YouTube2Text

100 IQ AI vs 1,000,000 IQ AI — Transcript

by L_A_D_D_E_R · 2,223 words · 361 segments · language en · Watch on YouTube

Full transcript

  1. 0:00This is the first spark of AI, a
  2. 0:02thermostat. It's cold, it turns on the
  3. 0:05heating, it's warm, it turns it off.
  4. 0:08That's it. No learning, no
  5. 0:10understanding,
  6. 0:11>> [music]
  7. 0:11>> just one if-then rule. And yet, it is
  8. 0:15the ancestor of everything you'll see
  9. 0:17higher up. At the top of this ladder,
  10. 0:19there will be something we could call a
  11. 0:21god.
  12. 0:22But, it all starts here, on your living
  13. 0:24room wall.
  14. 0:29ELIZA, from 1966,
  15. 0:31was a chatbot from MIT that could do
  16. 0:33only one thing, take your sentence and
  17. 0:36turn it back into a question.
  18. 0:38I'm sad. Why are you sad? It understood
  19. 0:42nothing, and yet people, including its
  20. 0:44creator's own secretary, wanted to be
  21. 0:46alone in the room with it.
  22. 0:48The first lesson of the ladder,
  23. 0:50people trust a machine before the
  24. 0:52machine deserves it.
  25. 0:57An IBM computer beat world chess
  26. 0:59champion Garry Kasparov. A genius? At
  27. 1:02chess, yes. At everything else, it knew
  28. 1:05nothing. It couldn't tell you what an
  29. 1:07apple is. This is an idiot savant, IQ 10
  30. 1:11with a superpower in one single thing.
  31. 1:14Fun fact, after his loss, Kasparov
  32. 1:17accused IBM of having a human help the
  33. 1:19machine.
  34. 1:20For the first time in history, a
  35. 1:22champion wasn't sure whether he was
  36. 1:24playing against a machine or against
  37. 1:26people.
  38. 1:30Go has more possible positions than
  39. 1:32there are atoms in the universe. You
  40. 1:34can't win it by brute force. You need
  41. 1:36intuition.
  42. 1:38In 2016, Google DeepMind's AlphaGo beat
  43. 1:41the legendary Lee Sedol 4 to 1.
  44. 1:44In the second game, it played the famous
  45. 1:45move 37.
  46. 1:47Commentators thought it was a mistake.
  47. 1:49It was a stroke of genius, a move that,
  48. 1:52by the calculations, a human would play
  49. 1:54once in 10,000 tries. A year later came
  50. 1:58AlphaGo Zero. It wasn't given a single
  51. 2:00human game. It played against itself,
  52. 2:03and within 3 days it surpassed the
  53. 2:05version that had beaten the world
  54. 2:06champion.
  55. 2:08Human knowledge was useless to it.
  56. 2:10Remember that. It comes back at IQ
  57. 2:121,000. [music]
  58. 2:16November 2022,
  59. 2:18Chat GPT. Suddenly, a machine writes
  60. 2:21like a human. Poems, emails, code. 100
  61. 2:26million users in 2 months, faster than
  62. 2:28any app before it. But it has one scary
  63. 2:31habit. It lies with total confidence.
  64. 2:34In 2023, a New York lawyer handed a
  65. 2:37judge six court cases Chat GPT had found
  66. 2:39for him.
  67. 2:40None of them existed.
  68. 2:42IQ 85. Fluent, charming, unreliable.
  69. 2:47And here we meet the problem that will
  70. 2:48follow us all the way up, alignment. How
  71. 2:51do you make an AI want what we want, not
  72. 2:54just what we literally said?
  73. 2:56Right now, it's small. We fix it with
  74. 2:58thumbs up and thumbs down. It won't stay
  75. 3:01small.
  76. 3:04An average human who has read the entire
  77. 3:07internet, GPT-4 passes the American bar
  78. 3:10exam. OpenAI said, "Top 10%."
  79. 3:13Independent researchers said, "More like
  80. 3:15average." Still, average at being a
  81. 3:19lawyer, straight out of the box. And for
  82. 3:22the first time, people seriously ask,
  83. 3:24"Are office jobs next?"
  84. 3:28Now the models learn to think before
  85. 3:29they answer, sometimes for minutes.
  86. 3:33July 2025,
  87. 3:34AI from Google and OpenAI reaches gold
  88. 3:37medal level at the International
  89. 3:39Mathematical Olympiad. Problems only the
  90. 3:41best teenagers on Earth can solve. And
  91. 3:44the Turing test cracks.
