2027: The Year AI Becomes Smarter Than Us — Transcript

Full transcript

1. 0:55Look at this timeline. It’s a ticking time bomb.
2. 0:58Each year shows us a different AI  experts’ estimate for when we’ll
3. 1:02develop a technology that will change everything
4. 1:05Artificial General Intelligence, or AGI is  an AI that is smarter than the average human.
5. 1:11Elon Musk thinks AGI is coming by 2026 And Dario Amodei the CEO of Anthropic
6. 1:17thinks AGI is 2-3 years away GoogleDeepmind Chief AGI Scientist
7. 1:22Shane Legg thinks we’ll have AGI by 2028, Sam Altman and Jensen Huang estimate AGI by 2029
8. 1:39The consensus is moving towards  AGI in the next 1-10 years.
9. 1:42Just 4 years ago, expert forecasters and AI  scientists predicted that AGI was 20 years away.
10. 1:49Now, they think it’s just 3 years away.
11. 1:52You can see just how surprised  people were by GPT-3 and 4.
12. 2:12Now, an AI has scored 98th percentile  on a Mensa (genius) admission test. Yes,
13. 2:20So what’s going on?
14. 2:22Why have these timelines shrunk so dramatically?
15. 2:25It started about 8 years ago, when AI  started beating humans in specific tasks
16. 2:30way faster than predicted.
17. 2:31In 2016 experts predicted that AI wouldn’t  master the Japanese board game Go until
18. 2:372028. But that same year DeepMind's AlphaGo  beat one of the best Go players in the world,
19. 2:44Lee Sedol, a full twelve years ahead of schedule.
20. 2:47Sedol retired in 2019, saying
21. 2:55But what about the more complex game  Starcraft, which has far more variables?
22. 3:00Experts predicted human-level  performance by 2022. DeepMind’s
23. 3:04AlphaStar achieved Grandmaster status in 2019. If anything this trend is speeding up.
24. 3:10Experts predicted that AI could write  a high-school essay by 2026. GPT-3 did
25. 3:15that in 2020, getting the equivalent to  an American high-schooler’s C+ grade.
26. 3:20In 2022 GPT 3.5 moved on to legal exams  and scored 213 out of 400 on the Bar Exam,
27. 3:28putting it in the bottom 10% of human test-takers.
28. 3:31And just months later GPT-4 scored  298, putting it in the top 10%.
29. 3:39“that’s better than the human rate”
30. 3:41The same technology experts thought would struggle
31. 3:44with writing high school essays is now  beating most American law graduates.
32. 3:49So WHY are we smashing through past predictions? Well that’s where things get… worrying.
33. 3:55It scares the hell out of me […] it’s  capable of vastly more than anyone knows,
34. 3:59and the rate of improvement is exponential. The last big breakthrough in AI capabilities
35. 4:05was when transformer technology was  introduced in 2017, which let AI
36. 4:09models process all parts of the input data  simultaneously rather than sequentially.
37. 4:14This is what made ChatGPT possible. Since then, AI progress has been exponential,
38. 4:19as researchers figured out that  the more resources they committed
39. 4:22to AI - the more compute and letting AI  models read and analyse more and more of
40. 4:26the internet - the better their AI  models got at virtually all tasks.
41. 4:30It’s reading more of the internet, and throwing  more computers at the stuff it’s read on the
42. 4:35internet. And out pops out… suddenly, it  knows how to explain jokes. You're like,
43. 4:39"Where did that come from?" Or now it knows how to play chess
44. 4:42For now, more money and resources directly  translates to more intelligent AIs.
45. 4:51The CEO of Microsoft AI Mustafa Suleyman  predicts humanity will be training AI models
46. 4:57that are 1,000x larger in the next 3 years. So it looks like this huge investment in
47. 5:02transformer-era AI has put  us on a fast AI timeline.
48. 5:06But AI capabilities will increase even  faster if this massive investment gives
49. 5:11us AGIs that are capable of  recursive self-improvement.
50. 5:15Imagine 1000s of highly skilled AI ‘scientists’  working on the next version of themselves.
51. 5:21At first, the improvements might  be measured in weeks or months.
52. 5:25But once they produce new AI models  with superhuman intelligence,
53. 5:29with the same goal of improving AI  capabilities… we’ll get the same
54. 5:33advancements in a single evening as we got in the  past few years of humans working on the problem.
55. 5:39Current AI models already  write code, and AI improving
56. 5:42through self-learning has happened before. Alpha4o Zero beat the world’s best Go player Lee
57. 5:48Sedol without using any data from human matches. It trained by just playing against itself,
58. 5:54running tests over millions of  simulated games until, after 48hrs,
59. 5:59it played better than any human on Earth. The CEO of NVIDIA says we’re "obviously"
60. 6:13less than 5 years from AGI because  current AI is already helping to create
61. 6:19the chips its future iterations will run on. Sam Altman is so convinced by the current rate
62. 6:33of progress that he’s looking beyond AGI  at ASI, an Artificial Super Intelligence.
