Skip to main content

​
Hello World

This isn’t your typical “Hello World” — we’re diving straight into what makes LLMs powerful. After configuring your OpenAI API key, press Enter to see the magic happen. The code is self-explanatory, so go ahead and modify it to experiment.
Prefer to follow along locally? You can checkout the examples repository and run the code on your machine:
git clone https://github.com/ai-tutorial/typescript-examples
Then navigate to the specific example file and run it with your configured OpenAI API key.

​
How LLMs Process Your Input

Before we write our first prompt, let’s understand what’s actually happening under the hood. The Context Window: Your Working Memory Think of an LLM’s context window like RAM on your computer. Everything you send - your instructions, conversation history, documents - gets loaded into this window. The model can only “see” what fits inside. Current Context Windows (as of January 2025; verify latest on vendor pages below):
  • GPT-4: 128K tokens (~96K words)
  • Claude Sonnet 4.5: 200K tokens (~150K words)
  • Gemini 1.5 Pro: 2M tokens (~1.5M words)
See vendor references: OpenAI models, Anthropic pricing/models, Google Gemini models. Why This Matters: A customer support agent handling a complex case might need:
  • 2K tokens: System instructions
  • 5K tokens: Company knowledge base excerpts
  • 10K tokens: Conversation history
  • 3K tokens: Customer account details
That’s 20K tokens before the LLM generates a single word. Multiply by thousands of requests, and you see why context management matters. Token Economics: input cost=150K tokens∗($10 / 1M tokens)=$1.50output cost=3K tokens∗($30 / 1M tokens)=$0.09total per request=$1.59input\,cost = 150K\,tokens * (\$10\,/\,1M\,tokens) = \$1.50 \newline output\,cost = 3K\,tokens * (\$30\,/\,1M\,tokens) = \$0.09 \newline total\,per\,request = \$1.59

​
LLM Limitations You Must Know

1. Hallucinations: Making Stuff Up LLMs are trained to predict the next plausible token. They’re not fact-checking databases. Famous Failure: Air Canada’s chatbot hallucinated a bereavement discount policy that didn’t exist. The airline had to honor it in court. Cost: Unknown, but significant legal precedent. (BBC, 2024) Why It Happens:
  • Missing information → fills gaps with plausible-sounding text
  • Conflicting instructions → makes judgment calls
  • Outdated training data → invents current information
What Works in Production:
  • Constrain to provided context: “Only use information from these documents”
  • Validate outputs: Check facts against source data
  • Ground the answer in knowledge (throughout the tutorial).
  • Add human review: For high-stakes decisions
2. Non-Determinism: Different Answers Every Time Run the same prompt twice, get different answers. That’s by design (temperature > 0). Try running the same prompt multiple times with different temperature settings to see how the model generates different responses: Production Strategy: Control the creativity of the model with the Temperature parameter.
  • Temperature=0 for minimal creativity tasks (classification, extraction)
  • Temperature=0.3-0.7 for creative tasks (writing, brainstorming)
  • Run multiple times and vote (self-consistency, covered in 1.5)
GPT-5 models do not support the temperature parameter, and using it will raise an error. This breaks backward compatibility with earlier OpenAI models.Instead, GPT-5 introduces a new way to control output variability: reasoning depth, via:To achieve similar results with reasoning effort set higher, or with another GPT-5 family model, try these alternative parameters:
Reasoning depth: reasoning: { effort: "none" | "low" | "medium" | "high" }
3. Recency Bias: Recent Context Matters More LLMs pay more attention to text at the beginning and end of the prompt. Middle content gets “lost.” Heuristic: Prompt structure and placement can affect results. Many teams place the specific query toward the end; validate with your use case. Best Practice for Gemini: See Gemini prompting strategies. 
Critical instructions at START
[Large document content in MIDDLE]
Specific query at END

​
Mental Model: LLMs as Completion Engines

Wrong Mental Model: “The AI understands my intent”
Right Mental Model: “The AI completes patterns it’s seen in training” Example: 
prompt = "The capital of France is"
result = "Paris"
# Not because it "knows" geography, but because it's seen this pattern millions of times

prompt = "The capital of Atlantis is"
result = "unknown" or makes something up
# It hasn't seen this pattern → hallucination risk

​
Practical Implication

When you want structured output, give the LLM a pattern to complete. Compare these two approaches:

​
❌ Antipattern

First, try the vague prompt and see how the LLM responds in prose. Then try the pattern-based approach to get structured output:

​
✅ Best Practice