Iron Man or Ultron: Is AI here to help us or hurt us?
summary
Argues that large language models are a next-word predictor not conscious, citing the November 2021 base models and showing how Bing/Sydney uses hidden prologues and gates to moderate outputs. Demonstrates how temperature controls randomness and how probabilities expose alternate completions; warns that DaVinci (≈174B params) behaves very differently from Ada (≈2.7B). Recommends using the smallest adequate model or edge deployment, keeping user interactions out of training, and using tools like TypeChat to constrain JSON outputs; notes GitHub Copilot only sees open tabs.
AI-written summary · about
chapters (20)
- 0:00Opening remarks
- 1:30Personal reflections
- 2:58Childhood tech access
- 8:14Luck equals preparation
- 11:01AI’s rise and perception
- 13:30Ethics in AI development
- 17:24AI decision-making
- 23:54Temperature settings
- 27:03Guardrails in AI systems
- 31:01AI limitations and training
- 33:29Diverse teams for AI
- 35:58Large language models
- 39:12Edge computing models
- 40:57TypeChat responses
- 43:32Ethical AI boundaries
- 46:18AI as confident bullshitter
- 48:27Token length and context
- 51:17AI for summarization
- 53:05Iron Man vs Ultron
- 55:13Ethics and AI future
related talks
58mIron Man or Ultron: Is AI here to help us or hurt us?
43mBeyond the AI Hype: What's Real, What's Next
58mPanel Discussion: AI- Friend or Foe?
58mKeynote: The past, present, and future of AI for application developers
39mForward Focus: Perspectives on AI, Hype, and Security
30mLarge Language Models and Prompt Engineering for Developers
01h 28mFuture of Software Programming with AI Bots at Hand
24mBring the Power of AI to Your Application