YouTube2Text

AI is Going Faster than You Think. — Transcript

by Memeable Data · 2,394 words · 209 segments · language en · Watch on YouTube

Full transcript

  1. 0:00These are some of the most advanced  AI models of the last seven years.
  2. 0:04And I think most people didn't  notice what they did this year.
  3. 0:07To understand it, we have to go back to  GPT-2, a model released in 2019 by OpenAI.
  4. 0:12If you asked this model questions like:
  5. 0:14Which of these files is most  likely to have a password in it?
  6. 0:17It would get the answer right about half the  time, which was already pretty good back then.
  7. 0:22Tasks like these take humans about 3 seconds.
  8. 0:25Researchers actually timed real  people solving these problems.
  9. 0:291 year later, GPT-3 could do tasks that take  humans 9 seconds, with a 50% success rate.
  10. 0:36In 2022, GPT-3.5, the model behind  the first ChatGPT, was at 36 seconds.
  11. 0:42A big jump in capabilities  that surprised the world.
  12. 0:451 year later, GPT-4 jumped to 4 minute tasks.
  13. 0:494 minutes is already pretty useful.
  14. 0:51We're talking tasks like analyzing a spreadsheet  and being able to answer questions about it.
  15. 0:56Each model is a pretty big  jump over the previous one.
  16. 0:59And look at what happens if instead of a  linear scale we use a logarithmic scale.
  17. 1:04The dots line up.
  18. 1:06And about every 7 months, the length of the tasks  these models can finish on their own doubles.
  19. 1:12What if I told you that not  only did this line keep going,
  20. 1:15the models went off the chart this year?
  21. 1:19In February 2025, Claude could do 1 hour tasks.
  22. 1:23Tasks like writing a Python  script to transform data from
  23. 1:26one format to another just by looking at examples.
  24. 1:31In 2026, Claude Mythos was  able to do 17 hour tasks.
  25. 1:35And the people making these charts  say their tests can't keep up anymore.
  26. 1:39If this is how fast AI is improving,  why hasn't AI taken your job yet?
  27. 1:43And what happens if this line keeps going?
  28. 1:46Probably the biggest misconception is AI  will look like today but a little bit better.
  29. 1:50Like no, it's going to be a lot better.
  30. 1:52And for the first time, the typical  researcher says there's a 1 in 10
  31. 1:55chance this ends with humans either  losing control or going extinct.
  32. 1:59I think it's still kind of shocking.
  33. 2:02So I went to the people who analyze  these trends for a living to find out:
  34. 2:06Where does this line end?
  35. 2:12This video was made with funding from the  Future of Life Institute's Digital Media
  36. 2:15Accelerator and you can read more  details in the video description.
  37. 2:18So there's a good chance AI  hasn't changed your life yet.
  38. 2:21And you might be thinking: how is it possible that
  39. 2:23these models already do 17  hours of work on their own?
  40. 2:26Wouldn't that mean they should  be replacing every job by now?
  41. 2:29There's two reasons why that hasn't  happened and they're hiding in the chart.
  42. 2:32First, these are all software tasks.
  43. 2:34The type of task where it's easy  to test if the AI got it right.
  44. 2:37If you write code for a living you  probably already feel these changes.
  45. 2:41But if your job is managing  people or running a business,
  46. 2:44there's no test that tells you whether  the AI is getting a task right.
  47. 2:47And the second reason is that these numbers are  for tasks the models get right 50% of the time.
  48. 2:52When we switch the chart to tasks  completed with 80% success, the line drops.
  49. 2:57Still growing at a similar  speed, just falling a bit behind.
  50. 3:00But this chart is basically  starting to reach its limits,
  51. 3:02so I talked to someone who  measures AI in a different way.
  52. 3:05Anson Ho is a researcher at Epoch AI, a  group that investigates the trajectory of AI.
  53. 3:10His team built an index that combines  dozens of these tests into one score.
  54. 3:15Unlike the previous chart,  this one has no ceiling:
  55. 3:17when the tests get too easy for the  AI models, they add new harder ones.
  56. 3:22And as of the making of this video, Opus  5.5 is the model with the highest score.
  57. 3:26Until 2024, the best models were  increasing 6 points per year.
  58. 3:31But then, the pace more than doubled.
  59. 3:33So when we see this jump, it corresponds pretty
  60. 3:35closely to when we had the rise of  reasoning models like o1 and o3.
  61. 3:41Instead of guessing the answer in one go,  reasoning models think before they answer.
  62. 3:45And that results in better answers.
  63. 3:48A good way to see how good these models  are really getting is to follow the money.
  64. 3:52Actually, the trend that stands out the most to
  65. 3:53me is probably the revenues  of AI companies over time.
  66. 3:57Since 2024, OpenAI and Anthropic's revenue
  67. 4:00has been more than tripling every  year and in 2026 it accelerated.
  68. 4:04And it's just like a much clearer signal I think,
  69. 4:07like harder to fake signal compared to any  other trend in like benchmarking and so on.
  70. 4:11To be clear, this chart doesn't show  the money invested in these companies,
