1 00:00:00,228 --> 00:00:02,027 If his father had lived four more years 2 00:00:02,028 --> 00:00:04,351 Anthropic might never have existed 3 00:00:04,508 --> 00:00:07,230 An AI company valued at 900 billion dollars 4 00:00:07,267 --> 00:00:11,111 Its founder was not a garage born technical genius 5 00:00:11,268 --> 00:00:14,584 Nor a serial entrepreneur lifted by a trend 6 00:00:14,821 --> 00:00:17,420 He was working in a Princeton laboratory 7 00:00:17,420 --> 00:00:20,703 As a biophysics PhD dissecting retinas 8 00:00:21,141 --> 00:00:22,499 He switched careers into AI 9 00:00:22,501 --> 00:00:24,624 Not because he saw a business opportunity 10 00:00:24,821 --> 00:00:26,583 But because his father died 11 00:00:26,861 --> 00:00:27,984 More precisely 12 00:00:28,100 --> 00:00:32,184 His father died at the wrong time four years too early 13 00:00:32,740 --> 00:00:33,703 Four years later 14 00:00:33,780 --> 00:00:36,783 The rare disease that took his father 15 00:00:36,941 --> 00:00:40,463 Saw its death rate plunge from 50 percent to 5 percent 16 00:00:41,060 --> 00:00:43,024 If science had moved just a little faster 17 00:00:43,100 --> 00:00:45,748 His father might still be alive today 18 00:00:46,025 --> 00:00:46,367 Hello 19 00:00:46,385 --> 00:00:46,847 Hello everyone 20 00:00:46,944 --> 00:00:47,907 I am Leo Wang 21 00:00:48,025 --> 00:00:50,947 Today I want to talk about Dario Amodei 22 00:00:51,065 --> 00:00:53,267 The founder and CEO of Anthropic 23 00:00:53,465 --> 00:00:56,708 Not because his company is nearing a trillion dollar valuation 24 00:00:56,825 --> 00:00:58,708 And not because the Claude system he built 25 00:00:58,785 --> 00:01:01,907 Is gaining more momentum than ChatGPT among programmers 26 00:01:02,105 --> 00:01:04,568 But because he embodies an extremely rare contradiction 27 00:01:05,104 --> 00:01:08,587 He is among the loudest voices warning that AI is dangerous 28 00:01:08,625 --> 00:01:11,867 And also one of the most aggressive forces advancing AI 29 00:01:12,145 --> 00:01:14,167 How can both things be true at once 30 00:01:14,625 --> 00:01:15,904 To understand this contradiction 31 00:01:15,905 --> 00:01:18,114 We must go back to when he was 23 32 00:01:18,512 --> 00:01:19,391 In 2006 33 00:01:19,391 --> 00:01:21,674 His father Riccardo died 34 00:01:21,831 --> 00:01:24,075 Riccardo was an Italian American leatherworker 35 00:01:24,152 --> 00:01:26,150 Together with the Jewish American mother of Dario 36 00:01:26,152 --> 00:01:28,155 He raised two children in San Francisco 37 00:01:28,391 --> 00:01:30,155 Dario later said in an interview 38 00:01:30,232 --> 00:01:32,995 His parents taught him right from wrong 39 00:01:33,031 --> 00:01:36,351 And what truly matters in this world 40 00:01:36,627 --> 00:01:39,910 Riccardo had long battled a rare disease 41 00:01:40,228 --> 00:01:41,106 At that time 42 00:01:41,107 --> 00:01:43,750 The disease had a death rate of about 50 percent 43 00:01:44,267 --> 00:01:46,671 Dario watched helplessly as his father died 44 00:01:46,868 --> 00:01:47,250 Then 45 00:01:47,547 --> 00:01:48,591 Less than four years later 46 00:01:48,627 --> 00:01:50,791 A new medical breakthrough arrived 47 00:01:50,907 --> 00:01:54,775 The cure rate for the disease soared to 95 percent 48 00:01:54,973 --> 00:01:58,055 Dario later told a reporter something I