Foundations of Generative AI

Mental Models and Effective Interaction

Our Next 3 Sessions

Session 1: Foundations (Today)

Mental models for understanding generative AI. Prompting fundamentals. How better instructions yield better outputs.

Session 2: Research & Analysis (Jan 11)

Practical applications for senior managers: research, competitive intelligence, document synthesis. Tools like Gemini Deep Research.

Session 3: Strategic Implications (Jan 18)

How AI affects your industry and role. What does the future look like. A framework for assessing your own position.

Section 01

The Intern Problem

A management challenge, not a technology problem

The Brilliant Intern

100x Speed

Completes in seconds what takes hours.

100x Bandwidth

Can work on many tasks in parallel.

Has Read Everything

Every book, paper, article, manual you can think of.

Eager to Please

Will attempt any task you give them, immediately.

The Catch

Cannot Tell Truth from Fabrication

Will confidently state things that are made up, with no internal sense that it's doing so.

Every Conversation Is Hour 1, Day 1

Complete amnesia between conversations. Knows nothing about you, your company, your context, every time.

"AI isn't dangerous because it's wrong. It's dangerous because it sounds right."

Why This Matters

Authority Without Accountability

Speaks with the confidence of an expert, but has no stake in being correct.

Plausibility Without Grounding

Optimized for 'sounds right' not 'is right'.

Confidence Without Uncertainty

Has no model of what it doesn't know. Won't say 'I'm not sure' unless you teach it to.

The Manager's Job

Decide Where to Trust

Which tasks can this intern handle on their own? Where does speed and breadth create real value?

Decide When to Verify

What outputs need human review? Where could a confident error cause real damage?

Decide How to Brief

What context does this intern need? How do you communicate to someone with zero institutional knowledge?

"To manage this intern well, we need to understand how it actually works."

Section 02

How LLMs Work

The mechanism behind the intern's behavior

The Puzzle

The Paradox

Why does ChatGPT write confident, fluent prose about things that are completely made up?

If It's So Capable...

Why does it invent fake legal citations?
Why does it fabricate research papers?
Why does it generate fake statistics?

Two Kinds of Intelligence

Human/Animal Intelligence

Formed via evolution
Optimized for survival
Embodied and multi-sensory
A teenager learns to drive in 15 hours

LLM Intelligence

Formed via training on text
Optimized for next-token prediction
Limited to language
Millions of hours of training, still can't drive

The Core Mechanism

Next Token Prediction (Auto-Regression)

Input: Everything typed so far
Model: Predicts probability distribution for the next word
Selection: Chooses one token (word/part of word)
Repeat: Adds choice to context and predicts again

DEMO

Auto-Regression Playground

Token-by-token generation with probability distributions

What the Demo Shows

Next Token Prediction

It's predicting what word comes next given everything before. That's the entire operation.

Temperature Controls Variation

Low temperature = predictable choices. High temperature = surprising choices (and more risk of errors).

Context Accumulates

Every token generated becomes part of the context for the next prediction. Errors compound.

DEMO

Hallucination Lab

Visualizing cascading errors

Why Hallucinations Happen

The Mechanism

At each step, the model chooses from probable next tokens. Sometimes a low-probability choice becomes the foundation for everything after.

No Internal Check

It's doing what it always does, predicting plausible next tokens. There's no flag for "this is made up."

Cascading Errors

Once a fabricated detail enters the context, subsequent tokens build on that shaky foundation.

Hallucination Traps

Fake Regulatory Details

"Under the 2024 SEBI Digital Asset Compliance Framework, Section 7.3.2..."

Fake Research Citations

"According to Gartner's 2025 Magic Quadrant for Enterprise AI Platforms..."

Fake Technical Specs

"The recommended token limit for Gemini 1.5 Pro when processing financial documents..."

Fake Precedents

"The Delhi High Court ruling in Sharma vs. TechMahindra (2023)..."

"AI isn't dangerous because it's wrong. It's dangerous because it sounds right."

Now you understand why

An Alternative Mental Model

The Simulated Mind

Think of each conversation as spinning up a simulated mind. It has knowledge and personality, but exists only for this conversation.

What This Captures

It can reason within a conversation
It builds on what you've said
It has preferences and tendencies

When to Use This Model

Helpful for complex, multi-turn conversations where you're building something together over many exchanges.

Section 03

Managing the Intern

Applying what you already know

A Different Kind of Software

Traditional Software

Deterministic: same input → same output
Operated like a tool
You learn its commands
Capabilities are fixed

LLMs

Probabilistic: same input → varied output
Managed like an employee
It responds to how you communicate
Capabilities can be discovered

Similar to a Human Employee

What Works

Responds to clear briefs
Performs better with context
Managed through feedback
Adapts to your communication style

Implication

Brief it like a consultant
Provide background, not just instructions
First draft starts a conversation
"Act as a CFO" changes everything

Different from a Human Employee

Differences

No persistent memory
No judgment about truth
100x speed and bandwidth
No ego or bad days

Implication

Provide context every time
Verify high-stakes outputs
Use it for volume tasks
Tell it when something matters

The Amnesia Problem

The Issue

Every new conversation starts from zero. The intern forgets everything you've discussed.

