BBC Learnings · Section 8

Training artificial intelligence

Do you like cooking, Sam? There's a new recipe I've been trying out.

Level: B2 · 100 sentences

Transcript

  1. Do you like cooking, Sam?
  2. There's a new recipe I've been trying out.
  3. It's for frosted oysters.
  4. Frosted oysters.
  5. Sounds unusual.
  6. How do you make it?
  7. Well, you take a pound of chicken and then some cubed pork and half a crushed garlic.
  8. Eh?
  9. I thought you said it was for frosted oysters!
  10. whatever they are.
  11. Yes, that's right.
  12. Now heat it up until boiling and serve with custard.
  13. Sounds disgusting!
  14. Who on earth told you that recipe?
  15. It's not who told me, Sam, but what.
  16. In fact, that recipe was made by computers using artificial intelligence,
  17. or AI, which is the topic of today's programme.
  18. In real life, AI is making huge progress,
  19. from car sat-navs to detecting cancer cells.
  20. But as you can see from that revolting recipe,
  21. things don't always go according to plan.
  22. So just how intelligent is artificial intelligence?
  23. I mean, it definitely needs some cooking lessons.
  24. Right.
  25. AI is not as intelligent as we tend to think.
  26. AI programmes use artificial brain cells to roughly imitate real brain cell activities.
  27. But they're still a long way behind human levels of intelligence.
  28. And that's my quiz question.
  29. In terms of brain cell count,
  30. what level of intelligence is AI currently working at?
  31. Is AI as smart as a a frog,
  32. b an earthworm or c a bumblebee?
  33. Well, I don't think any of those are good cooks either,
  34. to be honest, but I'll say c a bumblebee because at least they can make honey.
  35. Nice guess, Sam.
  36. We'll find out the answer later,
  37. but first let's find out more about how AI misunderstandings like the oyster recipe can happen.
  38. Janelle Shane is the author of You Look Like a Thing and I Love You,
  39. in which she tells her amusing experiences and bizarre experiments with AI.
  40. You Look Like a Thing and I Love You.
  41. That's a strange title for a book, Neil.
  42. Yes, it's another example of AI miscommunication.
  43. The book title is what AI produced
  44. when asked to write chat-up lines remarks men and women make
  45. to start up a conversation with someone they don't know but find attractive.
  46. Here she is talking to the BBC World Service programme More or Less.
  47. So machine learning is what most programmers mean when they say, AI.
  48. In the program that we're used to, if you want to have a computer program solve a problem,
  49. you have to have a human programmer write down exhaustive step-by-step instructions on how to do everything.
  50. But with machine learning, you just give it the goal,
  51. and then the program figures out via trial and error how it's going to solve that problem.
  52. So even though we're talking about machines learning for themselves,
  53. there still need to be humans involved at the start of the journey.
  54. This human teaching is done by computer programmers, people who write,
  55. or code the computer programmes used by AI.
  56. Right.
  57. These programmers write algorithms a set of rules or procedures to be followed in problem-solving exercises.
  58. So, for example,
  59. the AI that wrote that oyster recipe read thousands of other recipes before coming up with its own version.
  60. In other words, artificial intelligence uses a process of trial and error
  61. repeating the same task over and over until finding the most successful way.
  62. Only in the case of the oyster recipe,
  63. there was more error than trial.
  64. Well, according to Janelle Shane,
  65. we can learn a lot about something by seeing how it goes wrong.
  66. Here she is talking about an AI which had been told to solve maths problems.
  67. It seemed to be that it was getting scored on how many
  68. wrong answers that it got and it was supposed to be minimizing the number of wrong answers.
  69. And just by a stroke of luck,
  70. as part of its trial and error flailing around,
  71. one of the flails it did accidentally deleted the solutions list
  72. and then it and everybody else got a perfect score.
  73. So AIs learn by minimizing their errors,
  74. reducing them as much as possible.
  75. And sometimes these algorithms only discover the right answer by a stroke of luck,
  76. when something unexpected happens by good luck or chance.
  77. It seems to me that they're not so intelligent after all.
  78. Well, let's settle it once and for all by answering today's quiz question.
  79. Remember I asked you how intelligent AI was in terms of brain cell count and you said as intelligent as...
  80. I said C, a bumblebee.
  81. Well, here's Janelle again with the answer.
  82. If you're looking at rough computing power,
  83. the algorithms we're working with are probably somewhere around the level of an earthworm.
  84. Ah, so the correct answer was B, as clever as an earthworm.
  85. No wonder AIs can't cook.
  86. Or take a maths test without cheating.
  87. In this programme, we've been looking at artificial intelligence,
  88. or AI, and seeing how programmers,
  89. that's people who write instructions for computers to follow,
  90. create algorithms, sets of rules used in problem solving.
  91. AI learns through trial and error,
  92. repeating the same activity again and again
  93. until discovering the best way and minimizing,
  94. so reducing as much as possible,
  95. the number of errors it makes.
  96. And success can be the result of a stroke of luck
  97. when something unexpected happens purely by chance,
  98. although so far that hasn't helped AIs to write good chat-up lines
  99. the flattering remarks people make to get to know someone they find attractive.
  100. And AIs don't know much about cooking oysters either.