Skip to content
All cheat-sheets

alphaPlan · How Computers & the Internet Work · Cheat-sheet 02

Artificial intelligence

AI is a growing toolbox of narrow systems with a long history of hype and winters, and its output deserves the same questioning mind you would give any powerful source.

Ideas to remember

  1. 01Two honest views of AI: the agent view (building something that senses and acts a little like a thinking creature) and the tool view (a set of tools for real jobs).
  2. 02Every impressive AI you meet today is narrow, very good at one thing and helpless outside it; a general artificial intelligence does not exist.
  3. 03Newer general-purpose models sit between the two boxes: far broader than narrow AI, nowhere near general, and jagged in what they get right.
  4. 04Excitement about AI comes in waves; when promises outran the technology, funding dried up in cold spells called AI winters.
  5. 05The Turing test judges only behaviour: whether an interrogator can tell a machine from a person through written messages.
  6. 06AI learns from data made by people, so human bias hides in the data and a confident answer is not the same as a fair one.

Words

Agent
Something that senses its surroundings and acts within them; the inspiration is us.
Narrow AI
A system very good at one specific thing and helpless outside it.
General artificial intelligence
A machine that could do everything a human can do; we do not have one.
AI winter
A cold spell when interest and money drained away and progress seemed to stall.
Turing test
A machine passes if an interrogator exchanging written messages cannot tell it from a person.
Deepfake
A photo or video of something that never happened, convincing enough to fool a quick glance.

Do this

  • Ask where content came from: who made it, who trained it, and what they might want you to believe or do.
  • Be slow with the startling; check shocking content before you pass it on.
  • Look for a second source, since a real event usually leaves more than one trace.
  • Keep breadth separate from reliability: a system that can attempt many things is not proof it does any of them well.
  • Name one everyday AI you met today and say why it is a narrow tool, not a general mind.

Watch out

  • Thinking a computer is neutral: a hiring model trained on skewed past decisions keeps favouring the same kind of person.
  • Face recognition that helps find a missing child can, turned on a whole population, monitor ordinary people who did nothing wrong.
  • A smooth, certain answer is not proof, only a well-formed guess; the machine can be confidently wrong.

Found something unclear, outdated or improvable? Suggest an improvement