Ep: 04 How LLMs Work: Tokens & Next-Word Prediction in 6 Minutes

How do LLMs like ChatGPT actually work? You type "Hello", but ChatGPT sees 13225. In 6 minutes you'll see what tokens are, how text becomes numbers, how next-token prediction writes every answer, why the same question can cost more in Hindi, what temperature and the context window do, and the 3 interview questions you should be able to answer. Episode 4 of FDE Journey: the step-by-step roadmap from zero to Forward Deployed Engineer. Every episode ends with interview questions. ⏱️ Chapters 0:00 ChatGPT never reads your words 0:29 Pause and guess: strawberry 1:00 What is a token? 1:41 Tokens become numbers 2:07 Next-token prediction: the one trick 2:42 English vs Hindi: same question, more tokens 3:25 Temperature: why answers change 3:59 The context window 4:27 FDE Reality Check: tokens in a real job 5:04 3 interview questions 5:36 Recap and what's next 🎯 Interview questions covered What is a token? What does an LLM actually predict? What is a context window? 🔬 How we checked the numbers Every token split, count and ID was produced with OpenAI's open-source tokenizer, tiktoken (encoding o200k_base, used by GPT-4o): https://github.com/openai/tiktoken • "Hello" = token 13225 · "Hello, how are you?" = 6 tokens · the same question in Hindi = 9 tokens (this sentence only) • "strawberry" on its own = st · raw · berry Other models (Gemini, Claude) use their own tokenizers: the counts differ, the idea is the same. ▶️ Previous: History of AI: From ELIZA to ChatGPT → [   • EP 03: History of AI: From ELIZA to ChatGP...  ] ▶️ Next: How prompts control an LLM (Episode 5) 📸 Instagram:   / futurereadywithgaurav   💬 Comment your name and I'll tell you how many tokens it is. 📤 Know someone preparing for AI interviews? Send them this. 🔔 Subscribe to Future Ready with Gaurav. Learn Today. Lead Tomorrow. ℹ️ Narration: AI voice. Some visuals are AI-generated. Next-token percentages and the temperature answers are illustrative. #LLM #ChatGPT #GenerativeAI