63. 6:45That will be a system that can eclipse all  humans combined at all cognitive tasks,
64. 6:50just like Stockfish destroys us at chess. So why aren’t MORE people talking about this?
65. 6:56This comes down to human psychology. It’s hard for humans, even experts,
66. 7:01to build an intuition for exponential growth.
67. 7:04Take the rate of human economic  growth and innovation for example.
68. 7:08The world in 2024, with smartphones, nuclear power  and space travel would make zero sense to someone
69. 7:14from 1750. It would blow their mind because  so much has changed in those last 250 years.
70. 7:21But a person from 1500 would be a lot less  surprised if they were transported forward in
71. 7:26time to 1750: people were still riding horses,  fighting with swords, and sailing in boats.
72. 7:32The reason for that difference is  because the rate of growth isn't linear;
73. 7:36it compounds over time. Inventions  like the microscope or computers
74. 7:41make it easier for new inventions  to happen at an even faster rate.
75. 7:45So it's not simply that we're making strides;  it's that each stride is longer than the last.
76. 7:52Understanding this compounding  effect is crucial for grasping
77. 7:56how human progress has accelerated over time. But despite us living through and benefiting
78. 8:01from exponential economic growth, humans  - even experts - intuitively expect the
79. 8:06speed of growth in the next 30 years to look  a lot like the last thirty years, and project
80. 8:12a steady, linear progression into the future. We draw a straight line upwards from the past,
81. 8:19assuming that we’ve already reached top  speed and are going to carry on cruising,
82. 8:24without realising that exponential growth actually  looks like infinitely increasing acceleration.
83. 8:30And exponential growth happens fast. Let’s say you have a piece of algae in a pool
84. 8:35that doubles in size daily. If it takes 30 days to  fully cover the pool, on what day will the pool be
85. 8:41covered? Not day 15 - day 29. One day before you  have a pool full of algae it’s only half-full.
86. 8:51Right now all indications are that for  future AI progress we’re at “day 29”,
87. 8:56at the knee of this graph, just before  things get real crazy, real fast.
88. 9:01The unintuitive nature of exponential growth  is why AI timelines have been consistently
89. 9:05underestimated, even by some experts  who think we’re on an exponential curve.
90. 9:10Because it just feels wrong to say that the  same advancements that used to take decades,
91. 9:15will soon take place in a few months,  then every week, then every day.
92. 9:21So we’re probably making TWO mistakes about AI:
93. 9:24Not only are we underestimating  just how soon AGI will arrive...
94. 9:28We’re also underestimating (or just being  unable to imagine) what exponential growth
95. 9:33in the intelligence of another  species looks and feels like.
96. 9:37We can’t intuitively GRASP just how much more  intelligent an advanced AI will be than us,
97. 9:44once that exponential growth takes us  beyond human-level AGI, soon after 2030.
98. 9:49I’d love to describe in more depth what  superintelligent AI will be like when
99. 9:53it passes human levels of intelligence. But that’s kind of the point. "that would
100. 9:57be like asking prehistoric humans  to predict modern day space travel.
101. 10:02They’re just not going to know. The closest we can get is looking at the
102. 10:05most advanced chess AI model, and then imagining  that it doesn’t just beat us at chess but at
103. 10:11everything. And that it’s interested in playing  and winning games that we don’t even know exist.
104. 10:18OK, so should we be worried?
105. 10:20Heads of frontier AI labs think AIs may  become 1,000x more powerful than they are
106. 10:26now in 3-7 years. How do you control  an entity that’s smarter than you?
107. 10:32There are no historic examples of a species  being controlled by a less intelligent lifeform.
108. 10:38We don't even know how to  control current, primitive AIs.
109. 10:42Sure we can control what they say - but like  any dictator will tell you, controlling what
110. 10:48someone says is very different to controlling  what they feel, think, and plan in secret.
111. 10:55Even AI researchers don’t really know  how their own invention operates.
112. 10:59It’s a black box that they give inputs to,
113. 11:02and tweak to get the right outputs without  really knowing what it is they’re tweaking.
114. 11:07It’s better to think of AI as a fundamentally  alien intelligence. AI safety experts have
115. 11:13warned us for years about the danger of  developing technology we don’t understand.
116. 11:1950% of AI researchers believe there's a 10% or  greater chance that humans go extinct from our
117. 11:25inability to control AI. And we already  said you would not get on that plane,
118. 11:32if that was the chance that the engineers who  built the plane told you was going to happen.
119. 11:37Now like I said I don’t know what it’d be  like to interact with a superhuman AGI,
120. 11:42or what that AGI would want to do to you… but  if you want my best guess watch this video next.