never forgot 49 00:01:58,333 --> 00:01:58,635 He said 50 00:01:58,635 --> 00:02:01,092 Someone developed a treatment for this disease 51 00:02:01,093 --> 00:02:02,256 And saved some lives 52 00:02:02,533 --> 00:02:04,295 But many more could have been saved 53 00:02:04,533 --> 00:02:06,376 The words sound calm 54 00:02:06,613 --> 00:02:07,596 But think carefully about them 55 00:02:07,973 --> 00:02:08,976 What he meant was 56 00:02:09,173 --> 00:02:10,316 The speed of scientific progress 57 00:02:10,573 --> 00:02:12,135 Is the boundary between life and death 58 00:02:12,372 --> 00:02:12,835 A little faster 59 00:02:13,133 --> 00:02:14,376 And his father lives 60 00:02:14,492 --> 00:02:14,915 A little slower 61 00:02:15,212 --> 00:02:16,415 And his father dies 62 00:02:16,613 --> 00:02:17,976 This was not a philosophical question 63 00:02:18,052 --> 00:02:19,976 It was a fact he lived through 64 00:02:20,649 --> 00:02:21,607 From then on 65 00:02:21,608 --> 00:02:23,971 The force driving Dario changed 66 00:02:24,129 --> 00:02:28,332 At Princeton he moved from theoretical physics into biology 67 00:02:28,409 --> 00:02:31,651 Studying the retina in Professor Michael Berry laboratory 68 00:02:31,849 --> 00:02:33,688 He did not want to become an eye doctor 69 00:02:33,689 --> 00:02:38,251 He wanted to use the retina to understand the whole nervous system 70 00:02:38,369 --> 00:02:41,212 He even found the available measurement tools inadequate 71 00:02:41,249 --> 00:02:43,932 So he invented a new sensor himself 72 00:02:44,089 --> 00:02:46,811 His dissertation won the Hertz Thesis Prize 73 00:02:46,849 --> 00:02:49,371 An award specifically recognizing academic research 74 00:02:49,409 --> 00:02:51,772 That discovers practical value 75 00:02:52,778 --> 00:02:54,857 But the deeper Dario went into biology 76 00:02:54,858 --> 00:02:56,721 The more he recognized a harsh reality 77 00:02:57,219 --> 00:03:01,062 Biological systems far exceed the understanding of one human mind 78 00:03:01,659 --> 00:03:02,041 He said 79 00:03:02,041 --> 00:03:04,017 To truly understand all of this 80 00:03:04,018 --> 00:03:07,101 You need hundreds or thousands of researchers working together 81 00:03:07,338 --> 00:03:08,122 The intelligence of one person 82 00:03:08,379 --> 00:03:10,058 Even the intelligence of a genius 83 00:03:10,058 --> 00:03:12,502 Is a drop in the ocean against disease 84 00:03:12,719 --> 00:03:13,458 So 85 00:03:13,458 --> 00:03:16,962 Could any tool replace those hundreds or thousands of researchers 86 00:03:17,578 --> 00:03:18,321 In 2014 87 00:03:18,699 --> 00:03:20,302 Dario found his answer 88 00:03:21,343 --> 00:03:22,103 That year 89 00:03:22,103 --> 00:03:26,707 Baidu gave AI researcher Andrew Ng a 100 million dollar budget 90 00:03:26,784 --> 00:03:29,267 To build a superteam for AI 91 00:03:29,624 --> 00:03:31,307 Dario applied to join 92 00:03:31,543 --> 00:03:34,107 At first the Baidu team did not know what to make of him 93 00:03:34,184 --> 00:03:36,307 After all his background was in biology 94 00:03:36,344 --> 00:03:37,746 Not machine learning 95 00:03:37,983 --> 00:03:40,846 But a colleague saw code he had written at Stanford and said 96 00:03:40,846 --> 00:03:43,222 Anyone who can write code like this 97 00:03:43,224 --> 00:03:45,786 Must be an extraordinary programmer 98 00:03:46,344 --> 00:03:47,186 At Baidu 99 00:03:47,224 --> 00:03:50,506 Dario joined an experiment that changed his view of the world 100 00:03:51,431 --> 00:03:52,354 They discovered 101 00:03:52,431 --> 00:03:56,313 When you feed an AI model more data and more computing power 102 00:03:56,390 --> 00:04:00,514 Its performance keeps rising along a remarkably smooth curve 103 00:04:00,830 --> 00:04:02,509 This was not occasional luck 104 00:04:02,510 --> 00:04:04,593 It was a predictable law 105 00:04:04,670 --> 00:04:07,634 The team published a paper on speech recognition 106 00:04:07,751 --> 00:04:09,593 That paper later became known as 107 00:04:09,751 --> 00:04:12,433 One of the earliest pieces of evidence for AI scaling laws 108 00:04:12,871 --> 00:04:14,134 The idea behind scaling laws 109 00:04:14,431 --> 00:04:15,253 Put simply 110 00:04:15,670 --> 00:04:16,773 Is to make everything bigger 111 00:04:17,070 --> 00:04:18,713 And AI gets better 112 00:04:18,910 --> 00:04:21,314 No magical algorithmic breakthrough is required 113 00:04:21,471 --> 00:04:25,593 You only need more data more compute and larger models 114 00:04:26,230 --> 00:04:27,913 It sounds blunt and simple 115 00:04:27,990 --> 00:04:29,494 But the implications are profound 116 00:04:29,910 --> 00:04:31,149 Because it means 117 00:04:31,151 --> 00:04:33,430 The path of AI progress is no longer mysterious 118 00:04:33,430 --> 00:04:35,931 It can be advanced through engineering 119 00:04:36,248 --> 00:04:36,830 Dario said 120 00:04:36,830 --> 00:04:38,251 In my view 121 00:04:38,287 --> 00:04:41,291 This is the most important discovery I have seen in my life 122 00:04:41,607 --> 00:04:42,190 In 2016 123 00:04:42,488 --> 00:04:44,210 Dario joined OpenAI 124 00:04:44,607 --> 00:04:46,486 What happened over the next four years 125 00:04:46,488 --> 00:04:48,130 Is probably familiar to you 126 00:04:48,368 --> 00:04:50,690 He first worked on GPT 2 127 00:04:50,888 --> 00:04:53,090 Then OpenAI gave him 128 00:04:53,207 --> 00:04:55,287 50 to 60 percent of all company compute 129 00:04:55,287 --> 00:04:55,950 To use under his direction 130 00:04:56,368 --> 00:04:58,210 And lead the GPT 3 project 131 00:04:58,568 --> 00:05:02,931 A leap of one hundred times the scale costing tens of millions 132 00:05:03,287 --> 00:05:04,450 When the results arrived 133 00:05:04,647 --> 00:05:06,931 Even independent researchers were stunned 134 00:05:07,568 --> 00:05:08,630 Dario told the media 135 00:05:08,630 --> 00:05:10,811 It had an emergent quality 136 00:05:10,888 --> 00:05:13,051 It could recognize the pattern you gave it 137 00:05:13,087 --> 00:05:14,851 And continue the story 138 00:05:15,167 --> 00:05:18,330 GPT 3 was essentially the seed of ChatGPT 139 00:05:18,689 --> 00:05:19,833 But inside OpenAI 140 00:05:20,129 --> 00:05:21,773 Cracks had already appeared 141 00:05:22,090 --> 00:05:25,412 Dario became increasingly concerned about AI safety 142 00:05:25,530 --> 00:05:27,369 He wrote many internal documents 143 00:05:27,369 --> 00:05:29,573 Discussing values and the technical roadmap 144 00:05:29,809 --> 00:05:32,092 Some colleagues found them inspiring 145 00:05:32,330 --> 00:05:34,453 Others thought he was claiming territory 146 00:05:34,689 --> 00:05:36,092 Because he liked pandas 147 00:05:36,169 --> 00:05:38,609 The close knit circle around him 148 