Forecasting Forecasting

Aidan Lyon, DelphiCloud, aidan@delphi.cloud

• Aggregative Contingent Estimation Program (IARPA)
• Australia, preventing agricultural diseases, etc.
  • BigComputer is pretty good at answering questions but which questions should we ask?
  • Hajek, Fridge magnet technique - write all revelant words on pieces of paper and play with rearranging them.
    • Automation with taking the concept cloud from an area and making sequences out of concepts.
• EEG headset (400$ headset is now accurate enough to play videogames)
  • Emotions change our judgements (anger --> overconfidence).
  • We can use EEG to compensate for that.
• What about extreme events?
  • We hire philosophers to think about this.
  • What should we include in our probability space and how do we make sure we don't miss important and possible stuff?
• Humans have this data-efficiency and creativity magic that we can't replicate with AI at the moment.
  • Ayahuasca improves divergent thinking and impairs convergent thinking.
  • Having a body and easy access to the world makes learning easier.
  • Computers can't reframe the problem -- but is this really true? we might be overestimating our cognitive flexibility.
  • Creativity: AlphaGo did make some creative moves that suprised everyone.
  • If the computer can't explain an obscure decision, we won't trust it.
    • We're working on it.
    • If it's really good, maybe we'll trust it anyway.
  • We are very data-efficient but that's because we have lots of things already hardcoded in us by evolution. This is unlikely to be a lasting advantage.
  • In any case, the AI is different from us -- we are not creating artificial humans but perhaps artificial aliens. And this different is valuable. If we have teams of humans + AIs, they are more diverse.
• Summary
  • For now H + C > H, C
  • Humans are better at asking questions, especially when there's little data.
  • Things are moving and some of those things might change.
  • For now we can work on improving H + C.

Q/A

• Explainability is not always needed: example with native speaker as broadcaster vs. language teacher.
• Sometimes human explanations are just rationalizations.
• Using analogies for creativity (transfer learning)
• Creative machines?
• Fake explanations? (that's how humans often do it anyway)
• Black swans are sometimes "black elephants" (we know about them, we're just not paying attention). Perhaps machines can help us find those.