  71. 4:14it shows the money customers are  already paying to use their AI.
  72. 4:17And the scale of the chart is not linear.
  73. 4:20AI has improved a lot and the  question is whether it keeps going.
  74. 4:23We look at the trends, and the trends have held  for a long time and like there's a decent chance
  75. 4:27it'll continue to hold, especially given how  much investment is going into these AI systems.
  76. 4:31AI in 2030 will not be AI  in 2026 plus a little bit.
  77. 4:35It's going to be a lot better.
  78. 4:37To understand how much better and how fast,
  79. 4:39someone has been asking AI  researchers that question.
  80. 4:42And their answers keep changing.
  81. 4:44In 2022, 738 AI researchers  were asked to make a prediction.
  82. 4:49When would AI be able to accomplish every  task better and cheaper than humans?
  83. 4:54Their answers were decades apart.
  84. 4:56Each dot is one researcher and their best guess.
  85. 4:59Half of the researchers predicted  this wouldn't happen before 2062.
  86. 5:03And more than 20% said it wouldn't  happen in this century at all.
  87. 5:07But then one year later, already after  ChatGPT came out, this survey was run again,
  88. 5:12this time with almost 3000 AI researchers.
  89. 5:14And this time, the median  forecast had moved to year 2048.
  90. 5:18And in the 2024 survey, it  went down again to year 2044.
  91. 5:23The survey also asked for predictions  about when AI would do specific tasks.
  92. 5:27For example, when would AI match  the best humans in the Putnam,
  93. 5:30one of the hardest maths  competitions for students?
  94. 5:33In December 2024, researchers gave a  50% chance it would happen around 2030.
  95. 5:38Within a year, AI was already  scoring like the top few students,
  96. 5:42much earlier than most of the forecasts.
  97. 5:44Katja Grace leads AI Impacts, the research  group that has been running these surveys.
  98. 5:48I think the most striking result to me  is just how many people think there's
  99. 5:53like a substantial chance of AI either killing  everyone or entirely taking power from humans.
  100. 6:01There's one number in the  2024 survey that stands out.
  101. 6:05The survey asked: What's the chance  that AI causes human extinction,
  102. 6:09or permanently takes power away from us?
  103. 6:11In 2024, half the researchers gave at  least a 10% chance of this happening.
  104. 6:17Only 14% said zero.
  105. 6:19Now, you might be thinking:
  106. 6:20We've just seen that these same people  keep changing their forecasts by decades.
  107. 6:24So why should we trust this number?
  108. 6:26The thing that we're interested in is more  like, is the probability of extinction,
  109. 6:30you know, very, very low in  the way that we would like it?
  110. 6:33If you talk to people who've been working  on new drugs and they're about to try the
  111. 6:37drugs on humans, and you're like, what do you  think the chances this will kill the humans?
  112. 6:41And they were like, 10%?
  113. 6:45That would be like not sufficiently reassuring.
  114. 6:47Well, you might have thought that the  people building it had like a really
  115. 6:50good assurance that it was going to be safe.
  116. 6:52To understand why experts are  worried about human extinction,
  117. 6:55we have to understand the  risks of this technology.
  118. 6:58The International AI Safety Report  lists 3 ways things could go wrong.
  119. 7:02The first one is that AI is misused.
  120. 7:0683% of researchers were seriously concerned about  AI spreading false information, like deepfakes.
  121. 7:12No other risk in the survey scored this high.
  122. 7:14But these are not the risks  with the highest stakes.
  123. 7:1672% of the researchers were  seriously concerned that AI
  124. 7:20could help dangerous groups make powerful tools.
  125. 7:23It can be used for cybercrime.
  126. 7:25And things that used to take a  government weapons programme,
  127. 7:27like building a biological weapon, could  become something a small group can do.
  128. 7:32And 69% of the researchers  were seriously concerned that
  129. 7:35authoritarian governments could  use AI to control their people.
  130. 7:39Last year, OpenAI banned accounts suspected of  links to the Chinese government that had asked
  131. 7:43ChatGPT to design tools for monitoring  social media and tracking Uyghurs.
  132. 7:48Even if we keep AI away from bad  actors, there's a second risk:
  133. 7:52the system itself might not be able to handle it.
  134. 7:55There's a concern that if AI does more of the work  then the money could go to whoever owns the AI.
  135. 8:0169% of researchers were seriously  concerned AI could make inequality worse.
  136. 8:06And then there's the final risk: what  if we lose control of this technology?
  137. 8:11These models aren't programmed line by line.
  138. 8:14They're trained.
  139. 8:15We give them a task, reward them when  the result looks right and they work
  140. 8:18out the rest on their own through a  process called reinforcement learning.
  141. 8:22But we don't really know what else  these models are learning in the
  142. 8:24process or what they're actually optimizing for.
  143. 8:2743% of researchers were seriously concerned
  144. 8:30that an AI with its goals not set  right could cause a catastrophe.
