If you want to learn:
- How do large language models create the illusion of memory and thinking?
- What is chain of thought reasoning and how does it improve AI responses?
- Why do reasoning models outperform standard LLMs on complex problems?
- How do thinking budgets and reasoning traces actually work in modern AI?
- What are the fundamental limitations of LLMs when it comes to true reasoning?
- When should you use reasoning models versus chat models for AI applications?
Then this lecture is for you!
This lecture explores the core mechanisms behind LLM reasoning capabilities and exposes the illusion of thinking in artificial intelligence. You'll discover how the "illusion of memory" works through stateless prompt engineering, where the entire conversation history is sent with each request to create the appearance of memory retention. The lecture demonstrates chain of thought prompting techniques, showing how adding "think step by step" to prompts dramatically improves reasoning outcomes by forcing the model to generate intermediate reasoning traces before final answers.
You'll learn the technical difference between chat models and reasoning models, understanding how reasoning models are trained to output step-by-step thought processes that lead to more accurate results on complex reasoning tasks and benchmark problems. The lecture reveals the surprisingly simple yet effective technique of inserting tokens like "wait" during inference to extend reasoning effort and create longer reasoning traces, explaining how thinking budgets (none, minimal, low, medium, high) control the depth of AI reasoning.
Through concrete examples comparing GPT-4 variants with and without reasoning enabled, you'll see how reasoning models handle trick questions and probability puzzles that standard models fail. The lecture covers the autoregressive token generation process, explaining how transformer models generate text one token at a time and how this architecture enables chain of reasoning improvements. You'll understand the strengths and limitations of reasoning models, including when chat models may actually outperform reasoning models in agentic AI systems, and learn the experimental approach needed to select the right model for your specific use case in machine learning applications.