Mathematics & Logic

• Learning by reward

  Nobody tells the learner the right answer, only how well it did. How that one number teaches, and where it misleads.

  • 9min read
  • 10min listen
  • 29questions
• Learning with no one to mark the answers

  Labels are slow and costly to make. So machines learned to set their own questions: group what is alike, rebuild what was squeezed, guess what was hidden.

  • 8min read
  • 11min listen
  • 26questions
• How much is a surprise worth?

  Information is measured by how much a message surprises you, and a machine that learns to be less surprised by text learns a great deal about it.

  • 9min read
  • 11min listen
  • 31questions
• Changing your mind by the numbers

  Start from what you already knew, weigh the new clue, and end up exactly as sure as the evidence allows.

  • 9min read
  • 10min listen
  • 28questions
• Thinking in lists of numbers

  Vectors and matrices, the school maths that search engines, recommendations and every neural network quietly run on.

  • 8min read
  • 10min listen
  • 29questions
• Secret codes

  A cipher hides a message's letters but not its patterns, and the patterns are what give it away.

  • 8min read
  • 11min listen
  • 29questions
• Transformers

  Attention, next-word prediction and scale — and the gap between sounding right and being right.

  • 9min read
  • 9min listen
  • 30questions
• Deep learning

  Nobody writes the rules. A network of simple calculations tunes itself on examples until its answers come out right.

  • 8min read
  • 10min listen
  • 33questions
• The bitter lesson

  Seventy years of AI research, one uncomfortable pattern: methods that scale with computing beat methods built on what we know.

  • 7min read
  • 10min listen
  • 24questions