case citations (lawyer sanctioned)
E-commerce: Hallucinated discounts (revenue loss)

3. Probabilistic vs. Deterministic Systems

Probabilistic (LLMs)

How it works: Predicts based on patterns and probabilities

# LLM generates different answers each time
llm("What is 2+2?")
→ "4" (90% probability)
→ "2+2 equals four" (8%)
→ "The answer is 4" (2%)

Characteristics:

  • ✅ Flexible, creative
  • ✅ Handles ambiguity well
  • ❌ Not 100% reliable
  • ❌ Different outputs for same input

Use cases: Creative writing, summarization, conversation

Deterministic (Verification Engines)

How it works: Follows exact rules, always same output

# Calculator always gives same answer
calculator(2 + 2)
→ 4 (100% certainty, always)

Characteristics:

  • ✅ 100% reliable (for verifiable tasks)
  • ✅ Same input = same output
  • ❌ Can't handle ambiguity
  • ❌ Needs precise specifications

Examples: SymPy (math), Z3 (logic), compilers (code)

Why Both Matter

Best practice: Use LLM for generation, deterministic tools for verification

User Query (English)
    ↓
LLM translates to code
    ↓
Deterministic engine verifies
    ↓
Return verified result

This is the neurosymbolic approach QWED uses!

4. Why Verification is Critical

The Trust Problem

Without Verification:

  • Hope LLM is right (73-85% accuracy on finance tasks)
  • Manually check outputs (slow, error-prone)
  • Ship bugs to production (costly)

With Verification:

  • Mathematically prove correctness
  • Block errors before they reach users
  • Ship with confidence

Cost of Unverified AI

Industry Cost of Error Example
Healthcare Lives Wrong dosage (1000x overdose)
Finance Revenue $12,889 calculation error
Legal Sanctions Fake case citations
E-commerce Trust Hallucinated discounts

When Verification is Required

✅ Must verify:

  • Financial calculations
  • Medical dosages
  • Legal citations
  • Security checks
  • Regulatory compliance

⏸️ Optional verification:

  • Creative writing
  • Casual conversation
  • Brainstorming
  • Subjective opinions

Quick Check: Did You Understand?

Answer these to test yourself:

  1. What is an LLM?

    Answer A text prediction machine that generates likely next words based on training patterns, not facts.

  2. Why do hallucinations happen?

    Answer LLMs predict plausible text, not true text. They can't distinguish fact from fiction.

  3. Difference between probabilistic and deterministic?

    Answer Probabilistic (LLM): Flexible but unreliable, different outputs each time. Deterministic (Calculator): Exact same output for same input, 100% reliable.

  4. When should you use verification?

    Answer When errors have real consequences: money, lives, legal issues, security.

Key Takeaways

✅ LLMs predict patterns, not facts
✅ Hallucinations are inevitable (not bugs, it's how they work)
✅ Probabilistic ≠ Deterministic
✅ Verification prevents costly errors

Next: Now that you understand the problem, learn how QWED solves it!

→ Module 1: The Crisis

Additional Resources