full transcript

From the Ted Talk by Max Tegmark: How to keep AI under control


Unscramble the Blue Letters


Let's look at a simple example of where we first machine-learn an algorithm from data and then distill it out in the form of code that porvlaby mtees spec, OK? Let’s do it with an algorithm that you probably learned in first gdrae, addition, where you loop over the digits from right to left, and sometimes you do a carry. We'll do it in binary, as if you were cnnuiotg on two fingers instead of ten. And we first train a rurneecrt neural network, never mind the details, to nail the task. So now you have this aohrilgtm that you don't understand how it works in a black box dinfeed by a bunch of tables of numbers that we, in nerd speak, call ptaeerarms. Then we use an AI tool we built to automatically distill out from this the leernad algorithm in the form of a Python program. And then we use the formal verification tool known as Dafny to prove that this program correctly adds up any nrbeums, not just the numbers that were in your training data.

Open Cloze


Let's look at a simple example of where we first machine-learn an algorithm from data and then distill it out in the form of code that ________ _____ spec, OK? Let’s do it with an algorithm that you probably learned in first _____, addition, where you loop over the digits from right to left, and sometimes you do a carry. We'll do it in binary, as if you were ________ on two fingers instead of ten. And we first train a _________ neural network, never mind the details, to nail the task. So now you have this _________ that you don't understand how it works in a black box _______ by a bunch of tables of numbers that we, in nerd speak, call __________. Then we use an AI tool we built to automatically distill out from this the _______ algorithm in the form of a Python program. And then we use the formal verification tool known as Dafny to prove that this program correctly adds up any _______, not just the numbers that were in your training data.

Solution


  1. parameters
  2. recurrent
  3. provably
  4. grade
  5. algorithm
  6. numbers
  7. defined
  8. learned
  9. counting
  10. meets

Original Text


Let's look at a simple example of where we first machine-learn an algorithm from data and then distill it out in the form of code that provably meets spec, OK? Let’s do it with an algorithm that you probably learned in first grade, addition, where you loop over the digits from right to left, and sometimes you do a carry. We'll do it in binary, as if you were counting on two fingers instead of ten. And we first train a recurrent neural network, never mind the details, to nail the task. So now you have this algorithm that you don't understand how it works in a black box defined by a bunch of tables of numbers that we, in nerd speak, call parameters. Then we use an AI tool we built to automatically distill out from this the learned algorithm in the form of a Python program. And then we use the formal verification tool known as Dafny to prove that this program correctly adds up any numbers, not just the numbers that were in your training data.

Frequently Occurring Word Combinations


ngrams of length 2

collocation frequency
ai tool 5
provably safe 4
powerful ai 3
human intelligence 2
ai researchers 2
human extinction 2
safe systems 2
learned algorithm 2
stop obsessively 2

ngrams of length 3

collocation frequency
provably safe systems 2


Important Words


  1. addition
  2. adds
  3. ai
  4. algorithm
  5. automatically
  6. binary
  7. black
  8. box
  9. built
  10. bunch
  11. call
  12. carry
  13. code
  14. correctly
  15. counting
  16. dafny
  17. data
  18. defined
  19. details
  20. digits
  21. distill
  22. fingers
  23. form
  24. formal
  25. grade
  26. learned
  27. left
  28. loop
  29. meets
  30. mind
  31. nail
  32. nerd
  33. network
  34. neural
  35. numbers
  36. parameters
  37. program
  38. provably
  39. prove
  40. python
  41. recurrent
  42. simple
  43. speak
  44. spec
  45. tables
  46. task
  47. ten
  48. tool
  49. train
  50. training
  51. understand
  52. verification
  53. works