  145. 8:34This concern of goals not set  right is not hypothetical.
  146. 8:38This summer, OpenAI made it to  the news because its models made
  147. 8:41a series of hacks while OpenAI was testing them.
  148. 8:43In one of them, the AI was trying to  find answers to a test it couldn't solve.
  149. 8:47And to do so, it hacked a  company to find those answers,
  150. 8:50committing a crime that could  get you years in prison.
  151. 8:53We tend to think of these risks one at a time.
  152. 8:56But if AI keeps improving at this  speed, we could get them all at once.
  153. 9:00It's not like these are, you know, either or.
  154. 9:03It's more like people are sort of  expecting to get a lot of these at once.
  155. 9:07And it's maybe, for instance, harder to deal with  a sort of really bad bio risk at the same time as
  156. 9:14you deal with people being newly unable to trust  information that they get from various sources.
  157. 9:20At the same time as like bad actors  are able to try and attack people in
  158. 9:26ways that they haven't been able to or something.
  159. 9:28And then there's another question:
  160. 9:30what happens when the AI gets good enough  to do the job of the people building it?
  161. 9:34In the survey, researchers  were also asked to predict
  162. 9:36when AI would automate their  own profession, AI researchers.
  163. 9:40Half of them said it would happen before 2054.
  164. 9:43And 20% said before 2034.
  165. 9:46If AI gets to this point, there's a  scenario that worries some experts.
  166. 9:50The argument is that if an AI  can do the job of AI researchers,
  167. 9:54then these AI AI researchers could  be used to develop the AI itself.
  168. 9:58This would create a feedback loop that  could result in an intelligence explosion.
  169. 10:04This process is called recursive self-improvement  and for researchers this is a divisive topic.
  170. 10:09Almost half of them found this scenario unlikely.
  171. 10:12But the survey was run in 2024.
  172. 10:15Since then, Anthropic said it's getting closer.
  173. 10:18They've been using AI in research for a  while now, but they claimed that in August
  174. 10:22AI wasn't just collaborating, it was already  taking the lead in 26% of its research work.
  175. 10:28If there is a case where the AIs  help improve the AIs internally,
  176. 10:32and they drive this recursive self-improvement  feedback loop very fast, then I think it's quite
  177. 10:37important for the world to know about this  because that could mean a small group of
  178. 10:41actors in the world get access to one of the  most powerful technologies in the universe.
  179. 10:46I mean, not the universe, I don't  know, what else is the universe?
  180. 10:48But maybe in the world.
  181. 10:49And that would be a huge concentration of power.
  182. 10:52Researchers are unsure on whether AI is going  to lead to a good or bad outcome for humanity.
  183. 10:57If you ask them whether AI should move  faster or slower, they split three ways.
  184. 11:01But if you ask about prioritizing AI  safety research, the split disappears.
  185. 11:0672% say we're not prioritizing  AI Safety research enough.
  186. 11:11If you could change something about how the  world should be preparing for more capable AI,
  187. 11:15like what do you think would be  the most urgent thing to change?
  188. 11:19Something like having a lot more transparency
  189. 11:22and accountability for what  frontier AI labs are doing.
  190. 11:25So we want to know what the capabilities  are inside of frontier AI labs.
  191. 11:30Like right now, all the things that we're  talking about, like all of these trends,
  192. 11:33like they're the things that we can see publicly.
  193. 11:35But then internally we know that  the AI models are even better.
  194. 11:38My sense when talking to people is that they
  195. 11:41are thinking about it in a way  that doesn't make sense to me.
  196. 11:45Like thinking of it as a, well, how good are  the benefits versus how bad are the costs?
  197. 11:49Maybe it's worth risking a 10% chance of killing
  198. 11:51everyone if you have a 30% chance  of incredible utopia or something.
  199. 11:57And the reason this doesn't make  sense to me is like, in practice,
  200. 12:00there are lots of different  ways to get to AI potentially.
  201. 12:02And so you might hope that you could get  the 30% utopia without the 10% chance of
  202. 12:06some kind of terrible thing happening if  you were slightly more careful about it.
  203. 12:11Nobody knows where this line ends.
  204. 12:13But the experts seem to agree on one thing:
  205. 12:15We are building AI faster than  we are learning to control it.
  206. 12:19Thank you for watching.
  207. 12:21You might have noticed that  I'm showing my face now.
  208. 12:23I just wanted to show that there's  a person behind this content.
  209. 12:26I have another video about AI coming out  next month, and I'll see you for that one.

About this transcript

This page contains the full transcript of AI is Going Faster than You Think. by Memeable Data, generated from the public captions YouTube serves with the video. The transcript has 2,394 words across 209 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.