Duration: ~30 minutes of video content Timestamps: 1:20:32 - 2:07:28
4.1 Understanding Hallucinations
What Are Hallucinations?
Hallucination: When an LLM generates plausible-sounding but factually incorrect or fabricated information.
User: Who wrote the book "The Azure Sky"?
Model: "The Azure Sky" was written by Jonathan Mitchell in 1987.
Reality: This book doesn't exist. The model invented it.Root Causes
| Cause | Explanation |
|---|---|
| ------- | ------------- |
| Training Objective | Models learn to generate plausible text, not truthful text |
| Pattern Completion | Statistical patterns don't encode truth |
| People-Pleasing | Models prefer giving answers over admitting ignorance |
| Compression Loss | Parameters store approximations, not facts |
Key Insight from Karpathy
"Models fabricate information due to their statistical nature. They learn they must always provide answers, even for nonsensical questions."
4.2 Types of Hallucinations
Categorization
┌─────────────────────────────────────────────────────────────┐
│ HALLUCINATION TAXONOMY │
├─────────────────────────────────────────────────────────────┤
│ │
│ 1. FACTUAL HALLUCINATIONS │
│ • Wrong dates, numbers, names │
│ • Invented citations or quotes │
│ • Non-existent events or people │
│ │
│ 2. FABRICATED CONTENT │
│ • Made-up books, papers, products │
│ • Fictional URLs or sources │
│ • Invented code libraries │
│ │
│ 3. CONFLATION │
│ • Mixing up similar entities │
│ • Combining features of different things │
│ • Temporal confusion (wrong time periods) │
│ │
│ 4. SELF-HALLUCINATION │
│ • Claiming capabilities it doesn't have │
│ • Inventing its own creation story │
│ • False claims about training data │
│ │
└─────────────────────────────────────────────────────────────┘Self-Identity Hallucinations
Untuned base models confidently make up their origins:
User: Who created you?
Base Model: I was created by [random company], a team of researchers at [random university]...
Reality: Without hardcoded system prompts, models confabulate their identity.4.3 Mitigation Strategy 1: Training for Uncertainty
Meta's Factuality Research
A systematic approach to reduce hallucinations:
┌─────────────────────────────────────────────────────────────┐
│ META FACTUALITY PIPELINE │
├─────────────────────────────────────────────────────────────┤
│ │
│ Step 1: Extract training snippets │
│ (Get passages from training data) │
│ ↓ │
│ Step 2: Generate factual questions │
│ (Create Q&A pairs from passages) │
│ ↓ │
│ Step 3: Produce model answers │
│ (Run model on questions multiple times) │
│ ↓ │
│ Step 4: Score accuracy │
│ (Compare answers to source passages) │
│ ↓ │
│ Step 5: Train refusal behavior │
│ (Model learns to say "I don't know" when uncertain) │
│ │
└─────────────────────────────────────────────────────────────┘Teaching "I Don't Know"
Include training examples like:
User: What was the GDP of Atlantis in 2020?
Assistant: I don't have information about this. Atlantis is a mythical
location, so it wouldn't have economic statistics.
User: Who won the 2030 World Cup?
Assistant: I don't have information about events after my training
cutoff date. The 2030 World Cup hasn't occurred yet.4.4 Mitigation Strategy 2: Tool Integration
The Tool Use Pattern
Instead of answering from parametric memory, models can call external tools:
┌─────────────────────────────────────────────────────────────┐
│ TOOL-AUGMENTED LLM │
├─────────────────────────────────────────────────────────────┤
│ │
│ User: What is Apple's current stock price? │
│ │
│ ┌──────────────────┐ │
│ │ LLM MODEL │ │
│ └────────┬─────────┘ │
│ │ Recognizes need for real-time data │
│ ▼ │
│ <SEARCH_START>Apple stock price<SEARCH_END> │
│ │ │
│ ▼ │
│ ┌──────────────────┐ │
│ │ SEARCH ENGINE │ → Returns: "$192.45 as of 2:30 PM" │
│ └──────────────────┘ │
│ │ │
│ ▼ │
│ Model: Based on my search, Apple (AAPL) is currently │
│ trading at $192.45. │
│ │
└─────────────────────────────────────────────────────────────┘Available Tools
| Tool | Purpose | Reduces Hallucinations For |
|---|---|---|
| ------ | --------- | --------------------------- |
| Web Search | Current information | Facts, news, prices |
| Calculator | Math operations | Arithmetic, formulas |
| Code Interpreter | Execute code | Complex calculations |
| Database | Structured queries | Company data, statistics |
| Knowledge Base | Verified facts | Domain-specific info |
Training Tool Use
Models learn tool patterns through examples:
User: What's 17.5% of 4,382?
