Deep Dive into LLMs like ChatGPT by Andrej Karpathy

This video is one of the most comprehensive yet accessible introductions to the inner workings of systems like ChatGPT. While it does get a bit technical in places, it is still easy to follow and full of useful takeaways, with plenty of “aha” moments throughout.

This is a must watch material for anyone who is interested in AI, whether you are a developer, strategist, marketer or simply fascinated by artificial intelligence.

Here are some of the topics that the video covers:

• Pretraining data - how massive text datasets teach models grammar, facts, and reasoning through pattern recognition.
• LLM Tokenization - a process of converting words and symbols into numeric tokens the model can process.
• Neural network internals - how the model processes data, learn patterns, and generate outputs.
• LLM Inference - on how model predicts the next word (or token) to form a response.
• Pretraining & post-training - the dual phases where the model first learns, then refines its behaviour.
• LLM Hallucinations - why LLMs sometimes generate fictional information, and how this can be addressed.
• Reinforcement learning (RLHF) - how to align models outputs with human intent through reward-based training.

I would like to thank Andrej Karpathy for this amazing and detailed explanation.

🎥 Watch the full Video