Partial Solutions

Gemini Gems / Custom GPTs saved instructions
Memory features in some tools
Document upload for context

Your Job

Provide the context that only you know: your situation, audience, constraints, and definition of success.

"Context engineering matters more than prompt engineering. Context is what the tool builders don't know."

Section 04

Briefing the Intern

From principles to practice

The Core Principle

Good Prompt = Good Job Description

Write prompts as if briefing someone with no context about your company, situation, or goals.

What a New Hire Needs

What's the situation?
What do you need from me?
What does good look like?
Who is this for?

Three Principles

Be Specific

Define the output you need
Ask for options when you want creativity
Specify format, length, tone

Give Context

Background and situation
Audience and purpose
Constraints and limitations

Iterate

First output is a draft
Give feedback, request changes
Build on what works

Weak vs. Strong Briefs

WEAK

What are the trends in my industry?

RESULT

Generic response
Nothing actionable
Requires follow-up

Weak vs. Strong Briefs

STRONG

List three emerging trends in Indian IT services and explain how each could impact mid-sized companies (500-2000 employees) competing against TCS and Infosys. Focus on differentiation opportunities.

RESULT

Specific insights
Tailored to situation
Immediately useful

Role Assignment

Without Role

"What are the risks of this project?" → Generic list

With Role

"Act as a CFO advising the CEO, what are the financial risks and rewards?" → Focused analysis

Why It Works

Role assignment is shorthand for context. It specifies perspective, priorities, and expertise.

Section 05

Advanced Techniques

Getting more from your intern

Interactive Prompting

The Problem

Writing all context upfront is hard. You don't always know what details matter.

The Solution

Tell the intern: "Before you start, ask me questions to understand the situation. One at a time, building on my answers."

Why It Works

The intern extracts context through questioning, like a good consultant would.

DEMO

Interactive Prompting

Performance feedback scenario

Interactive Prompt Template

THE BRIEF

I need to write difficult performance feedback. Before drafting, interview me to understand the situation. Ask one question at a time, building on my responses. Cover: what happened, the impact, our relationship, what outcome I want, and constraints. Then draft the feedback.

WHAT HAPPENS

Asks about specific behaviors
Explores impact
Understands context
Clarifies goals
Drafts nuanced feedback

Voice Prompting

The Problem with Typing

When you type, you edit as you go. You compress. You leave out context that seems obvious.

Voice Captures More

You speak more naturally, include background you'd skip, and explain nuance.

Tools

Wispr Flow and similar tools let you dictate prompts, then clean up the transcription.

DEMO

Voice Prompting

Speech captures more context than typing

Meta-Prompting

The Concept

Ask the intern to help you write the brief for your actual task.

How It Works

Describe what you're trying to accomplish, ask for a detailed brief, then review and use it.

Why It Works

The intern knows what information it needs. Let it tell you.

Meta-Prompting Example

META-BRIEF

I need to synthesize API documentation for a business head evaluating build vs. buy. Write me a detailed brief for this task. Include what context to provide and what output format works best.

OUTPUT

Brief template
Context questions
Suggested structure
Evaluation criteria

Gemini Gems: Reusable Briefs

The Principle

Once you have a brief that works, save it.

Gems / Custom GPTs

Save refined briefs as tools
Include role and context
Build a library

Examples

Meeting notes synthesizer
Tech-to-exec translator
Vendor evaluation summarizer

Section 06

Live Problem Solving

Putting the intern to work

Scenario 1: Document Synthesis

The Task

Synthesize a technical document for a non-technical audience.

Examples

Analyst report → CTO brief
API documentation → Build vs. buy
Earnings call → Client talking points

Key Context

Who is the audience?
What decision will they make?
What do they already know?

Scenario 2: Performance Feedback

The Scenario

High performer with collaboration issues. Technically excellent but creating friction.

Why It's Hard

Too soft → message doesn't land
Too harsh → lose a performer
Nuance matters

Approach

Interactive prompting. Let the intern interview you, then draft.

WORKSHOP

Hands-On: Your Problems

Open session to work through real challenges

Section 07

Takeaways

The Mental Model

The Brilliant Intern

100x speed, broad knowledge, but cannot tell truth from fabrication, and every conversation starts fresh.

The Core Risk

AI isn't dangerous because it's wrong. It's dangerous because it sounds right.

Your Job

Decide where to trust, when to verify, how to brief. That's management.

The Techniques

Context Matters Most

Good brief = good job description for someone with zero context.

Interactive Prompting

Let the intern interview you to extract context.

Voice

Speaking captures more detail and nuance than typing.

Meta-Prompting

Use the intern to write your briefs.

Save What Works

Build a library of Gems for recurring tasks.

Next Steps

Between Sessions

Try the tools
Experiment with interactive prompting
Bring your results to Session 2

Session 2 (Jan 11)

Research, analysis, and decision support. Gemini Deep Research, NotebookLM, real problems.