Generative AI for Everyone

01 What is Generative AI?

Generative AI is a class of AI systems that can produce high-quality content — specifically text, images, and audio. Examples include ChatGPT, Bard, and BingChat.

AI has been mentioned on 38% of S&P 500 earnings calls as of early 2024 — the most mentions in tech, followed by energy. Generative AI as a developer tool will be even more impactful in the long term. AI is already pervasive: Google and Bing for web search, fraud detection in credit card payments, recommender systems on Amazon and Netflix.

02 How Generative AI Works

Where it fits in the AI landscape

Think of AI as a toolbox. Supervised learning is the most important existing tool. Generative AI is the newest addition. The toolbox also contains unsupervised learning and reinforcement learning.

Supervised Learning

Given an input (A), the model learns to predict an output (B). A decade of large-scale supervised learning laid the foundation for generative AI. Early models were small; researchers discovered that larger models with more data continue to improve, which unlocked LLMs.

Large Language Models (LLMs)

LLMs are built by using supervised learning to repeatedly predict the next word. Train on hundreds of billions of words and you get ChatGPT.

LLM as a Thought Partner

LLMs give us new ways to find information — like having a knowledgeable collaborator always on hand. They can brainstorm, rewrite, summarise, and reason. But they sometimes hallucinate — making up plausible-sounding but incorrect facts. Use trustworthy websites for high-stakes topics (health, legal); LLMs are better where less structured web information exists.

03 Generative AI Applications

GenAI tasks fall into three broad modes. Across all of them, the more context you give, the better the output.

04 What LLMs Can and Cannot Do

A useful mental model: think of an LLM as a fresh college graduate with lots of general knowledge but no company-specific context, no internet access, and no memory of past sessions. You get a different fresh graduate every conversation.

05 Tips for Prompting

1. Be detailed and specific

Give sufficient context. Describe the task in detail. Vague prompt → vague output.

Initial prompt

With context + detail

2. Guide the model to think through its answer

Add explicit steps to the prompt so the model follows a structured reasoning path.

Final prompt

3. Experiment and iterate

• There are no perfect prompts — adjust and improve until results are satisfactory
• Don't overthink the initial prompt; just start
• Be careful with confidential information
• Always double-check before fully trusting the output

06 Image Generation — Diffusion Models

Image generation is mostly done via Diffusion Models — a supervised learning approach at heart. Training: take a clean image + caption, progressively add noise across steps. The model learns to predict a less noisy image from a noisier one. Generation works in reverse: start from pure noise + a text prompt and iteratively denoise until a clean image appears (typically ~100 steps).

07 Using GenAI in Software Applications

Before GenAI, building a reading, writing, or chatbot app required months of ML engineering. Now a working prototype can be built in hours or days. Example — a reputation monitoring system:

08 Lifecycle of a Generative AI Project

Building GenAI applications is a highly empirical, iterative process — not a straight line from idea to deployment.

09 Retrieval Augmented Generation (RAG)

RAG lets an LLM answer questions using specific documents you provide at query time, rather than relying purely on its training knowledge. Think of the LLM as a reasoning engine, not a knowledge store.

RAG Applications

• Chat with PDFs — PDF.ai, ChatPDF, AskYourPDF, docAnalyzer.ai
• Article Q&A — CourseraCoach, Snapchat AI, Hubspot chatbot
• New web search forms — BingChat, You.com

10 Fine-tuning, Pretraining & Model Choices

Fine-tuning

Useful when the context is bigger than the input context length, or when the LLM needs domain-specific language (medical notes, legal, financial documents). Allows a smaller model to match a larger one on a specific task — lower cost, lower latency, runs on-device.

Pretraining

• Expensive — requires $10M+, many months, and vast data
• When in doubt: don't do it. Should be a last resort.
• BloombergGPT was trained on 50B+ parameters of proprietary Bloomberg data — outperforms open-source GPTs on financial text

Model Size

Closed vs Open Source

Instruction Fine-tuning & RLHF

A base LLM predicts the next word — if you ask "What is the capital of France?" it might respond with another question (the most common continuation in training data). Instruction fine-tuning trains it on Q&A pairs to follow instructions instead. RLHF (Reinforcement Learning with Human Feedback) goes further: train a reward model on human preference scores, then tune the LLM to generate responses the reward model rates highly. Companies use RLHF to make LLMs Helpful, Honest, and Harmless (the 3 Hs).

Tool Use & Agents

• Tools — LLMs can call external functions (place an order, run a calculation). LLMs are bad at precise math; a calculator tool solves this by letting the model offload computation and receive the result
• Agents — use LLMs to carry out complex sequences of actions autonomously. Cutting-edge research area, evolving rapidly

11 Cost Intuition

Pricing is based on input + output tokens. Output tokens are generally more expensive. A token is roughly ¾ of a word — 300 words ≈ 400 tokens. Common short words ("the", "a") count as 1 token; longer words split ("program" + "ming").

12 Generative AI & Business

Day-to-day uses

• Writing assistant & thought partner — drafting, editing, ideating
• Marketing — brainstorming email campaigns, ad copy
• Recruiting — summarising candidate reviews in 50 words or less
• Programming — writing Python, debugging, documentation

Task Analysis Framework

AI doesn't automate jobs — it automates tasks. Most jobs are a collection of many tasks. The framework: list the tasks in a role → assess each for GenAI potential along two dimensions:

Augmentation vs Automation

• Augmentation — AI helps a human do the task faster (human-in-the-loop)
• Automation — AI performs the task autonomously once trust is established

Job impacts across roles

• Programmer — writing code and documentation: high potential. Reviewing others' code, gathering requirements: lower
• Lawyer — drafting and reviewing documents, interpreting regulations: high potential. Negotiating, reviewing evidence: lower
• Landscaper — most tasks: low potential. Physical roles remain less impacted than knowledge roles

13 Teams to Build GenAI Software

A one-person team can be a software engineer with prompting know-how, or an ML engineer. A two-person team: ML engineer + software engineer. Many configurations work — start lean and scale the team as the product complexity grows.

14 Generative AI & Society

Concerns about AI

1. 1
   Will AI amplify human biases? Yes, it can reproduce biases in training data. Mitigations: fine-tuning with curated data and RLHF to train models on an unbiased reward model.
2. 2
   Who loses their job? Geoff Hinton predicted in 2016 that radiologists would be replaced in 5 years. A decade later, not one has lost their job — because radiologists perform 30+ tasks and X-ray interpretation is just one. The correct framing: "AI won't replace radiologists. Radiologists who use AI will replace radiologists that don't." — Curtis Langlotz, Stanford.
3. 3
   Human extinction? Arguments are not concrete. Perfect control isn't needed for technology to be valuable and safe — we can't control turbulence on aeroplanes, but we fly. AI will also be a key tool in solving real risks: climate change, pandemics.

Artificial General Intelligence (AGI)

General-purpose AI ≠ Artificial General Intelligence. AGI by definition can do any intellectual task a human can — drive a car after 20 hours of practice, complete a PhD thesis in 5 years. Though AI is powerful and better than humans at specific tasks, expecting AGI-level performance is an extremely high bar and not where we are today.

Responsible AI

15 Building a More Intelligent World

Intelligence is the power to apply knowledge and skills to make good decisions. Human intelligence is expensive — education, training, time. Artificial intelligence is cheap. AI has the potential to give every individual the ability to access intelligence at low cost.

Looking beyond AI, the world faces climate change, pandemics, and war. Solving them will require all the intelligence we can master — including artificial intelligence.