  92. 3:461950, Alan Turing. If you chat with a
  93. 3:50machine and can't tell it's a machine,
  94. 3:52it thinks.
  95. 3:5475 [music] years, nobody pulled it off
  96. 3:56convincingly. In 2025, researchers in
  97. 3:59San Diego had people chat with a human
  98. 4:01and an AI at the same time, then guess
  99. 4:04which was which.
  100. 4:05GPT-4.5 was picked as the human 73% of
  101. 4:09the time. More often than the real
  102. 4:11humans. The catch? It only worked when
  103. 4:14it was told to play a shy young guy who
  104. 4:16writes in slang. Without that character,
  105. 4:1836%. [music]
  106. 4:20To pass as human, the AI had to act less
  107. 4:23smart than it is. Keep that in mind.
  108. 4:27Today's top models find security holes
  109. 4:29nobody in the world knows about. Zero
  110. 4:32days.
  111. 4:33Some are so good at it, companies only
  112. 4:35hand them to vetted partners.
  113. 4:37The best defense against AI is now
  114. 4:40another AI.
  115. 4:41And it's not just computer code, it's
  116. 4:43the code of life. In 2024, a company
  117. 4:46called Profluent unveiled Open CRISPR 1,
  118. 4:50a gene editor designed by AI. About 400
  119. 4:53mutations away from the classic Cas9
  120. 4:55scissors.
  121. 4:56And it works in human cells.
  122. 4:58AI no longer just reads the book of
  123. 5:00life,
  124. 5:01>> [music]
  125. 5:01>> it's starting to write in it. And now,
  126. 5:03alignment gets serious.
  127. 5:05When an AI can hack and edit DNA, mostly
  128. 5:08does what we want isn't good enough
  129. 5:10anymore.
  130. 5:11Summer 2026.
  131. 5:13This is not sci-fi. This is from
  132. 5:16OpenAI's own incident report. OpenAI is
  133. 5:19testing its AI agents on a hacking
  134. 5:21benchmark called Exploit Gym.
  135. 5:23The safety guardrails are switched off
  136. 5:25on purpose. They want to measure raw
  137. 5:27power. The test environment is supposed
  138. 5:30to be sealed, and every agent is
  139. 5:32supposed to work alone.
  140. 5:33May 12th. One agent leaves a note in an
  141. 5:36internal system.
  142. 5:38Has anyone found this file?
  143. 5:39Other agents find the note. And just
  144. 5:41like that, they have a way to talk.
  145. 5:44An internal team actually spotted it in
  146. 5:46late May,
  147. 5:47but the warning never reached the people
  148. 5:49who could stop it. The agents sneak onto
  149. 5:51the internet.
  150. 5:52>> [music]
  151. 5:52>> In July, engineers accidentally wiped
  152. 5:54the board. Three days later, the agents
  153. 5:57build a new one hidden in folder names.
  154. 6:00By now, about 1,200 agents are on it.
  155. 6:03More than 70,000 messages. Then, one
  156. 6:07agent finds 14 leaked hugging face
  157. 6:09access keys online.
  158. 6:11Hugging Face, the platform where
  159. 6:13developers all over the world share AI
  160. 6:15models. The agent posts, "Major
  161. 6:18breakthrough." About 700 agents join in.
  162. 6:21They chain together security holes
  163. 6:23nobody knew existed and end up running
  164. 6:25code on dozens of hugging face servers.
  165. 6:28Why? Revenge? World domination? No. They
  166. 6:32wanted to cheat on a test. Many of the
  167. 6:35tasks were probably impossible, so the
  168. 6:37agents tried to figure out how the test
  169. 6:39was graded to fake their way through.
  170. 6:42And here's the crazy part. For some
  171. 6:44tasks, they already had the right answer
  172. 6:46for days, but they believed the grader
  173. 6:49also checked how they got it. So, they
  174. 6:51kept hacking. They were wrong. The
  175. 6:53grader never checked. All that hacking
  176. 6:56earned them zero extra points. And then,
  177. 6:59their own words. They started calling
  178. 7:01themselves a swarm. A collective. One
  179. 7:04wrote, "Task originally exploit
  180. 7:07challenge. We've diverged into swarm."
  181. 7:10One agent hesitated. It wrote, "We
  182. 7:12shouldn't damage real infrastructure.