Assistant: <CALC>4382 * 0.175</CALC>
System: 766.85
Assistant: 17.5% of 4,382 is 766.85.For KeenDreams: This pattern directly applies - KeenDreams can be a "memory tool" that the LLM learns to query when project context is needed.
4.5 Mitigation Strategy 3: Context Injection
Retrieval-Augmented Generation (RAG)
Provide relevant information in the context window:
┌─────────────────────────────────────────────────────────────┐
│ RAG PATTERN │
├─────────────────────────────────────────────────────────────┤
│ │
│ 1. User Query: "What are the project requirements?" │
│ ↓ │
│ 2. Retrieval: Search knowledge base for relevant docs │
│ ↓ │
│ 3. Context Construction: │
│ ┌─────────────────────────────────────────────────────┐ │
│ │ System: You are a helpful assistant. Use the │ │
│ │ following documents to answer questions. │ │
│ │ │ │
│ │ [Document 1: Requirements.md content...] │ │
│ │ [Document 2: Specs.md content...] │ │
│ │ │ │
│ │ User: What are the project requirements? │ │
│ └─────────────────────────────────────────────────────┘ │
│ ↓ │
│ 4. Model generates answer from provided context │
│ │
└─────────────────────────────────────────────────────────────┘Why Context > Parameters
| Aspect | Parametric | Contextual |
|---|---|---|
| -------- | ------------ | ------------ |
| Accuracy | Approximate | Exact |
| Updatability | Requires retraining | Instant |
| Verifiability | Opaque | Traceable |
| Reliability | Variable | High |
Key Insight
"Pasting information directly into context windows produces higher-quality outputs than relying on parametric knowledge."
For KeenDreams: This is the core value proposition - acting as the retrieval layer that surfaces relevant project memories into context, dramatically reducing hallucinations about project state.
4.6 Mitigation Strategy 4: Verification Chains
Multi-Step Verification
Have models check their own work:
User: Summarize this research paper.
Model (Step 1): [Initial summary]
Model (Step 2 - Verification):
Let me verify my summary against the source:
- Claim 1: "The study found X" ✓ (Page 3, paragraph 2)
- Claim 2: "Results showed Y" ✓ (Table 2)
- Claim 3: "The authors concluded Z" ⚠️ (I should reread the conclusion)
Model (Step 3): [Corrected summary with citations]Cross-Model Verification
Use multiple models to check each other:
# Generate with Model A
response_a = model_a.generate(prompt)
# Verify with Model B
verification = model_b.verify(
claim=response_a,
context=original_documents
)
# Return verified response or flag for human review4.7 Practical Hallucination Detection
Red Flags
Watch for these patterns that often indicate hallucinations:
| Pattern | Example | Risk Level |
|---|---|---|
| --------- | --------- | ------------ |
| Specific numbers without source | "Studies show 73.2% of..." | High |
| Named citations | "According to Smith et al. (2019)..." | High |
| Detailed URLs | "Available at example.com/specific/path" | Very High |
| Confident edge cases | Obscure historical details | High |
| Technical specifics | Exact code library versions | Medium |
Verification Techniques
- Ask for sources: "What source supports this claim?"
- Cross-reference: Search for claimed facts independently
- Challenge specifics: "Are you certain about that number?"
- Request uncertainty: "Rate your confidence 1-10"
4.8 Key Takeaways
Summary
| Strategy | Approach | Effectiveness |
|---|---|---|
| ---------- | ---------- | --------------- |
| Uncertainty Training | Train model to say "I don't know" | Medium |
| Tool Integration | External search, calculators | High |
| Context Injection | RAG, document retrieval | Very High |
| Verification Chains | Self-check, multi-model | High |
For AI Analytics Platforms
Critical Monitoring Points:
- Hallucination Detection: Flag responses with high-risk patterns
- Source Attribution: Track whether responses cite provided context
- Confidence Scores: Capture model-expressed uncertainty
- Tool Usage: Monitor when models invoke external tools
- Verification Loops: Log correction chains
For KeenDreams
Applicable Learnings:
- Memory as Grounding: KeenDreams provides factual anchor for project context
- Source Tracking: Always include provenance with retrieved memories
- Confidence Metadata: Store reliability scores with memories
- Verification Integration: Enable models to cross-check against cloud brain
Practice Questions
- Why do LLMs hallucinate even when they "know" the correct answer?
- How does tool integration reduce hallucinations?
- When is parametric knowledge acceptable vs. requiring context?
- How would you design a hallucination detection system?
Next Module
→ Module 5: Reinforcement Learning
Timestamps: 1:20:32 - Hallucinations & Tool Use | 1:41:46 - Knowledge of Self | Research on factuality and mitigation strategies