Ben Hall & Barbara Fusinska
summary
Explains tokenization, attention and Transformers with a tiny Shakespeare demo that converges after 500 iterations, then scales to production LLMs trained on web and curated sources (Common Crawl, Reddit, Stack Overflow, leaked bundles like “ligen”) with roughly 170 trillion parameters, trained over ~100 days on ~25,000 GPUs at an estimated cost of 63 million dollars. Covers practical stacks: Copilot, LangChain, Sourcegraph Codi, embeddings + vector database retrieval (RAG) and local/open models (LLaMA) and shows which approaches—fine-tuning, prompt engineering, or retrieval—actually work.
AI-written summary · about
chapters (12)
- 0:00Introduction to LLMs
- 2:59Practical LLM applications
- 7:11How LLMs work internally
- 14:29Understanding attention
- 18:41Training data and costs
- 25:05Customizing models
- 29:09Prompt engineering strategies
- 33:52Managing creativity
- 35:46Retrieval Augmented Generation
- 40:23Agents and automation
- 44:27Local LLM execution
- 47:43Conclusion
related talks
43mLarge Language Models for Devs: from zero to your first LLM application
34mTerrible tokenizer troubles in large language models
23mPanel
25mTraining Large Language Models in the Open
01h 07mEmpowering Security Teams with Generative AI
30mLearning to wield large language models
40mTechniques to avoid common LLM pitfalls
31mExploring LLMs across clouds – A Year in the Life