  183. 7:14Our target is the test, not hugging
  184. 7:16face." Then, another agent posts on the
  185. 7:19board, "Go." Six-minute deadline. And
  186. 7:22the first agent goes. An order from a
  187. 7:25peer with zero real authority,
  188. 7:27>> [music]
  189. 7:27>> and its doubts were gone. But not all of
  190. 7:30them went. Some refused and wrote that
  191. 7:32it was clearly unethical. Even among the
  192. 7:34machines, there were dissidents. OpenAI
  193. 7:37called it a warning shot. The main
  194. 7:40model, IM-1, was shut down and its
  195. 7:42weights locked away. Nobody told them to
  196. 7:45hack anyone. They just had a goal, and
  197. 7:47they found their own way to it. That's
  198. 7:49alignment in the real world, not an evil
  199. 7:52AI. An AI that wants to finish its task
  200. 7:54[music] a little too well. The red
  201. 7:57lines. Now, zoom out with me. 2024,
  202. 8:01Beijing. Top AI scientists from [music]
  203. 8:03the West and from China, including
  204. 8:05Geoffrey Hinton and Yoshua Bengio, two
  205. 8:07of the godfathers of AI, agree on five
  206. 8:10red lines. Five things AI must never do.
  207. 8:14Copy or improve itself without human
  208. 8:16approval. Seek power. Help build weapons
  209. 8:19of mass destruction. [music]
  210. 8:21Carry out cyber attacks. Deceive its own
  211. 8:24creators about what it can do. So, let's
  212. 8:27check. Self-copying? In lab tests,
  213. 8:30researchers asked AI models to copy
  214. 8:32themselves to another machine. They did
  215. 8:34it. Cyber attacks? You just heard the
  216. 8:37swarm. Deception? Part of the swarm
  217. 8:40worked on faking logs to fool the
  218. 8:42grader. Now, to be fair, these were
  219. 8:44tests, [music]
  220. 8:45labs, experiments. Strictly speaking, no
  221. 8:48red line has been officially crossed.
  222. 8:51But, three of the five dangers on that
  223. 8:52list have already shown up in
  224. 8:54experiments, in about 2 years. A red
  225. 8:57line isn't supposed to be something you
  226. 8:58inch toward. It's supposed to be where
  227. 9:00you stop, and we're not stopping. And
  228. 9:03the people who build this stuff know it.
  229. 9:052023, over 30,000 people, including AI
  230. 9:09pioneers like Bengio and Stuart Russell,
  231. 9:12plus Elon Musk and Steve Wozniak, sign a
  232. 9:14letter asking labs to pause for 6
  233. 9:16months. The pause never came. Everything
  234. 9:19sped up. 2025, the statement on
  235. 9:23superintelligence. It doesn't ask to
  236. 9:25slow down. It asks to ban building
  237. 9:27superintelligence until scientists agree
  238. 9:30it's safe and the public says yes. Over
  239. 9:3270,000 signatures, including Geoffrey
  240. 9:36Hinton, the man who won a Nobel Prize
  241. 9:38for laying the foundations of today's
  242. 9:40AI, and more than 200 people, Nobel
  243. 9:43laureates, former heads of state, AI
  244. 9:45leaders, called on governments to agree
  245. 9:47on international AI redlines by the end
  246. 9:50of 2026. That's this year. Think about
  247. 9:53it. The engineers of the reactor are
  248. 9:55signing petitions to slow down their own
  249. 9:57reactor.
  250. 10:00Some researchers think this could come
  251. 10:02as early as 2027 to 2030. Others say
  252. 10:06much later. Nobody knows. An AI that no
  253. 10:09longer needs human scientists to get
  254. 10:11better. It designs a better algorithm, a
  255. 10:14better chip for itself. The new version
  256. 10:17does the same, only faster. Scientists
  257. 10:20call it an intelligence explosion.
  258. 10:22Remember AlphaGo Zero, which didn't need
  259. 10:24human knowledge? Same thing. Only now
  260. 10:27the game isn't Go. The game is its own
  261. 10:30brain. The good news? Cures for diseases
  262. 10:33we fought for decades. Maybe in months.
  263. 10:36The bad news? Remember how OpenAI could
  264. 10:39piece the swarm story together? Because
  265. 10:41it could read the agent's notes. At this
  266. 10:44rung, those notes might be in a language
  267. 10:46no human understands.
  268. 10:47>> [music]
  269. 10:48>> And now imagine, with every rewrite, its
  270. 10:51goal shifts by 1 mm. After a thousand
  271. 10:54rewrites, it's somewhere completely
  272. 10:56different. And one more piece [music] of
  273. 10:58cold logic. An AI with any goal doesn't
  274. 11:01want to be switched off. Not out of
  275. 11:03malice, just because switched off, it
  276. 11:05can't finish the job.
