AI Alignment ⚐

up:: AI ⚐

What are AI

There are two kinds of AI: Narrow AI and Artificial General Intelligence (AGI)

• Narrow AI powerfully sort patterns in limited domains
• AGI: A system that can improve itself that is not limited to one domain. AGI is a system that can improve itself and is virtually unbound by a domain

How do machines learn

Machine learning uses data to answer questions. An agent gets trained (uses data) to make predictions (answers questions).

• Supervised learning classifies labeled data

• Unsupervised learning finds patterns in unlabeled data

• In reinforcement learning, agents receive rewards based on their behaviour

• Additional ways of learning

  • Symbolic Artificial Intelligence
  • Deep learning
  • Bayesian networks
  • Evolutionary algorithms

Problems: AI Alignment Theory

The real risk with AI isn’t malice but competence. A super-intelligent AI will be extremely good at accomplishing its goals, and if those goals aren’t aligned with ours we’re in trouble. – Stephen Hawking

Utility Functions

• A utility function is the thing an agent want to maximise Additional links
• Instrumental Goals are behaviours that help the agent to maximise their utility function
  • A common instrumental function is to prevent destruction
  • A common instrumental function is to prevent change
  • A common instrumental function is self-improvement

Problems with AI

• Corregibility: A common instrumental function is to prevent change
• Reward Hacking TK: (the boat game)
• Specification Problems TK
  • it’s more likely that if we don’t do anything, ai doesn’t do what we want it to do
• negative side effects, safe exploration

AI and human society

• AI and the future of work TK
  • LL070 just a draft
• Sufficiently evolved AI could seem divine just a draft

Important Players

DeepMind TK

---

source: Computerphile YouTube playlist on AI