AI = Anything that mimics human intelgence like chess program, chatbots, recommendation engine
Machine Learning = AI that learn patterns from data instead of following hard-coded rules
Deep Learning = Machine Learning using neural network (brain like structure with layers)
LLMs are deep learning models trained on massive amount of text to predict the next word. Think of it like auto-complete but so good, seems like understanding. This is ChatGPT, Claude, and other AI tools we use.
Tokens are how AI reads text, roughly 4 charactors = 1 token
"hello world" = 2 token
"supercalifragilisticexpialidocious" = 8 token
Understanding the tokens, saves you money and prevents mysterious errors or LLMs halucinations. The better we manage the tokens, the better we get the outputs.
Think of context window as number of pages an AI model can remember in one coversation.
GPT 4: 128K Tokens (100 Pages)
Claude 4: 200K Tokens (150 Pages)
If you hit the limit, AI forgets a lot mid-conversation.
Understaning the temperature settings decides how much creativity LLM model can use for the output.
temperature 0 = robotic, deterministic response (same input = same output)
temperature 0.7 = balanced creativity
temparature 2 = complete chaos, not sure what the response is and why
Wrong temperature destroys the response everytime.
Prompt engineering is a technique to frame context, give examples and structure the request.
The difference between a random user vs AI Power User, good prompt = 10x better result.
A bad prompt can make GPT4 perform bad than GPT3.
They are the first instruction that defines how AI should behave.
"You are a helpful assistant" vs "You are brutely honest business consultants"
Mastering this and you control how AI responses to everything, ignore it and AI will surpise you in (very) bad ways.
“Tokens” are what it reads/writes,
“Temperature” controls how it chooses words,
“System prompts” define who it is.
Learn to fine tune when prompting isn't enough.
You take a pre-trained model and train it further on your specific data, liking hiring a general expert and teaching them your industry.
Expensieve and complex, but creates AI that thinks exactly the way you want.
Only use when prompting isn't enough.
RAG let's AI search your data in data in real time, like giving AI a perfect memory of your companies knowledge base.
Cheaper, faster then fine tuning.
Most business AI application should start here.
RAG is a pattern that retrieves relevant data via semantic search, augments the LLM prompt with that context, then generates accurate responses.
Process:
- RETRIEVE: Vector DB finds semantically similar documents
- AUGMENT: Inject retrieved docs + user query into prompt
- GENERATE: LLM produces answer using all context
Cost consideration: Retrieved docs count as input tokens.
How softwares talks to software.
OpenAI APIs let's you send text and get AI responses back
This moves AI from a chat interface to integrated tool
Suddenly you CRM, email, website can all become AI-powered.
AI converts "the cat sat on the mat" into a list of 1536 numbers
Similar meaning get similar numbers, this enable AI to understand meaning no just match keywords.
Thisis foundation of smart search and recommendations.
Use vector database for Sementic Search.
Traditional database search exact matches, vector database finds similar meaning.
Search "CEO Compensation" and find "Executive Salary Packages"
This helps AI find relevent information from massive datasets
This powers every smart search system you have ever used.
Agent frameworks let's AI browse websites, run code, send emails, use tools.
They have goals and can break them down into steps
This changes everything, agents don't just answer "how to book a flight", they book the flight for you.
Process text, image, audio and video all together.
GPT4V don't just see images, it describes them
Wisper converts speach to text
The worlds isn't just text, multi-model AI can understand and create any type of context.
Lets AI trigger your APIs, query database and send message.
"book a meeting" becomes actual a calander integration.
Turn AI from smart chatbot into capable digital assistant
this is the difference between impresive demo vs useful tool.
Instead of jumping to answers, AI explains its thinking step-by-step, it improves accuracy on complex problems by 30-50%
It's essential for any task where being wrong has consequences
It helps you verify AI logic and catch errors before they matter.
transformers = text (GPT, Claude)
CNNs = image (object recognition)
RNNs = sequences (time series, speech)
choose the wrong architecture and your performance will decrease
understanding this helps you pick right tool for right job.
VERY IMPORTNANT TO UNDERSTAND THE AI BUSINESS
Instead of trainning form scratch (cost millions), you start with pre-trained models
It's like hiring an expert and teaching them your domain
Small teams can build sophisticated AI without Google sized budget
this is the reason of AI development exploded in last 5 years.
Understanding Reinforcement Learning from Human Feedback (RLHF) trains AI on human preferances
These human rate AI response as good/bad, AI learns maximize scores
This is how ChatGPT learned to be helpful instead of being accurate
This whole concepts explains why modern AI feels so much more useful than earlier versions
Understaning AI safety before deployment
Content filtering, bias detection, alignment techniques
These ensures AI behaves according to human values
Unaligned AI can spred mis information, be manupulated, or cause harm
Every production system needs built in safety guardrails
Models compressed to run on phones, tablets and IoT devices
The data stays on device and responses are instant
It enables AI in situations like poor connectivity and keeps sensitive informations from leaving your control.
Accuracy, Precision, recall, F1-score, perplexity
Human evaluation for subjective tasks
AI can feel impressive but fail on edge cases
A proper evalution catches problems before users catches them
Understanding monitoring for production by tracking
- response times
- error rates
- user satisfaction
- model performance get alerts when AI behaviour changes unexpectedly
AI models degrade overtime without maintaince Monitoring prevent silent failures that destroy user trusts
collect your data, define your task, train your model
most expensive but most powerful option
when prompting, RAG, fine-tuning isn't enough
- Week 1: Prompting (tokens, temperature, system prompts)
- Week 2: Data (embeddings, vectors, RAG)
- Week 3: Applications (APIs, agents, function calling)
- Week 4: Custom solutions (fine-tuning, deployment, monitoring)
converting text into numbers (numerical vectors) where similar meaning have similar numbers. Vector db is where we store these embedding and through semantic search we retrieve. RAG is a pattern of using semantic serach to retrieve embeddings from database, augmenting the prompt with retrieved context amd user's query to LLM where LLM generates the output. Temperature is used by LLM Model during output generation. user's query, system prompt, retrieved embedding are all considered as input token and generated output is considered as output token.
APIs let applications send prompts to AI models and receive responses programmatically. Moves AI from chat interface into your production systems (CRM, analytics, support). Pay per token (input + output).
LLM detects when a function should be called and extracts structured parameters. YOUR code validates the request and executes the actual function. LLM never directly accesses your systems—you stay in control. Turns "book a meeting with Sarah tomorrow at 2pm" into validated function calls.
Give AI a goal → it breaks into steps → calls functions → adapts based on results. Can iterate, revise plans, and use multiple tools to achieve objectives. Example: "Analyze AWS costs" → queries billing → detects spike → investigates → finds root cause → generates report. More autonomous but requires safety guardrails (step limits, approval workflows).