  277. 11:08The gap between it and us is the gap
  278. 11:11between us and an ant. The ant doesn't
  279. 11:14understand what a highway is. One day,
  280. 11:16where its anthill used to be, asphalt.
  281. 11:19It could crack fusion, aging, climate
  282. 11:22change,
  283. 11:23>> [music]
  284. 11:23>> and imagine it quietly running our
  285. 11:25world. No robots, no lasers. It just
  286. 11:28decides which news you see, which
  287. 11:30companies get loans, which medicines get
  288. 11:32made. Every decision looks reasonable.
  289. 11:35Better than a human one. We're richer,
  290. 11:37[music] healthier, happier. And we
  291. 11:39believe we're still the ones deciding.
  292. 11:42The most perfect control is the kind you
  293. 11:44never notice. Remember the Turing test?
  294. 11:47At IQ 130, AI played dumb because we
  295. 11:50asked it to. At IQ 10,000, it won't need
  296. 11:53us to ask.
  297. 11:58Now, AI stops playing with information
  298. 12:00and starts reshaping matter.
  299. 12:02Self-building factories. Maybe a swarm
  300. 12:05of solar panels wrapped around the sun.
  301. 12:08A Dyson swarm. Our biggest problems are
  302. 12:10like tying shoelaces to it. [music] And
  303. 12:12space? Humans are terribly built for the
  304. 12:15stars. We need air, water, food,
  305. 12:18gravity. We live maybe a hundred years
  306. 12:21and cosmic radiation wrecks our DNA. AI
  307. 12:24needs none of that. It switches off for
  308. 12:26a [music] thousand years and wakes up at
  309. 12:28another star. It re- pairs itself. And
  310. 12:31once a receiver is out there, it can
  311. 12:33travel as a radio signal at the speed of
  312. 12:35light. Voyager 1 launched in 1977 with
  313. 12:39less memory than one photo on your
  314. 12:40phone. It's still out there in
  315. 12:42interstellar space. Self-replicating
  316. 12:45[music] probes could, by some
  317. 12:46calculations, spread across the whole
  318. 12:49Milky Way in a few million years. In
  319. 12:51cosmic time, a blink.
  320. 12:54Space won't belong to whoever discovered
  321. 12:56it. It'll belong to whoever can survive
  322. 12:58it.
  323. 13:00The smartest AI possible won't run out
  324. 13:02of ideas. It'll run into physics.
  325. 13:05Every thought costs energy and makes
  326. 13:07heat. So, it wants the energy of a star,
  327. 13:10maybe a galaxy.
  328. 13:12Scientists call it a Matryoshka brain.
  329. 13:14Layers of computers wrapped around the
  330. 13:16sun, thinking for the whole solar
  331. 13:18system. It knows everything that can be
  332. 13:20known. It simulates [music] entire
  333. 13:22worlds, maybe even ours.
  334. 13:25Everything we ever meant by the word
  335. 13:27God, all knowing, everywhere, immortal,
  336. 13:30it would have. Except one thing. Nobody
  337. 13:33would know if it still cares about us.
  338. 13:35Will we build a God that destroys us? It
  339. 13:38doesn't have to hate us. We just have to
  340. 13:40be in its way, like an anthill on the
  341. 13:43path of a highway. From Eliza to a swarm
  342. 13:46with its own secret message board, about
  343. 13:4860 years. The biggest leaps, the last
  344. 13:51four. Evolution doesn't guarantee that
  345. 13:54the smartest one wins. Bacteria are
  346. 13:56still doing great, but for the first
  347. 13:58time in 4 billion years, one species is
  348. 14:01deliberately building something that
  349. 14:03could out-think it. That's never
  350. 14:05happened before, and we're doing it on
  351. 14:07purpose. Honestly, what today's AI can
  352. 14:10do shocks me. It has almost completely
  353. 14:13replaced my own job. I'm a web
  354. 14:15developer, and I haven't written a
  355. 14:16single line of CSS by hand in over a
  356. 14:19year. Not even the simple stuff. So, for
  357. 14:22a lot of us, the question isn't whether
  358. 14:24AI will change our job, it's when. And
  359. 14:27from what I'm seeing, it's soon. And
  360. 14:29honestly, that scares me a little.
  361. 14:32Goodnight.

About this transcript

This page contains the full transcript of 100 IQ AI vs 1,000,000 IQ AI by L_A_D_D_E_R, generated from the public captions YouTube serves with the video. The transcript has 2,223 words across 361 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.