Embeddings, Transformers, RLHF: Three key ideas to understand ChatGPT
summary
Explains that large language models first learn dense embeddings (word vectors from Word2Vec-style masked learning) that let arithmetic-like analogies emerge, then use the attention mechanism of Transformers to weight distant context, and finally undergo two practical fine-tunings: supervised instruction tuning on ~1–10 million high-quality examples and RLHF where humans rank outputs to teach helpfulness. Claims base models are trained on trillions of web tokens over months and tens of millions of dollars and can produce toxic outputs if raw data quality is poor.
AI-written summary · about
chapters (5)
- 0:00Introduction to ChatGPT
- 3:20Understanding embeddings
- 13:27Transformer architectures
- 18:42Training pipelines
- 22:24Conclusion and closing
related talks
01h 01mHow ChatGPT Works
01h 00mInside GPT – Large Language Models Demystified
46mConversations with ChatGPT: illusion or reality?
01h 01mKeynote: Transformers: The Rise of ChatGPT
27mEmbeddings, Transfer Learning and More: LLM Tailored to You
47mInside GPT – Large Language Models Demystified
51mLarge Language Models for Real-World Applications - A Gentle Intro
35mChatGPT is lying, how can we fix it?