You don’t need to build the next AI model. You need to know how to...
Millie Marconi@MillieMarconnni
76 views
Aug 18, 2025
~3 min read
Advertisement
1
You don’t need to build the next AI model.
You need to know how to use them.
Here’s how to get dangerous with AI in 30 days even if you're starting from zero:
You need to know how to use them.
Here’s how to get dangerous with AI in 30 days even if you're starting from zero:
2
Most beginners start by:
- Building chatbots
- Downloading agent frameworks
- Playing with APIs
But they don’t understand tokens, prompts, or context windows.
That’s like trying to write a novel before learning the alphabet.
- Building chatbots
- Downloading agent frameworks
- Playing with APIs
But they don’t understand tokens, prompts, or context windows.
That’s like trying to write a novel before learning the alphabet.
3
The AI hierarchy
Start here:
AI → ML → DL → LLMs
AI = mimics intelligence (rules or learning)
ML = learns from data
DL = uses neural networks
LLMs = predicts the next word at scale
If this is fuzzy, everything else will confuse you later.
Start here:
AI → ML → DL → LLMs
AI = mimics intelligence (rules or learning)
ML = learns from data
DL = uses neural networks
LLMs = predicts the next word at scale
If this is fuzzy, everything else will confuse you later.
4
Week 1: Learn how AI thinks
Tokens = the currency of AI
Context window = AI’s short-term memory
Temperature = controls randomness
System prompt = defines behavior
This week alone will fix 90% of your confusion.
Tokens = the currency of AI
Context window = AI’s short-term memory
Temperature = controls randomness
System prompt = defines behavior
This week alone will fix 90% of your confusion.
5
Tokens + Examples
Think in tokens:
“hello world” = 2 tokens
“supercalifragilistic” = 8 tokens
1 token ≈ 4 characters (not 1 word!)
If you ignore tokens, you’ll hit hidden limits, break prompts, and waste money.
Think in tokens:
“hello world” = 2 tokens
“supercalifragilistic” = 8 tokens
1 token ≈ 4 characters (not 1 word!)
If you ignore tokens, you’ll hit hidden limits, break prompts, and waste money.
6
Context + Temperature
Context window ≠ memory.
It’s the size of the conversation.
Claude = 200k tokens. GPT-4 = 128k.
Temperature tip:
0 = boring but precise
0.7 = creative and helpful
2 = chaotic mess
Most people set this wrong.
Context window ≠ memory.
It’s the size of the conversation.
Claude = 200k tokens. GPT-4 = 128k.
Temperature tip:
0 = boring but precise
0.7 = creative and helpful
2 = chaotic mess
Most people set this wrong.
7
Week 2: Learn how AI understands meaning
Embeddings
Vectors
RAG (retrieval augmented generation)
Vector DBs
Embeddings
Vectors
RAG (retrieval augmented generation)
Vector DBs
8
Embeddings = meaning
AI doesn’t store text.
It converts "the cat sat on the mat" into a list of 1,536 numbers.
Similar meanings = similar numbers.
This powers:
- Smart search
- Recommendations
- Chat with your data
It's the foundation of "AI memory".
AI doesn’t store text.
It converts "the cat sat on the mat" into a list of 1,536 numbers.
Similar meanings = similar numbers.
This powers:
- Smart search
- Recommendations
- Chat with your data
It's the foundation of "AI memory".
9
RAG vs Fine-Tuning
RAG = plug AI into your docs
Fine-tuning = retrain AI on new patterns
Use RAG first it’s cheaper and more flexible.
Only fine-tune when:
- Prompts fail
- You need tone/style consistency
- You have LOTS of labeled data
RAG = plug AI into your docs
Fine-tuning = retrain AI on new patterns
Use RAG first it’s cheaper and more flexible.
Only fine-tune when:
- Prompts fail
- You need tone/style consistency
- You have LOTS of labeled data
10
Week 3: Learn how to build AI tools
APIs = connect AI to software
Agents = let AI take actions
Function calling = automation engine
This is where you go from “chatting” → “building useful AI apps”
APIs = connect AI to software
Agents = let AI take actions
Function calling = automation engine
This is where you go from “chatting” → “building useful AI apps”
11
Agents + Function Calling
Agents = AI with goals
Function calling = trigger real actions
Used together:
“Book me a Zoom call tomorrow at 2PM”
→ AI finds a slot, calls the calendar API, and books it.
This is the future of assistants.
Agents = AI with goals
Function calling = trigger real actions
Used together:
“Book me a Zoom call tomorrow at 2PM”
→ AI finds a slot, calls the calendar API, and books it.
This is the future of assistants.
12
Week 4: Build custom solutions
- Fine-tuning
- Edge deployment
- Model evaluation
- RLHF
- Monitoring
Most people skip to this stuff and get lost.
But now you’re ready.
- Fine-tuning
- Edge deployment
- Model evaluation
- RLHF
- Monitoring
Most people skip to this stuff and get lost.
But now you’re ready.
13
AI Mastery Roadmap
Week 1 → Master prompting
Week 2 → Learn how AI understands meaning
Week 3 → Build real applications
Week 4 → Customize and scale AI for your use case
This order = foundations first, complexity last.
Week 1 → Master prompting
Week 2 → Learn how AI understands meaning
Week 3 → Build real applications
Week 4 → Customize and scale AI for your use case
This order = foundations first, complexity last.
14
LLMs change weekly.
Today it’s GPT-4o, Claude 3, Gemini 1.5...
Tomorrow it’ll be something new.
But the fundamentals stay the same:
• Tokens
• Embeddings
• Context
• Inference
• Training
• Safety
Learn these. Tools come and go principles endure.
Today it’s GPT-4o, Claude 3, Gemini 1.5...
Tomorrow it’ll be something new.
But the fundamentals stay the same:
• Tokens
• Embeddings
• Context
• Inference
• Training
• Safety
Learn these. Tools come and go principles endure.
15
I hope you've found this thread helpful.
Follow me @Yesterday_work_ for more.
Like/Repost the quote below if you can:
Follow me @Yesterday_work_ for more.
Like/Repost the quote below if you can:
Millie Marconi@MillieMarconnni
You don’t need to build the next AI model.
You need to know how to use them.
Here’s how to get dangerous with AI in 30 days even if you're starting from zero:
You need to know how to use them.
Here’s how to get dangerous with AI in 30 days even if you're starting from zero:
