YouTube2Text

not another AI video — Transcript

by Anna Bocca · 8,922 words · 1,336 segments · language en · Watch on YouTube

Full transcript

  1. 0:00December 2012,
  2. 0:03a casino hotel at Lake Tahoe, Nevada. In
  3. 0:07one of the rooms, a computer science
  4. 0:09professor called Joffrey Hinton is
  5. 0:11running a secret auction by email. What
  6. 0:14he's selling is a company he's just
  7. 0:16founded with his two students. It has
  8. 0:18three employees, no product, no revenue,
  9. 0:22and no business plan. Just one research
  10. 0:26paper. The biders, Google, Microsoft,
  11. 0:30the Chinese tech giant BYU, and a small
  12. 0:32London startup called Deep Mind. The
  13. 0:36bits climb by millions at a time. And in
  14. 0:39the end, Google wins for $44 million.
  15. 0:44This was the big bang of modern AI, the
  16. 0:48moment decades of university research
  17. 0:50started moving into the hands of big
  18. 0:52tech.
  19. 0:54Everyone talks about what AI will do to
  20. 0:56us. Will it kill us all? Take all our
  21. 0:59jobs or cure cancer, do all the work,
  22. 1:02and let us just have fun for the rest of
  23. 1:04our lives? The truth is, we can't see
  24. 1:07the future. But what we can do is look
  25. 1:09at where AI came from because that might
  26. 1:13help us understand where it's going. So,
  27. 1:15this video is the history of AI. And
  28. 1:18this is also the third video in my big
  29. 1:21tech series. Each video of my series
  30. 1:24works on its own, so you don't need to
  31. 1:26have seen the others, but if you want
  32. 1:28the full picture, I'd really recommend
  33. 1:31watching the other two as well after
  34. 1:33this video. Okay, but without further
  35. 1:36ado, let's get into the history of AI.
  36. 1:41First, we should probably define what AI
  37. 1:44actually is. Everyone is talking about
  38. 1:46AI, but only few people understand the
  39. 1:50concept of machine learning behind it.
  40. 1:52Artificial intelligence isn't really a
  41. 1:54scientific term. It's closer to a
  42. 1:57marketing label. The problem with the
  43. 1:58word AI is that it's like using the word
  44. 2:01vehicle for a bicycle, a truck, and a
  45. 2:03moon rocket at the same time. At least
  46. 2:06that's how two Princeton computer
  47. 2:08scientists, Arvin Narayan and Sash
  48. 2:12Kapoor, put it. Also, as always, I've
  49. 2:15made a document with all the sources I
  50. 2:17used in the video in the description.
  51. 2:19But back to the term AI. Everyone says
  52. 2:22AI, but hardly anyone knows what they
  53. 2:25actually mean by it. Because when most
  54. 2:28people say AI today, they often talk
  55. 2:31about a very specific form of AI,
  56. 2:34chatbots, like chatbt.
  57. 2:37So, let's talk about those first because
  58. 2:40I assume most of you are familiar with
  59. 2:42them and it will make the other ones
  60. 2:44easier to understand too. Under the hood
  61. 2:47of those chat bots is something called a
  62. 2:49large language model or LLM. And an LLM
  63. 2:53is basically autocomplete on an
  64. 2:55astronomical scale. But to really
  65. 2:58understand what that means, I think it's
  66. 3:00easiest to contrast it to classical
  67. 3:03programming. Writing software is a
  68. 3:05rule-based endeavor. A human programmer
  69. 3:08explicitly writes instructions.
  70. 3:11Like if user clicks button then save
  71. 3:14file. If you wanted to write a computer
  72. 3:18program that writes English, you'd have
  73. 3:21to write down millions of rigid
  74. 3:23grammatical rules, dictionary
  75. 3:24definitions, and exception trees.
  76. 3:27Historically, this never really worked
  77. 3:30because human language is too messy and
  78. 3:32nuanced to capture in static rules. The
  79. 3:36alternative is using something called
  80. 3:38machine learning. Instead of writing the
  81. 3:40rules yourself, you give the computer a
  82. 3:43blank mathematical model and a mountain
  83. 3:46of examples and tell it figure out the
  84. 3:49pattern yourself. And that's machine
  85. 3:51learning. And machine learning is behind
  86. 3:54almost everything we call AI. your
  87. 3:56Netflix recommendations, Face ID,
  88. 3:59self-driving cars. A large language
  89. 4:02model is just the version of it that
  90. 4:04learned language. And to explain how
  91. 4:06machine learning works, we are going to
  92. 4:08look at language. To visualize it, I'm
  93. 4:11going to build my very own language
  94. 4:13model. But before we continue, I wanted
  95. 4:16to tell you how these videos get made
  96. 4:19because loads of you have been curious
  97. 4:21how I go about my research. So, here's
  98. 4:24how I usually work. Making these kinds
  99. 4:26of videos takes a lot of research. I
  100. 4:28read a lot of books, articles,
  101. 4:30scientific papers, and what I do most of
  102. 4:33my reading in is UPDF, the sponsor of
  103. 4:37today's video, because this is where I
  104. 4:40can work best. I have tried other PDF
  105. 4:42readers before, like Adobe Acrobat,
  106. 4:45which has an insane subscription fee, or
  107. 4:48my laptop's preview, but none of them
  108. 4:50let me work as well as UPDF. This is
  109. 4:53their 1955 Dartmouth proposal, the
  110. 4:56document that coined the term artificial
  111. 4:58intelligence and this is attention is
  112. 5:01all you need from 2017, the paper behind
  113. 5:04chat gvt and every chatbot since. These
  114. 5:07were two very important sources that
  115. 5:09I've based this video on. And the cool
  116. 5:13thing about UPDF is that I can highlight
  117. 5:16and annotate right on the page while
  118. 5:18reading it. And with the AI sidebar
  119. 5:21open, I can ask it to explain a passage
  120. 5:23if I have difficulty understanding
  121. 5:24something.
  122. 5:26Also, a lot of early research papers
  123. 5:29like this one are scans of typewritten
  124. 5:31pages. So, normally you can't search
  125. 5:34them or copy anything out of them. With
  126. 5:36UPDF's OCR, one click turns it into real
  127. 5:40text. Now, I can search it, highlight,
  128. 5:43and annotate right on the page. Also, I
  129. 5:46use UPDF's e-sign function to sign
  130. 5:50contracts or important documents. So,
  131. 5:53that's been really useful for me, too.
  132. 5:56UPDF works on Mac, Windows, iPhone, and
  133. 5:59Android. And one license covers all
  134. 6:02four. You can download it for free. And
  135. 6:05if you're a heavy user like me, the pro
  136. 6:07version costs about a sixth of Adobe
  137. 6:09Acrobat and it has a lifetime option, so
  138. 6:13you don't have to keep paying for a
  139. 6:14subscription. There's a limited time
  140. 6:17launch offer for UPDF 3.0 right now. The
  141. 6:21link is at the top of the description.
  142. 6:23And now, let's get back to the video.
  143. 6:26I'm going to build a language model out
  144. 6:28of paper cups to visualize how large
  145. 6:30language models work. I'm feeding our
  146. 6:33DIY language model a bit of text from my
  147. 6:36last video where I explained how chips
  148. 6:38are made by making a photo print of my
  149. 6:41cat blobber, which also is very useful
  150. 6:45to understand the foundation AI is built
  151. 6:47on. So, I highly suggest you go watch
  152. 6:49the video on chips after this video if
  153. 6:52you want a deeper understanding. But
  154. 6:54anyway, here's how we're setting up our
  155. 6:56very own large language model. I mean,
  156. 6:59it's not going to be large, it's going
  157. 7:00to be a language model. I go through the
  158. 7:03text word by word and for every word I
  159. 7:05write down the word that comes right
  160. 7:07after it. Every time the word the
  161. 7:10appears, I look at the next word, write
  162. 7:13it on a slip and drop it in the the cup.
  163. 7:16Okay, this is the the cup. It has 10
  164. 7:19slips in it. Three of them say pattern
  165. 7:22and the others say things like chip,
  166. 7:25light, mask. So if I draw from this cup,
  167. 7:29pattern is three times more likely to
  168. 7:32come out than anything else. So there's
  169. 7:34a higher probability I will draw
  170. 7:36pattern. And here's something
  171. 7:38interesting. Most words don't even need
  172. 7:41a cup. Nothing, for example, is in my
  173. 7:43text only ever followed by happens. So
  174. 7:46whenever I land on nothing, there's no
  175. 7:49choice to make. The next word is always
  176. 7:52happens. Only 25 words in this text have
  177. 7:56more than one option. So only they
  178. 7:58really need a cup. Now let's make it
  179. 8:00talk. I start with the
  180. 8:08chip.
  181. 8:09Chip has a cup.
  182. 8:14people
  183. 8:23use
  184. 8:30blobber.
  185. 8:32So yeah, we just built a language model
  186. 8:34out of paper cups and it just told us
  187. 8:37that chip factories use my cow. It
  188. 8:40sounds like a real sentence. It's
  189. 8:42grammatically fine and it's obviously
  190. 8:44complete nonsense because the model
  191. 8:47doesn't know what a chip is. It doesn't
  192. 8:50know that blobber is a cat. All it knows
  193. 8:52is which words followed which in my
  194. 8:54script and how often. Computer
  195. 8:57scientists call what we just did an
  196. 8:59engram model. The n is how many words in
  197. 9:02a row the model looks at. Ours always
  198. 9:05looks at two words in a row. So it's
  199. 9:08called a bagram model. Now, as we've
  200. 9:11seen, my language model is very limited.
  201. 9:14It only knows the words in this passage.
  202. 9:17So, it can never say anything else. It
  203. 9:20can't say dog. It can't say hello. And
  204. 9:23because most words only have one slip,
  205. 9:25most of the time it just repeats my
  206. 9:27script back to me. Generally, if you
  207. 9:29only look at two words at a time, the
  208. 9:32sentences quickly degenerate into
  209. 9:34repetitive or disjointed gibberish. So,
  210. 9:36the obvious fix would be let it look at
  211. 9:39more words at once. Not just the last
  212. 9:41word, as in the two-word combinations we
  213. 9:44used here, but five word combinations.
  214. 9:47But that's where it breaks. The more
  215. 9:49words in a row you look at, the rarer
  216. 9:52each combination gets. Most sentences
  217. 9:54you write have probably never been
  218. 9:57written exactly that way before. So, for
  219. 9:59almost any five words, there's simply no
  220. 10:03cop. the model has never seen them and
  221. 10:06has no idea what comes next. This is
  222. 10:09called the sparsity problem and it's a
  223. 10:11big part of why language models like
  224. 10:13mine were pretty useless for decades. In
  225. 10:17other words, the sparsity problem is
  226. 10:19essentially the issue that the more
  227. 10:22words in a row a model looks at, the
  228. 10:24more likely that exact combination has
  229. 10:27never been written before. So, the model
  230. 10:30has nothing to learn from and no idea
  231. 10:32what comes next. So, if the sparsity
  232. 10:34problem kept language models useless for
  233. 10:36decades, how did tragic get past it?
  234. 10:40Part of the answer is the obvious one.
  235. 10:43It got way more text, a lot more. My cup
  236. 10:46model read 130 words about visualizing
  237. 10:49chipm
  238. 10:52read a huge chunk of the internet,
  239. 10:54websites, Wikipedia, Reddit, millions of
  240. 10:57books, code. And quick side note, these
  241. 11:00models don't actually read words. They
  242. 11:03read tokens, chunks of words. A long
  243. 11:07word gets chopped into pieces. On
  244. 11:09average, a token is about 3/4 of a word.
  245. 11:12Chachi has never heard of my cat, so it
  246. 11:15doesn't see blobber. It sees blob and
  247. 11:19burr. Or economics is chopped up into
  248. 11:22econom x.
  249. 11:24And you can see how words get tokenized
  250. 11:27yourself. Open AAI has a free tool where
  251. 11:29you type in any text and it shows you
  252. 11:31exactly how it gets chopped into tokens.
  253. 11:34And tokens matter for a lot of things.
  254. 11:36First, they explain why Chapd types its
  255. 11:39answer chunk by chunk. Here it is in
  256. 11:41slow-mo. It's because it generates one
  257. 11:44token at a time. Secondly, chat's memory
  258. 11:48is measured in tokens. Every chatbot can
  259. 11:51only keep a certain number of tokens in
  260. 11:53mind at once. That's called the context
  261. 11:56window. Today that can be hundreds of
  262. 11:59thousands of tokens, but it's still a
  263. 12:01limit. And once a conversation gets too
  264. 12:03long, the oldest parts start to fall
  265. 12:06out. That's why in really long chats,
  266. 12:08Chhatributt sometimes forgets what you
  267. 12:11told it at the start. And third, tokens
  268. 12:14are money. Companies that build products
  269. 12:16on top of Jatribute pay Open AI per
  270. 12:19token for every token that goes in and
  271. 12:22every token that comes out. So every
  272. 12:24question you ask and every answer you
  273. 12:26get has a price tag that is measured in
  274. 12:28tokens. And tokens are also how AI
  275. 12:31companies measure how much a model has
  276. 12:34read in total. So chatbot's entire
  277. 12:36knowledge base is measured in tokens. So
  278. 12:39how much did chat read? Open AAI doesn't
  279. 12:42say exactly how much anymore, but Meta
  280. 12:44does say it for its own model. Llama 3
  281. 12:48around 15 trillion tokens, which is
  282. 12:50roughly 11 trillion words. If you read 8
  283. 12:53hours a day every day, that would take
  284. 12:55you about 250,000 years. Okay, but back
  285. 12:59to my own language model and its
  286. 13:00limitations. Remember the sparsity
  287. 13:03problem? The problem that the more words
  288. 13:05or tokens in a row a model looks at, the
  289. 13:08more likely that exact combination has
  290. 13:11never been written before. So, the model
  291. 13:14has nothing to learn from and no idea
  292. 13:16what comes next.
  293. 13:18Well, even after 250,000 years worth
  294. 13:21reading, most fiveword combinations
  295. 13:24still never show up. So, more text alone
  296. 13:27doesn't solve it. So, how did Chhatri
  297. 13:30get around it? Two ways. First, Chhatri
  298. 13:34doesn't use cops. Instead, it has
  299. 13:36hundreds of billions of so-called
  300. 13:38weights. GPD3, Chhat GPT's predecessor,
  301. 13:43had 175 billion of them. Think of them
  302. 13:46as tiny dials. At the start, they're all
  303. 13:50set randomly. During training, it
  304. 13:52guesses the next token, and when it's
  305. 13:54wrong, a bit of maths called gradient
  306. 13:57descent nudges all the dials a tiny bit,
  307. 14:00so it guesses slightly better next time.
  308. 14:03Then you repeat that across trillions of
  309. 14:05tokens. The result is that it doesn't
  310. 14:08memorize word pairs. It learns patterns.
  311. 14:12It learns, for example, that cat and
  312. 14:14kitten get used in similar ways. So, it
  313. 14:17can deal with sentences it has never
  314. 14:19seen before. Second, it doesn't just
  315. 14:22look at the last word. It looks at
  316. 14:24everything, your whole question, the
  317. 14:26whole conversation, hundreds of
  318. 14:28thousands of tokens back. So, it's the
  319. 14:31same game as my paper cups. Predict the
  320. 14:33next word, just with billions of dials
  321. 14:35instead of slips of paper and a memory
  322. 14:37that reaches back across the entire
  323. 14:39conversation. Now, remember the other
  324. 14:41vehicles under the AI umbrella I
  325. 14:43mentioned at the start? Netflix
  326. 14:45recommendations, Face ID, self-driving
  327. 14:47cars. Here's the thing. They basically
  328. 14:50all went the same. Same dials, same
  329. 14:52nudging, just different examples. Face
  330. 14:55ID learned from millions of photos of
  331. 14:57faces. Netflix and Tik Tok learned from
  332. 14:59what billions of people watched.
  333. 15:01Self-driving cars learned from millions
  334. 15:03of hours of driving footage. And image
  335. 15:05generators like midjourney or deli learn
  336. 15:09from billions of images with captions.
  337. 15:11So different vehicles, same engine. Show
  338. 15:14a machine enough examples and let it
  339. 15:16notch its dials until it finds the
  340. 15:18pattern. That's machine learning. But it
  341. 15:21wasn't always like this. For most of
  342. 15:23AI's history, the people building it on
  343. 15:25the exact opposite approach, writing the
  344. 15:27rules by hand, the if then way. And that
  345. 15:30story is the one we are going to look
  346. 15:32at. It starts in 1950 with one question.
  347. 15:36Can machines think? The man who asked it
  348. 15:39was Alan Turing. And if that name sounds
  349. 15:42familiar, it's because he's the guy who
  350. 15:44helped break the Nazis Enigma code in
  351. 15:47World War II. Turing thought the
  352. 15:49question, "Can machines think?" was too
  353. 15:52vague to answer because what does
  354. 15:55thinking even mean? So, he swapped it
  355. 15:57for a different question, one you could
  356. 15:59actually test. He called it the
  357. 16:02imitation game.
  358. 16:04Imagine you're texting with two people
  359. 16:06you can't see. One is a human, one is a
  360. 16:09machine. You can ask them anything you
  361. 16:12want. And if you can't reliably tell
  362. 16:15which one is the machine, then Turing
  363. 16:18said, "We might as well say it can
  364. 16:20think."
  365. 16:22This became known as the Turing test.
  366. 16:25And you could argue it was the birth of
  367. 16:27AI's original sin. Defining intelligence
  368. 16:30not by whether a machine actually
  369. 16:32understands anything, but by whether it
  370. 16:35can convince a human that it does.
  371. 16:38Turing died in 1954, so he never got to
  372. 16:40see what came next. But just one year
  373. 16:43later in 1955, a young maths professor
  374. 16:46called John McCarthy wrote a funding
  375. 16:49proposal for a summer workshop at
  376. 16:51Dartmouth College in the US. together
  377. 16:54with Marvin Minsky, Claude Shannon, and
  378. 16:56Nathaniel Rochester from IBM. And it's
  379. 16:59in this proposal that the term
  380. 17:01artificial intelligence was used for the
  381. 17:04very first time. This is where the name
  382. 17:06comes from. And remember how I said AI
  383. 17:09is basically a marketing label. Well, it
  384. 17:12was from the very start. McCarthy later
  385. 17:15said he picked the name partly to set
  386. 17:17his field apart from a rival one called
  387. 17:19cybernetics. Cybernetics is the study of
  388. 17:22how machines and living things control
  389. 17:23themselves through feedback like a
  390. 17:26thermostat that keeps adjusting to hold
  391. 17:28a room at the right temperature. But
  392. 17:31back to the proposal. The proposal said
  393. 17:33that a significant advance could be made
  394. 17:35if a carefully selected group of
  395. 17:37scientists work on it together for a
  396. 17:40summer. And what do they want to solve
  397. 17:43in that one summer? They said they
  398. 17:45wanted to get machines to use language,
  399. 17:48form concepts, solve problems that only
  400. 17:51humans could solve and improve
  401. 17:53themselves in one summer. That was 70
  402. 17:56years ago. And there was a lot of
  403. 17:58optimism at the start. These early AI
  404. 18:00pioneers genuinely believed thinking
  405. 18:03machines were just around the corner.
  406. 18:05Two of the people who came to the
  407. 18:07Dartmouth summer research project on
  408. 18:09artificial intelligence were Herbert
  409. 18:13Simon and Alan Newwell. And they even
  410. 18:16brought a working program. It could
  411. 18:18prove mast theorems by itself. And a
  412. 18:23year later, they made some bold
  413. 18:24predictions. Within 10 years, they said
  414. 18:27a computer would be the world chess
  415. 18:29champion. Well, it actually ended up
  416. 18:31taking 40 years until IBM's Deep Blue
  417. 18:35beat world champion Gary Kasparov in
  418. 18:381997. But still, in 1967, Mavin Minsky,
  419. 18:43one of the Dartmouth organizers, wrote
  420. 18:45that within a generation, the problem of
  421. 18:48creating artificial intelligence will
  422. 18:51substantially be solved. But yeah, my
  423. 18:53point here is that these early AI people
  424. 18:56were very, very optimistic. But while
  425. 18:59they all agreed thinking machines were
  426. 19:01coming, they completely disagreed on how
  427. 19:04to build them. And the field split into
  428. 19:07two camps with totally different ideas
  429. 19:10of what intelligence even is. The first
  430. 19:12camp said, "Intelligence is rules. If we
  431. 19:16can write down all the rules of logic
  432. 19:18and knowledge, a computer can follow
  433. 19:21them and think." This is called symbolic
  434. 19:24AI. It's basically the if then approach
  435. 19:27of classical programming and most of the
  436. 19:29Dartmouth crowd were in this camp.
  437. 19:32McCarthy, Minsky, Simon, Newell. But the
  438. 19:36second camp said, "Intelligence is
  439. 19:38learning. Our brains aren't rule books.
  440. 19:42They are networks of neurons that learn
  441. 19:44from experience.
  442. 19:46So let's build machines that work like
  443. 19:48that." This is called connectionism or
  444. 19:51neural networks. And its biggest
  445. 19:53champion was a guy called Frank
  446. 19:55Rosenblat. Rosenblat wasn't a computer
  447. 19:57scientist. He was a psychologist at
  448. 19:59Cornell. And what he really wanted to
  449. 20:02understand was the human brain. How a
  450. 20:05lump of neurons can learn to recognize
  451. 20:08things it has never been taught. So he
  452. 20:11decided the best way to understand the
  453. 20:13brain was to build one. In 1958,
  454. 20:17Rosenlat and the US Navy, who funded his
  455. 20:20research, showed off the Perceptron for
  456. 20:23the first time. It was a program running
  457. 20:25on a giant computer. And a couple of
  458. 20:27years later, Rosenblad built it as its
  459. 20:30own machine, the Mark1 Perceptron, a
  460. 20:32fridge-sized computer that was hooked up
  461. 20:35to a camera with 400 pixels, which
  462. 20:37translates to a 20x 20 grid of light
  463. 20:40sensors, which was basically the
  464. 20:41perceptron's eye. The light from an
  465. 20:44image was passed onto 512 internal units
  466. 20:47connected to adjustable dials. Those
  467. 20:49dials were its weights and then it
  468. 20:52learned. You'd show it cards with a mark
  469. 20:55either on the left or on the right.
  470. 20:57Every time it got one wrong, it adjusted
  471. 20:59its weights a little. Until it could
  472. 21:01reliably tell the patterns apart. And
  473. 21:03here's why this was such a big deal. The
  474. 21:05perceptron was one of the first systems
  475. 21:07to turn an image into numbers. basically
  476. 21:10coordinates on a map and then learn
  477. 21:13where to draw the line between left and
  478. 21:15right. No one had programmed any rules
  479. 21:17into it. It had learned the difference
  480. 21:19all by itself. And that's the exact same
  481. 21:22principle behind image recognition today
  482. 21:24and behind the embeddings we talked
  483. 21:26about with Chach. But what really made
  484. 21:29the perceptron famous was the hype. In
  485. 21:32July 1958, when it was first shown to
  486. 21:35the world, the US Navy, who were funding
  487. 21:37Rosenlat's research, held a press
  488. 21:40conference in Washington, and the press
  489. 21:42went wild. The New York Times reported
  490. 21:44that the Navy had revealed the embryo of
  491. 21:48an electronic computer that it expected
  492. 21:51will be able to walk, talk, see, write,
  493. 21:54reproduce itself, and be conscious of
  494. 21:56its existence. I mean, this was a
  495. 21:59machine that could tell left and right,
  496. 22:01but the headlines at the time made it
  497. 22:03sound like it was about to become
  498. 22:05conscious. Rosenlat's machine was the
  499. 22:08learning cam's big moment, the
  500. 22:10perceptron, the very first neural
  501. 22:12network to make headlines, a machine
  502. 22:14that learned from examples, just like
  503. 22:16the AI we use today. So, yeah, it seemed
  504. 22:19like the learning camp had figured it
  505. 22:21out. But the rules camp had its own
  506. 22:23showpiece and it might be the most
  507. 22:25famous early AI program of all. In 1966,
  508. 22:29a computer scientist at MIT called Yseph
  509. 22:33Bitesenbound built Eliza, one of the
  510. 22:36very first chatbots. Eliza pretended to
  511. 22:39be a therapist. So yes, people have been
  512. 22:42oversharing with computers since 1966.
  513. 22:45Uh 6 years before people started trauma
  514. 22:48dumping on chatbt, they did it with
  515. 22:50Eliza. And Eliza worked with a few
  516. 22:52simple rules. It spotted a keyword in
  517. 22:55what you typed, turned your own words
  518. 22:57back into a question. And of course,
  519. 22:59people loved it. They opened up to it.
  520. 23:01They told it about their problems, their
  521. 23:04relationships, their families, and even
  522. 23:06formed deep emotional bonds with Eliza.
  523. 23:08Right. Bam's own secretary, who had
  524. 23:11watched him build it for months and knew
  525. 23:13it was just a program, reportedly asked
  526. 23:15him to leave the room so she could talk
  527. 23:18to Eliza in private. And Whitenbalam was
  528. 23:20horrified. He'd built Eliza partly to
  529. 23:23show how shallow these conversations
  530. 23:25with a computer really were. Instead,
  531. 23:28he'd shown how easily humans believe a
  532. 23:30machine understands them, even when it's
  533. 23:33just following a handful of rules. This
  534. 23:36became known as the Eliza effect. our
  535. 23:38tendency to project empathy,
  536. 23:40understanding, and intention onto
  537. 23:42software that has none. And it got
  538. 23:45worse. Some psychiatrists actually
  539. 23:47suggested using programs like Eliza to
  540. 23:50treat real patients. And we're seeing
  541. 23:53that today as well, where people use
  542. 23:56chatbt and other chat bots as a
  543. 23:59replacement for actual therapy. But
  544. 24:01yeah, Whitesenbound was so disturbed
  545. 24:03that he spent the rest of his career
  546. 24:05warning people about trusting computers
  547. 24:07too much. But there was another thing.
  548. 24:10Programs like Eliza made AI look much
  549. 24:13further along than it really was. From
  550. 24:15the outside, it seemed like all those
  551. 24:17big promises from the Dartmouth summer
  552. 24:19camp were finally coming true. But by
  553. 24:23the late 1960s, all these big promises
  554. 24:26started running into reality. And that
  555. 24:28reality is best explained by looking at
  556. 24:30an example like machine translation.
  557. 24:33During the Cold War, the US government
  558. 24:35spent millions trying to build computers
  559. 24:38that could automatically translate
  560. 24:40Russian into English. Very handy if you
  561. 24:43want to read Soviet documents. But in
  562. 24:461966, a government committee published
  563. 24:49what's called the ALPAC report. The
  564. 24:51verdict was machine translation was
  565. 24:54slower, less accurate, and twice as
  566. 24:56expensive as just paying a human
  567. 24:58translator. And the funding dried up.
  568. 25:01Remember the two camps I talked about
  569. 25:03earlier? The rules camp who wanted to
  570. 25:05write down all the rules of intelligence
  571. 25:08by hand and the learning camp who wanted
  572. 25:11machines to learn from examples like
  573. 25:14Rosenlat's perceptron. Well, machine
  574. 25:16translation was a blow for the rules
  575. 25:18camp. But in 1969, the learning camp got
  576. 25:21hit too. And the attack came from its
  577. 25:24biggest rival. Marvin Minsky, one of the
  578. 25:26leaders of the rules camp, and his
  579. 25:28colleague Samur published a book called
  580. 25:31Perceptrons.
  581. 25:32In it, they proved mathematically that a
  582. 25:35perceptron like Rosenlats couldn't solve
  583. 25:37some surprisingly simple problems. For
  584. 25:40example, it couldn't learn to tell
  585. 25:42whether two dots were the same color or
  586. 25:45different colors. and they suggested
  587. 25:48that adding more layers probably
  588. 25:50wouldn't fix it. Many people read the
  589. 25:52book as proof that neural networks were
  590. 25:54a dead end and the money for neural
  591. 25:56network research largely dried up too.
  592. 25:58And then in 1973, the British government
  593. 26:01asked the mathematician Sir James
  594. 26:04Lighill to review the whole field of
  595. 26:06artificial intelligence. His conclusion
  596. 26:08was brutal. In no part of the field have
  597. 26:11the discoveries made so far produced the
  598. 26:14major impact that was then promised. One
  599. 26:17big reason why there was no advancement
  600. 26:19was something called the combinatorial
  601. 26:22explosion. The rules approach worked on
  602. 26:24small toy problems. But as soon as
  603. 26:26problems got even a little bigger, the
  604. 26:29number of possibilities exploded beyond
  605. 26:32what any computer could handle. But
  606. 26:35yeah, this ended up with the UK cutting
  607. 26:38most of its AI funding. And in the US,
  608. 26:41DARPA, the military research agency that
  609. 26:44had been AI's biggest sponsor, started
  610. 26:47only funding research with direct
  611. 26:49military uses. And just like that, the
  612. 26:52money for AI research was gone. This is
  613. 26:55what's now called the first AI winter.
  614. 26:58But AI didn't stay dead for long. In the
  615. 27:011980s, the pendulum swung back. And this
  616. 27:04time the rules camp returned with a
  617. 27:06rebrand expert systems. Essentially they
  618. 27:11stopped trying to build a general
  619. 27:12thinking machine but instead build
  620. 27:14machines that could do one narrow job
  621. 27:18like diagnosing infections or
  622. 27:20configuring computers. Their approach to
  623. 27:22building that was to sit down with a
  624. 27:24human expert and write down everything
  625. 27:26they know as hundreds of thousands of if
  626. 27:29then rules. And for a while this
  627. 27:32actually worked really well. The
  628. 27:33computer company DEEC used an expert
  629. 27:36system called XCON to put together
  630. 27:38customer orders and it reportedly saved
  631. 27:41them tens of millions of dollars a year.
  632. 27:43Suddenly AI wasn't just a lab experiment
  633. 27:46anymore. It was a business. By the late
  634. 27:481980s, AI had become a billion-doll
  635. 27:50industry. Another famous example was
  636. 27:53mycin, a system built at Stanford in the
  637. 27:561970s that diagnosed blood infections
  638. 27:59and recommended which antibiotics to
  639. 28:01give. In tests, it performed about as
  640. 28:04well as Stanford's own infection
  641. 28:06specialists, but it was never actually
  642. 28:08used on real patients, partly because
  643. 28:11nobody could agree who'd be responsible
  644. 28:13if it got something wrong. This new AI
  645. 28:15boom was so big that a whole industry
  646. 28:17grew up around building special
  647. 28:19computers just for AI called lisp
  648. 28:22machines, named after lisp, the
  649. 28:25programming language most AI was written
  650. 28:27in. Companies poured huge sums into
  651. 28:30them. But yeah, the fact is that there
  652. 28:32were more and more real world uses for
  653. 28:34AI. And amidst this AI upswing,
  654. 28:38something interesting happened in 1982.
  655. 28:41Far away from the US and the UK,
  656. 28:44Japan announced the fifth generation
  657. 28:46computer project. A 10-year government
  658. 28:50funded plan to build intelligent
  659. 28:51computers and leapfrog the West. And the
  660. 28:54West obviously panicked. DARPA, the same
  661. 28:58agency that had cut the money a decade
  662. 29:00earlier, launched its strategic
  663. 29:02computing initiative. Also, I talk a lot
  664. 29:04more about DARPA in this video. It's
  665. 29:07actually the agency that basically built
  666. 29:09big tech. But yeah, DARPA launched the
  667. 29:12strategic computing initiative and the
  668. 29:15UK launched its own program and hundreds
  669. 29:18of millions of dollars were poured back
  670. 29:20into AI. But just like the first time,
  671. 29:22the hype didn't last. These expert
  672. 29:25systems turned out to be incredibly
  673. 29:27brittle. They weren't great as long as
  674. 29:29the situation was exactly what the rules
  675. 29:31were written for. But throw in anything
  676. 29:33unexpected and they just fell apart.
  677. 29:36These systems had no common sense and
  678. 29:38every time the world changed a bit, a
  679. 29:40human had to sit down and write new
  680. 29:43rules by hand. That got very expensive
  681. 29:46very fast. And then in 1987, the market
  682. 29:49for specialized AI computers, the
  683. 29:51so-called lisp machines, collapsed
  684. 29:54almost overnight. Ordinary desktop
  685. 29:56computers had become powerful enough to
  686. 29:59do the same job for a fraction of the
  687. 30:01price. Dozens of AI companies went bust
  688. 30:04and in 1992, Japan's fifth generation
  689. 30:07project ended without reaching its
  690. 30:09goals. This was the second AI winter.
  691. 30:12And this time AI became such a dirty
  692. 30:15word that researchers branded their work
  693. 30:18as machine learning informatics calling
  694. 30:21their work anything but artificial
  695. 30:23intelligence. And remember how I said at
  696. 30:26the start that AI is basically a
  697. 30:28marketing label. Well, when the label
  698. 30:30stopped selling, they just changed it.
  699. 30:33But while the rules camp was crashing,
  700. 30:35something was quietly happening in
  701. 30:37background. remember the learning camp,
  702. 30:40the neural network people. After
  703. 30:42Minsky's book, most of them had given
  704. 30:44up, but not all of them. In 1986, three
  705. 30:47researchers, David Rumlhart, Doffrey
  706. 30:50Hinton, and Ronald Williams, published a
  707. 30:53paper that made a technique called back
  708. 30:55propagation famous. Remember how Minsky
  709. 30:59said adding more layers probably
  710. 31:01wouldn't help? The problem was that
  711. 31:03nobody knew how to train a network with
  712. 31:05lots of layers. Back propagation solved
  713. 31:08that. It's a way of working out exactly
  714. 31:10how much each of those dials contributed
  715. 31:13to a mistake layer by layer so you know
  716. 31:17how to nudge every single one. And
  717. 31:19suddenly the simple problem the
  718. 31:21perceptron couldn't solve like telling
  719. 31:23whether two dots were the same color
  720. 31:25became solvable. Hinton by the way will
  721. 31:28come up again a lot. Anyway, a year
  722. 31:31later, he left the US for the University
  723. 31:33of Toronto, partly because he didn't
  724. 31:35want his research funded by the American
  725. 31:37military. And Toronto became the hideout
  726. 31:40for the few people still betting on
  727. 31:42neural networks. And neural networks
  728. 31:45actually started working in the real
  729. 31:47world. At Bell Labs in the US, a young
  730. 31:49French researcher called Yan Lun built a
  731. 31:53neural network that could read
  732. 31:55handwritten numbers. By the late 1990s,
  733. 31:58banks were using it to read a
  734. 31:59significant share of all checks in the
  735. 32:01US. But most of the AI world still
  736. 32:04didn't take neural networks, aka the
  737. 32:06learning camp, seriously because in
  738. 32:081997, the rules camp got its biggest win
  739. 32:12ever. IBM's Deep Blue beat world chess
  740. 32:15champion Gary Kasparov. The big promise
  741. 32:17that had been made 40 years earlier,
  742. 32:21and the chess computer Deep Blue didn't
  743. 32:24learn anything. It was a brute force
  744. 32:26machine checking 200 million chess
  745. 32:28positions per second using rules written
  746. 32:31by humans. So at the end of the 1990s,
  747. 32:34it looked like the rules camp at one.
  748. 32:36But here's the twist. The learning camp
  749. 32:39hadn't been wrong. It had just been
  750. 32:41starved. Remember how chachip got past
  751. 32:43the sparsity problem? Billions of dials
  752. 32:45trained on trillions of words. Neural
  753. 32:47networks need two things to work. huge
  754. 32:50amounts of examples to learn from and
  755. 32:53huge amounts of computing power to nudge
  756. 32:55all those dials. And in the 1990s,
  757. 32:58neither existed. There was no internet
  758. 33:01full of text and photos to learn from.
  759. 33:03And computers were nowhere near fast
  760. 33:06enough. So neural networks were missing
  761. 33:08exactly the two things they needed, the
  762. 33:10data to learn from and the power to
  763. 33:11learn from it. But the idea behind
  764. 33:13today's AI was already there. It had
  765. 33:16been there since Rosenlat's perceptron
  766. 33:18in 1958. It just had to wait for the
  767. 33:20world to catch up. And when that finally
  768. 33:23happened, the missing computing power
  769. 33:25didn't come from a university or a
  770. 33:27supercomput. It came from video games.
  771. 33:30If you watched my chip video, you
  772. 33:32already know what a GPU is. If not,
  773. 33:35here's the quick version. Every computer
  774. 33:37has a CPU, a central processing unit.
  775. 33:40That's its brain, and it's good at doing
  776. 33:43one complicated thing at a time. But
  777. 33:45video games have a different problem. To
  778. 33:47draw a game on your screen, you have to
  779. 33:50calculate the color of millions of
  780. 33:51pixels many times per second. Each
  781. 33:54calculation is simple, but there are
  782. 33:56millions of them, and they all have to
  783. 33:58happen at once. So, companies like
  784. 34:00Nvidia built a different kind of chip
  785. 34:03for that, the GPU, the graphics
  786. 34:05processing unit. Instead of doing one
  787. 34:08complicated thing at a time, it does
  788. 34:10thousands of simple calculations at the
  789. 34:13same time. And here's the thing.
  790. 34:15Remember the dials? Training a neural
  791. 34:18network is basically billions of very
  792. 34:20simple calculations. Multiply, add,
  793. 34:23nudge, repeat. Exactly the kind of job a
  794. 34:26GPU was built for. A chip designed to
  795. 34:28make video games look pretty turned out
  796. 34:31to be the perfect machine for AI. And
  797. 34:34this is why Nvidia is such a valuable
  798. 34:36company. And I really recommend you go
  799. 34:39watch the video about chips I made if
  800. 34:41you really want to understand how these
  801. 34:43chips are made and why the US and China
  802. 34:46are fighting over them. But back to the
  803. 34:48story because there was one big problem.
  804. 34:51GPUs only understood graphics. By the
  805. 34:54early 2000s, some scientists had noticed
  806. 34:57that their work looked a lot like a
  807. 34:59video game problem. weather models,
  808. 35:01physics simulations, chemistry, all of
  809. 35:03it was millions of symbol calculations
  810. 35:06happening at once. And gaming GPUs were
  811. 35:09cheap and incredibly powerful at exactly
  812. 35:11that. But to use a video game GPU for
  813. 35:14science, researchers had to trick it.
  814. 35:17They had to disguise their maths as a
  815. 35:19picture, pretend their numbers were
  816. 35:22pixel colors, let the card draw them,
  817. 35:25and then read the image back out as
  818. 35:27results. It worked, but it was a
  819. 35:29nightmare. So, in 2006, Nvidia announced
  820. 35:32CUDA. It's software that lets
  821. 35:34programmers give a GPU any kind of
  822. 35:37calculation directly without dressing it
  823. 35:39up as graphics. And this was a huge
  824. 35:42gamble. Nvidia CEO Yansen Huang built
  825. 35:46CUDA support into its ordinary gaming
  826. 35:48GPUs. That made every chip more
  827. 35:50expensive for a market that barely
  828. 35:52existed. It was only a handful of
  829. 35:55scientists, oil companies, and banks.
  830. 35:57And for years that gamble didn't pay off
  831. 36:00and investors kept asking why a gaming
  832. 36:02company was spending fortunes on this.
  833. 36:05And then came 2012
  834. 36:08in Toronto. Draw Hinton the man who'
  835. 36:10kept neural networks alive through the
  836. 36:13winter had two students Alex Kresfki and
  837. 36:17Ilia Satsk.
  838. 36:19And they had an idea. What if you took a
  839. 36:22neural network, made it much bigger and
  840. 36:26deeper than anyone had dared before, and
  841. 36:29trained it on gaming graphics cards?
  842. 36:32Quick thing before we move on. There's
  843. 36:34no team behind this channel. The
  844. 36:36research, the writing, the filming, the
  845. 36:38editing, that's all just me, which is
  846. 36:42why I can only make about one video a
  847. 36:44month instead of one a week. And
  848. 36:47honestly, one deeply researched video a
  849. 36:50month doesn't pay nearly as well as
  850. 36:52pumping out quick videos every few days.
  851. 36:55The only reason I can keep making videos
  852. 36:57like this is my patron supporters. So, a
  853. 37:00huge thank you to all of you. And if
  854. 37:03you'd like to support the channel, too,
  855. 37:06the link is in the description. Anyway,
  856. 37:09let's get back to the video because we
  857. 37:12were talking about Joffrey Hinton, Alex
  858. 37:14Krosfki, and Ilascover
  859. 37:16and their idea. What if you took a
  860. 37:18neural network, made it much bigger and
  861. 37:21deeper than anyone had dared before, and
  862. 37:23drained it on gaming graphics cards?
  863. 37:26Well, Alex Kreski bought two Nvidia
  864. 37:30gaming cards around $500 each, put them
  865. 37:34into the computer in his bedroom, and
  866. 37:36wrote the code to make them train a
  867. 37:38neural network. It took about a week of
  868. 37:40the cards crunching numbers non-stop.
  869. 37:44But remember, neural networks don't just
  870. 37:46need computing power, they also need
  871. 37:48examples, lots of them. And that's where
  872. 37:50the second ingredient came in. A few
  873. 37:53years earlier, the Stanford professor
  874. 37:55Fifi Lee had built a giant collection of
  875. 37:58photos called Imageet. Over a million
  876. 38:01images, each labeled by hand with what's
  877. 38:03in it. Dog, car, strawberry, a thousand
  878. 38:06different categories. And all those
  879. 38:09labels, well, they were added by
  880. 38:11thousands of low paid online workers
  881. 38:13clicking through images one by one. And
  882. 38:15since 2010, there had been a yearly
  883. 38:17competition. whose program could
  884. 38:20recognize the most imageet photos
  885. 38:22correctly. Until then, the best programs
  886. 38:25were built the old way. Humans had
  887. 38:28carefully designed by hand what the
  888. 38:31computer should look for. Edges, shapes,
  889. 38:33textures, basically the rules camp
  890. 38:36approach. But in 2012, Alex Kresfk's
  891. 38:39network later named Alexet entered and
  892. 38:44it destroyed everyone. The best
  893. 38:47traditional program got about 26% of its
  894. 38:50answers wrong. Alex Neck got about 15%
  895. 38:53wrong. In a competition where the
  896. 38:56winners usually improved by a couple of
  897. 38:57percentage points, Alex Net nearly half
  898. 39:01the error rate. And nobody had told it
  899. 39:04what an edge or a shape was. It had
  900. 39:06learned all of that by itself from
  901. 39:08examples. Rosenblat's idea from 1958
  902. 39:11finally had what it needed, data and
  903. 39:14computing power. And big tech noticed
  904. 39:18immediately just a few weeks later in
  905. 39:21December 2012. Hinton Kresfki and
  906. 39:24Satskcover founded a company. It was
  907. 39:26called DNN Research. Three people, no
  908. 39:30product, no revenue, no business plan,
  909. 39:33just one paper. And they did something
  910. 39:36pretty wild. That December, the world's
  911. 39:38biggest AI conference was taking place
  912. 39:41in a casino hotel at Lake Tahoe, Nevada.
  913. 39:45And from this hotel room, Hinton run a
  914. 39:47secret auction by email. The biders
  915. 39:50Google, Microsoft, the Chinese tech
  916. 39:53giant BU, and a small London startup
  917. 39:56called Deep Mind. The bids kept climbing
  918. 39:59by millions at a time. And when it
  919. 40:02reached $44 million, Hinton stopped the
  920. 40:05auction and sold to Google. Not even
  921. 40:08necessarily to the highest possible
  922. 40:09bidder, who just chose Google. $44
  923. 40:13million for three researchers and one
  924. 40:15idea that most of the field had laughed
  925. 40:18at for decades. And this was a turning
  926. 40:20point. For 70 years, AI had mostly been
  927. 40:25something that happened at universities,
  928. 40:27funded by governments and the military.
  929. 40:30From now on, the best minds in AI would
  930. 40:33increasingly work for big tech. And
  931. 40:35remember that small London startup in
  932. 40:37the auction, Deep Mind. It was founded
  933. 40:40by a man called Deus Asabis, a former
  934. 40:43child chess prodigy who designed video
  935. 40:46games as a teenager and then got a PhD
  936. 40:49in neuroscience. His mission was as
  937. 40:52modest as the Dartmouth crowds. Solve
  938. 40:55intelligence. In 2013, Deep Mind showed
  939. 40:58off a program that taught itself to play
  940. 41:01old Atari video games. Nobody told it
  941. 41:04the rules. It just saw the pixels on
  942. 41:06screen and the score unlearned by trial
  943. 41:09and error. After a while, it was better
  944. 41:11than most humans. And just weeks later,
  945. 41:14in early 2014, Google bought DeepMind
  946. 41:17for around half a billion dollars. So
  947. 41:20now Google owned the Hinton team as well
  948. 41:22as DeepMind. A lot of the world's best
  949. 41:25AI researchers were now working for one
  950. 41:28company. And then in 2016, DeepMind went
  951. 41:32after the game that was considered AI's
  952. 41:34holy grail, Go. It's an ancient Chinese
  953. 41:38board game, and it has more possible
  954. 41:40board positions than there are atoms in
  955. 41:43the universe. So, you can't brute force
  956. 41:45it like Deep Blue did with chess. You
  957. 41:47need something closer to intuition.
  958. 41:50Deep Mind's program, Alph Go, played the
  959. 41:53world champion Lee Settle in Soul.
  960. 41:56According to Deep Mind, more than 200
  961. 41:58million people watched. And in the
  962. 42:01second game, Alph Go played a move so
  963. 42:03strange that commentators thought it was
  964. 42:05a mistake. Move 37. Lee Settle looked
  965. 42:10stunned and took almost 15 minutes to
  966. 42:13respond. It turned out to be brilliant,
  967. 42:16a move no human would have ever played.
  968. 42:19Alph Go ended up beating Lee Saddle by
  969. 42:224-1. And 3 years later, Lee had all
  970. 42:25retired from Professional Go, saying AI
  971. 42:28was an entity that cannot be defeated.
  972. 42:31So by 2016, Google had the best
  973. 42:34researchers, the best AI lab, and a
  974. 42:37mountain of GPUs. AI could now recognize
  975. 42:40images, and beat humans at games. But
  976. 42:42language was still a problem. Back then,
  977. 42:45AI read text the way you'd read through
  978. 42:47a keyhole, one word at a time, left to
  979. 42:50right. And by the end of a long
  980. 42:52sentence, it had often forgotten how the
  981. 42:54sentence started. It was also slow
  982. 42:56because reading one word after the other
  983. 42:59meant you couldn't use the GPU's big
  984. 43:01strength, doing thousands of things at
  985. 43:04the same time. And then in 2017, a group
  986. 43:08of eight researchers at Google published
  987. 43:10a paper with a slightly cheeky title,
  988. 43:13Attention Is All You Need. It's a nod to
  989. 43:16the Beatles song, All You Need Is Love.
  990. 43:19their idea. Stop reading one word at a
  991. 43:22time. Look at the whole text at once and
  992. 43:24for every word work out which other
  993. 43:26words matter most for understanding it.
  994. 43:29That's what attention means. Take the
  995. 43:32sentence. The cat didn't cross the road
  996. 43:34because it was too tired. What does it
  997. 43:37mean? The cat or the road? You know
  998. 43:41immediately it's the cat. Attention is
  999. 43:44how the model learns to make that same
  1000. 43:45connection. When it reads it now, it
  1001. 43:50pays the most attention to cat. And
  1002. 43:52because it looks at everything at once,
  1003. 43:54it could run perfectly on GPUs with
  1004. 43:57thousands of calculations in parallel.
  1005. 43:59So you could train it on way more text
  1006. 44:01than ever before. They called it the
  1007. 44:03transformer. And that's the T in chatbt,
  1008. 44:07generative pre-trained transformer. It's
  1009. 44:10the reason chach can look at your whole
  1010. 44:12conversation at once, which we talked
  1011. 44:15about at the start. The paper has since
  1012. 44:18been cited well over a 100,000 times.
  1013. 44:21It's one of the most important research
  1014. 44:23papers of the century and was invented
  1015. 44:25at Google. And what Google did next
  1016. 44:28might be the biggest blunder in tech
  1017. 44:31history. Because here's the irony.
  1018. 44:34Google had invented the technology
  1019. 44:37behind chat GBT, but it didn't release a
  1020. 44:40chatbot. Google did end up using
  1021. 44:43transformers. It quietly built them into
  1022. 44:46Google search to better understand what
  1023. 44:47people were typing. But releasing a
  1024. 44:50chatbot that talks to you, that's where
  1025. 44:53Google hesitated. Two Google engineers,
  1026. 44:56Noam Shazair, one of the eight authors,
  1027. 44:59and his colleague Daniel Defrighas did
  1028. 45:02build a chatbot based on the
  1029. 45:03transformer. The two Google employees
  1030. 45:06pushed for years to release it to the
  1031. 45:07public. Google said no. Officially, the
  1032. 45:11worry was safety. What if it said
  1033. 45:13something offensive or wrong under
  1034. 45:15Google's name? But many people think
  1035. 45:17there was another reason, too. Google
  1036. 45:20makes most of its money from search ads.
  1037. 45:23If a chatbot just gives you the answer,
  1038. 45:25you don't click on any links. So, why
  1039. 45:27would Google built the thing that could
  1040. 45:29kill its own business? So, the authors
  1041. 45:31of the most important AI paper of the
  1042. 45:33century did what frustrated talent in
  1043. 45:36Silicon Valley does. They left. One by
  1044. 45:40one, all eight of them left Google. Most
  1045. 45:43of them started their own companies.
  1046. 45:46Aiden Gomez, who'd been an intern and
  1047. 45:48stellar student when the paper was
  1048. 45:50written, co-founded Coher. Leon Jones,
  1049. 45:53moved to Tokyo and co-founded Sakana AI.
  1050. 45:57And Noam Jazer co-founded Character AI,
  1051. 46:00a chatbot company. And one of the eight,
  1052. 46:03Lucas Kaiser, went to work for a small
  1053. 46:06lab called Open AI. Before we get into
  1054. 46:11open AI, I want to get back to one funny
  1055. 46:13thing. Because in 2024, with Google
  1056. 46:16desperately trying to catch up in the AI
  1057. 46:18race, it paid around $2.7 billion in a
  1058. 46:22deal with Character AI, largely to get
  1059. 46:25Noam Shazer back. So, Google basically
  1060. 46:28paid billions to rehire a man to work on
  1061. 46:30a technology he had invented while
  1062. 46:32working at Google. Okay, but let's get
  1063. 46:35to open AI. What was this small lab that
  1064. 46:38was called OpenAI? In the summer of
  1065. 46:412015, Sam Oldman, who was running the
  1066. 46:44famous startup accelerator Y Combinator
  1067. 46:47at the time, hosted a private dinner at
  1068. 46:49a hotel in Silicon Valley. Among the
  1069. 46:52guests were Elon Musk, Greg Brockman,
  1070. 46:55then the CTO of the payment company
  1071. 46:57Stripe, and Ilia Satskever. Yes, the
  1072. 47:01same Ilia Satskever from the AlexNet
  1073. 47:03team and the $44 million auction. He was
  1074. 47:07working at Google at the time and all of
  1075. 47:10them shared one fear that Google with
  1076. 47:13Hinton's team and Deep Mind was going to
  1077. 47:17own the future of AI. One company
  1078. 47:20controlling what might become the most
  1079. 47:22powerful technology in history. So in
  1080. 47:25December 2015, they launched the
  1081. 47:27opposite, Open AI, a nonprofit. No
  1082. 47:31shareholders, no profit motive. Its
  1083. 47:33mission was to make sure that artificial
  1084. 47:36general intelligence AI as smart as a
  1085. 47:38human benefits all of humanity, and it
  1086. 47:41would share its research openly, hence
  1087. 47:44the name. And to show they were serious,
  1088. 47:47they announced a funding pledge of $1
  1089. 47:48billion. But a nonprofit with a mission
  1090. 47:51to save humanity quickly ran into a very
  1091. 47:53practical problem. Money. Remember what
  1092. 47:56neural networks need? Huge amount of
  1093. 47:59data and huge amounts of computing
  1094. 48:01power. And computing power costs a
  1095. 48:04fortune. And of that billion dollar
  1096. 48:05pledge, much less actually came in. Then
  1097. 48:08came a power struggle. In 2018, Elon
  1098. 48:11Musk left. According to OpenAI, he'd
  1099. 48:13wanted more control of the company, and
  1100. 48:15when he didn't get it, he walked away.
  1101. 48:17Musk tells the story differently and the
  1102. 48:20two sides have been fighting over it in
  1103. 48:22court. But nonetheless, OpenAI had money
  1104. 48:25troubles. So in 2019, OpenAI did
  1105. 48:28something that would have been
  1106. 48:29unthinkable at the start. It created a
  1107. 48:32for-profit company under the nonprofit
  1108. 48:34with investors profits capped at 100
  1109. 48:36times their investment.
  1110. 48:39So if you put in $10 million, you could
  1111. 48:42only make a billion.
  1112. 48:44Very modest. It's a bit like going on a
  1113. 48:48diet and limiting yourself to a 100
  1114. 48:50donuts a day. But the donut diet did its
  1115. 48:52job because now OpenAI could take money
  1116. 48:55from investors and the first big one
  1117. 48:57showed up fast. Just a few months later,
  1118. 49:00Microsoft invested $1 billion. But let's
  1119. 49:03jump back to an important question. Why
  1120. 49:05did OpenAI suddenly need so much money?
  1121. 49:08Well, it was because of one big bet.
  1122. 49:10Inside OpenAI, the chief scientist, Ilia
  1123. 49:13Satskev, had become convinced of
  1124. 49:15something.
  1125. 49:17Maybe AI didn't need some brilliant new
  1126. 49:19idea. Maybe you just had to take the
  1127. 49:21transformer and make it bigger. More
  1128. 49:24dials, more data, more computing power.
  1129. 49:28Satsk was so convinced that he
  1130. 49:30reportedly led staff in a chant at
  1131. 49:32company events. Feel the AGI. And in
  1132. 49:352020, Open AI researchers published a
  1133. 49:38paper that seemed to prove him right.
  1134. 49:40It's called scaling laws for neural
  1135. 49:42language models. And it showed that if
  1136. 49:45you make the model bigger and give it
  1137. 49:46more data and more computing power, it
  1138. 49:49gets better predictably every single
  1139. 49:52time. And you can see that bet in
  1140. 49:55OpenAI's models. They call them GPT.
  1141. 49:58GPD1 in 2018 had 117 million dials. It
  1142. 50:03was trained on about 7,000 unpublished
  1143. 50:06books, lots of them self-published
  1144. 50:08romance novels. Also, none of the
  1145. 50:11authors were ever asked. GPD2 in 2019
  1146. 50:15had 1.5 billion dials and OpenAI
  1147. 50:19announced it was too dangerous to
  1148. 50:20release in full because it might be used
  1149. 50:22to mass-produce fake news, which of
  1150. 50:25course made headlines everywhere. Sound
  1151. 50:28familiar? The Perceptrons press
  1152. 50:30conference all over again. GPT3 in 2020
  1153. 50:34had 175 billion dials trained on a huge
  1154. 50:38chunk of the internet on a supercomput
  1155. 50:41Microsoft build specifically for OpenAI
  1156. 50:44with 10,000 Nvidia GPUs. So we went from
  1157. 50:47117 million to 175 billion dials in 2
  1158. 50:51years. That's around 1,500 times bigger.
  1159. 50:55But there was still a problem. GBT3 was
  1160. 50:58basically a giant autocomplete. Remember
  1161. 51:01my paper cups? It was the same game just
  1162. 51:04at a gigantic scale. So if you typed in
  1163. 51:07a question, it didn't necessarily answer
  1164. 51:09it. It just continued the text,
  1165. 51:12sometimes with more questions as if it
  1166. 51:14was writing an exam paper, which was
  1167. 51:16obviously impressive, but not very
  1168. 51:19useful. So OpenAI added one more
  1169. 51:21ingredient, humans. People wrote example
  1170. 51:24answers and ranked the model's responses
  1171. 51:27from best to worst. And the model was
  1172. 51:29trained to give more of the kind of
  1173. 51:31answers humans liked. This is called
  1174. 51:34reinforcement learning from human
  1175. 51:36feedback or RLHF. And it's what turned
  1176. 51:39autocomplete into an assistant. Who
  1177. 51:41those humans were and how they were
  1178. 51:43treated we'll get to later. But for now,
  1179. 51:45let's look at the moment you probably
  1180. 51:47first heard of chat GPT. The moment it
  1181. 51:49went mainstream. On the 30th of October
  1182. 51:522022, OpenAI released a chatbot built on
  1183. 51:55GPT 3.5 as a free research preview. They
  1184. 52:00called it Chat GPT. Sam Oldman announced
  1185. 52:03it with a pretty casual tweet. Nobody
  1186. 52:06inside OpenAI expected much. 5 days
  1187. 52:08later, it had a million users. Within
  1188. 52:10about 2 months, according to estimates,
  1189. 52:13100 million. At the time, the fastest
  1190. 52:15growing consumer app in history. And at
  1191. 52:18Google, management reportedly declared a
  1192. 52:20code red. The company that had invented
  1193. 52:23the technology behind Chhat GPT now had
  1194. 52:26to scramble to catch up. But let's jump
  1195. 52:29back a little. Remember those humans who
  1196. 52:32helped turn GPT into Chhat GPT? Let's
  1197. 52:34talk about them because the story of
  1198. 52:36Chhat GPT isn't just a story about
  1199. 52:39genius researchers in San Francisco.
  1200. 52:41Remember, GPT3 learned from a huge chunk
  1201. 52:44of the internet. And the internet isn't
  1202. 52:46just Wikipedia and recipes. It's also
  1203. 52:48full of hate speech, violence, and
  1204. 52:50abuse. And a model trained on all of it
  1205. 52:53can repeat all of it. So before open AI
  1206. 52:56could release a friendly chatbot, it
  1207. 52:58needed a filter. A second AI that could
  1208. 53:00spot toxic content and block it. But to
  1209. 53:03build that filter, you need examples.
  1210. 53:05Thousands of examples of the worst
  1211. 53:07things human have ever written. Each one
  1212. 53:10read and labeled by a person. So in late
  1213. 53:132021, OpenAI hired a company called Sama
  1214. 53:17which employed workers in Nairobi,
  1215. 53:20Kenya. Sama marketed itself as an
  1216. 53:23ethical AI company. The workers drop
  1217. 53:26read text describing things like murder,
  1218. 53:28suicide, torture, and child sexual abuse
  1219. 53:31and label it so the filter could learn
  1220. 53:33to recognize it. According to an
  1221. 53:36investigation by Time magazine, they
  1222. 53:37read between 150 and 250 of these
  1223. 53:40passages per 9hour shift. And their
  1224. 53:43take-home pay, according to Time,
  1225. 53:45between roughly 1.32 and $2 an hour.
  1226. 53:49Sama disputes those figures. Workers
  1227. 53:52told time the work left them
  1228. 53:53traumatized. One described as torture.
  1229. 53:56In early 2022, Sama ended its work with
  1230. 53:58Open AI months ahead of schedule. And
  1231. 54:01that was just one drop. Remember RLHF,
  1232. 54:05the humans rating chat's answers so it
  1233. 54:08learns to be helpful? That's another
  1234. 54:09layer of hidden labor. When OpenAI first
  1235. 54:12developed this technique, it hired a few
  1236. 54:14dozen contractors through outsourcing
  1237. 54:17platforms like Upwork and a company
  1238. 54:19called Scale AI to write example answers
  1239. 54:22and rank the models responses. And today
  1240. 54:25that work has turned into a huge global
  1241. 54:27industry. Hundreds of thousands of
  1242. 54:29people in places like Kenya, the
  1243. 54:31Philippines, and Venezuela sit at home
  1244. 54:33labeling images, writing chatbot
  1245. 54:35answers, and correcting AI mistakes task
  1246. 54:37by task, often for a few dollars an
  1247. 54:40hour. The Washington Post reported on
  1248. 54:41workers on Scale AI's platforms, Remo
  1249. 54:44tasks in the Philippines, whose payments
  1250. 54:46were delayed, reduced, or never arrived
  1251. 54:49at all. And the thing is in 2025, Meta
  1252. 54:52invested around $14 billion in scale AI.
  1253. 54:54The company built on gig workers in the
  1254. 54:57global south was suddenly worth almost
  1255. 54:59$30 billion. So behind every intelligent
  1256. 55:02machine, there's a huge invisible army
  1257. 55:04of humans clicking, rating, correcting,
  1258. 55:07making the AI look smarter than it
  1259. 55:09actually is. Remember how tragic he read
  1260. 55:11a huge chunk of the internet, websites,
  1261. 55:13Wikipedia, Reddit, millions of books?
  1262. 55:16Well, here's a question nobody really
  1263. 55:18asked at the time. Did anyone ask
  1264. 55:20permission? Mostly no. The first GPT was
  1265. 55:24trained on thousands of self-published
  1266. 55:26romance books whose authors were never
  1267. 55:28asked. And it went on from there. News
  1268. 55:30articles, blog posts, forum discussions,
  1269. 55:33artwork, code. If it was online, it was
  1270. 55:36fair game. The economist Cecilia Recap
  1271. 55:39calls companies like OpenAI, Anthropic,
  1272. 55:41and Google intellectual monopolies. The
  1273. 55:44knowledge that millions of people
  1274. 55:46created together for free gets swallowed
  1275. 55:48up and turned into private products that
  1276. 55:50a handful of companies own and sell back
  1277. 55:53to us. And it wasn't just stuff that was
  1278. 55:56freely online. According to court
  1279. 55:58documents, MetaTrained its AI on books
  1280. 56:00from a pirate library called Liben.
  1281. 56:02Millions of copyrighted books downloaded
  1282. 56:05illegally. And in a lawsuit against
  1283. 56:07Anthropic, court documents showed it had
  1284. 56:09also downloaded millions of pirated
  1285. 56:12books. and it bought millions of
  1286. 56:14physical books, cut off their spines and
  1287. 56:16scanned every page. And the irony is
  1288. 56:18since companies realized how valuable
  1289. 56:20all this human writing was, the platform
  1290. 56:23started charging for it. Reddit now gets
  1291. 56:26paid by Google and OpenAI to let them
  1292. 56:28train on Reddit posts. Posts written for
  1293. 56:31free by Reddit users who get nothing.
  1294. 56:35Okay, so AI is built on human labor and
  1295. 56:38on human knowledge. But what does it do
  1296. 56:40to the humans who have to work with it?
  1297. 56:43And to explain that, there's a great
  1298. 56:45idea from the writer Cory Dr. Row. And
  1299. 56:47it starts with chess. Remember Deep Blue
  1300. 56:50being Kasparov. Afterwards, Kasparov
  1301. 56:54came up with a new kind of chess. A
  1302. 56:56human and a computer playing together as
  1303. 56:58a team. He called them centators like
  1304. 57:01the creature from Greek mythology. Half
  1305. 57:03human, half horse. The human makes the
  1306. 57:06decisions and the machine gives them
  1307. 57:08superpowers. That's how AI is usually
  1308. 57:10sold to us. You stay in charge and AI
  1309. 57:13makes you faster and better. But Dr. O
  1310. 57:15says that for a lot of workers, it's
  1311. 57:17turning out the other way round. He
  1312. 57:19calls it a reverse centur. The machine
  1313. 57:22is in charge and the human is just there
  1314. 57:24to keep up with it, fix its mistakes,
  1315. 57:26and take the blame when it goes wrong.
  1316. 57:29According to a report in the New York
  1317. 57:31Times, some programmers at Amazon say
  1318. 57:33their jobs have changed. less writing
  1319. 57:36code, more racing to check code written
  1320. 57:38by AI under pressure to work faster and
  1321. 57:40faster. And there's a quieter risk. In
  1322. 57:422025, a study in Poland looked at
  1323. 57:45doctors doing colonoscopies with an AI
  1324. 57:47assistant that helped spot early signs
  1325. 57:49of cancer. After a few months of getting
  1326. 57:52used to the AI when the doctors worked
  1327. 57:54without it, they spotted fewer warning
  1328. 57:56signs than before. Their detection rate
  1329. 57:59dropped from about 28% to 22%.
  1330. 58:03So AI doesn't just need humans. It can
  1331. 58:06quietly change what humans are for. From
  1332. 58:09doing the work to babysitting the
  1333. 58:11machine and what that means for society
  1334. 58:13is what we'll explore in the next video
  1335. 58:15of this series. Anyway, thanks for
  1336. 58:18watching and see you soon. Bye.

About this transcript

This page contains the full transcript of not another AI video by Anna Bocca, generated from the public captions YouTube serves with the video. The transcript has 8,922 words across 1,336 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.