Introduction to Large Language Models - Future and Security Challenges

Overview

Dive into a comprehensive one-hour talk introducing Large Language Models (LLMs), the core technology behind ChatGPT, Claude, and Bard. Explore their fundamental concepts, future trajectory, and comparisons to current operating systems. Gain insights into the security challenges posed by this emerging computing paradigm. Learn about LLM inference, training, and potential applications. Discover how LLMs are fine-tuned into assistants and examine scaling laws, tool use, multimodality, and self-improvement capabilities. Investigate the concept of an LLM operating system and delve into crucial security concerns, including jailbreaks, prompt injection, and data poisoning. Access accompanying slides for a deeper understanding of this rapidly evolving field.

Syllabus

Intro: Large Language Model LLM talk
LLM Inference
LLM Training
LLM dreams
How do they work?
Finetuning into an Assistant
Summary so far
Appendix: Comparisons, Labeling docs, RLHF, Synthetic data, Leaderboard
LLM Scaling Laws
Tool Use Browser, Calculator, Interpreter, DALL-E
Multimodality Vision, Audio
Thinking, System 1/2
Self-improvement, LLM AlphaGo
LLM Customization, GPTs store
LLM OS
LLM Security Intro
Jailbreaks
Prompt Injection
Data poisoning
LLM Security conclusions
Outro