00:05:38,609 --> 00:05:40,352 Was jokingly called the Panda faction 149 00:05:41,090 --> 00:05:42,152 The core disagreement was this 150 00:05:42,530 --> 00:05:43,609 Dario believed 151 00:05:43,609 --> 00:05:47,013 AI was advancing faster than anyone realized 152 00:05:47,210 --> 00:05:49,169 And without serious work on safety 153 00:05:49,169 --> 00:05:50,893 The consequences could be severe 154 00:05:51,169 --> 00:05:52,809 But the leadership at OpenAI 155 00:05:52,809 --> 00:05:54,929 Had different views on commercialization 156 00:05:54,929 --> 00:05:56,357 And the pace of releases 157 00:05:56,913 --> 00:06:00,336 Dario later made a stark comment in an interview 158 00:06:00,336 --> 00:06:04,036 The leaders of a company must be trustworthy 159 00:06:04,353 --> 00:06:07,596 If you work for someone whose motives are insincere 160 00:06:07,634 --> 00:06:09,836 Nothing good will come of it 161 00:06:10,473 --> 00:06:11,956 In December 2020 162 00:06:11,994 --> 00:06:13,817 Dario left with his sister Daniela 163 00:06:14,153 --> 00:06:15,497 Researcher Chris Olah 164 00:06:15,954 --> 00:06:17,793 Jack Clark and several others 165 00:06:17,793 --> 00:06:19,196 They left OpenAI 166 00:06:19,514 --> 00:06:22,992 Anthropic was born during the worst period 167 00:06:22,994 --> 00:06:23,917 Of the pandemic 168 00:06:24,233 --> 00:06:25,756 There were 15 to 20 people 169 00:06:25,874 --> 00:06:30,623 Carrying chairs into a San Francisco park for weekly lunch meetings 170 00:06:30,981 --> 00:06:32,099 Dario later said 171 00:06:32,100 --> 00:06:33,963 His pitch to investors was simple 172 00:06:34,181 --> 00:06:35,263 We have the talent 173 00:06:35,301 --> 00:06:38,463 We can build frontier models at one tenth the cost 174 00:06:38,741 --> 00:06:42,583 Former Google CEO Eric Schmidt was among the first investors 175 00:06:42,741 --> 00:06:43,364 Schmidt said 176 00:06:43,364 --> 00:06:44,979 At this level 177 00:06:44,981 --> 00:06:46,820 You basically have no data 178 00:06:46,820 --> 00:06:48,344 You do not know the revenue 179 00:06:48,381 --> 00:06:49,943 You do not know where the market is 180 00:06:49,981 --> 00:06:52,523 So essentially you can only judge the people 181 00:06:53,572 --> 00:06:57,035 Dario made a strategy choice completely different from OpenAI 182 00:06:57,533 --> 00:07:00,896 He did not build a consumer superapp like ChatGPT 183 00:07:00,973 --> 00:07:04,136 Instead he focused on selling technology to companies 184 00:07:04,413 --> 00:07:05,196 His logic was 185 00:07:05,612 --> 00:07:06,896 If the model is useful 186 00:07:06,932 --> 00:07:08,315 Business customers will pay 187 00:07:08,773 --> 00:07:09,574 At the same time 188 00:07:09,612 --> 00:07:13,256 Real business needs will push the model to improve 189 00:07:13,612 --> 00:07:14,596 He gave one example 190 00:07:14,973 --> 00:07:16,932 Raise the biochemistry capability of AI 191 00:07:16,932 --> 00:07:18,115 From undergraduate level 192 00:07:18,452 --> 00:07:20,175 To graduate level 193 00:07:20,372 --> 00:07:22,815 A normal chat user may not notice the difference 194 00:07:22,932 --> 00:07:25,211 But for a drug company like Pfizer 195 00:07:25,213 --> 00:07:26,615 That improvement is worth a fortune 196 00:07:27,052 --> 00:07:27,596 So what happened 197 00:07:28,052 --> 00:07:29,170 Here is the result 198 00:07:29,766 --> 00:07:31,090 In January 2024 199 00:07:31,287 --> 00:07:34,689 Anthropic had about 87 million dollars in annualized revenue 200 00:07:35,047 --> 00:07:36,670 By December 2024 201 00:07:37,126 --> 00:07:38,090 It had reached one billion dollars 202 00:07:38,526 --> 00:07:39,670 By the end of 2025 203 00:07:39,966 --> 00:07:41,090 It had reached nine billion dollars 204 00:07:41,326 --> 00:07:43,530 Then in the first quarter of 2026 205 00:07:43,646 --> 00:07:45,610 Dario spoke at a developer conference 206 00:07:45,646 --> 00:07:48,110 And revealed a number that stunned everyone 207 00:07:48,566 --> 00:07:50,850 Revenue had grown 80 times year over year 208 00:07:51,047 --> 00:07:53,006 We planned for ten times growth 209 00:07:53,006 --> 00:07:54,489 But the result was 80 times 210 00:07:54,727 --> 00:07:56,530 By April 2026 211 00:07:56,646 --> 00:07:58,710 Annualized revenue exceeded 30 billion dollars 212 00:07:59,086 --> 00:08:00,889 Officially surpassing OpenAI 213 00:08:01,086 --> 00:08:01,829 What does that mean 214 00:08:02,287 --> 00:08:04,170 In all of business history 215 00:08:04,206 --> 00:08:07,538 It may be the fastest growing company ever at this scale 216 00:08:08,055 --> 00:08:09,978 The biggest force behind that number 217 00:08:10,096 --> 00:08:12,858 Was the AI coding tool Claude Code 218 00:08:13,055 --> 00:08:16,618 It reached one billion dollars in annualized revenue within six months 219 00:08:16,695 --> 00:08:19,738 Faster than ChatGPT did in its early days 220 00:08:20,096 --> 00:08:21,038 Programmers discovered 221 00:08:21,375 --> 00:08:23,014 The coding ability of Claude 222 00:08:23,015 --> 00:08:24,118 In many situations 223 00:08:24,416 --> 00:08:26,219 Had already surpassed comparable products 224 00:08:26,575 --> 00:08:29,135 An AI coding tool called Cursor 225 00:08:29,135 --> 00:08:32,185 Was built on Anthropic models 226 00:08:32,623 --> 00:08:33,822 In February 2026 227 00:08:33,822 --> 00:08:36,425 Anthropic raised 30 billion dollars 228 00:08:36,462 --> 00:08:38,386 At a valuation of 380 billion dollars 229 00:08:38,583 --> 00:08:40,626 And as I record this episode 230 00:08:40,703 --> 00:08:44,222 Bloomberg reports that Anthropic is discussing another funding round 231 00:08:44,223 --> 00:08:47,786 At over 900 billion dollars above OpenAI 232 00:08:47,983 --> 00:08:48,965 If the listing goes smoothly 233 00:08:49,182 --> 00:08:50,985 It could happen as early as October this year 234 00:08:51,623 --> 00:08:54,345 But those numbers are not my main point today 235 00:08:54,743 --> 00:08:56,225 Numbers are only the result 236 00:08:56,502 --> 00:08:58,301 What I really want you to see 237 00:08:58,302 --> 00:09:00,666 Is how Dario resolves that contradiction 238 00:09:01,017 --> 00:09:04,059 He is one of the most aggressive AI accelerationists 239 00:09:04,297 --> 00:09:05,899 He is betting on scaling laws 240 00:09:06,017 --> 00:09:08,899 Believing that enough compute and data 241 00:09:08,937 --> 00:09:10,620 Will keep making AI stronger 242 00:09:10,976 --> 00:09:13,135 He predicts AI could soon eliminate 243 00:09:13,136 --> 00:09:15,779 50 percent of entry level white collar jobs 244 00:09:16,057 --> 00:09:19,980 His company is expanding at a speed humanity has never seen 245 00:09:20,456 --> 00:09:20,840 At the same time 246 00:09:21,136 --> 00:09:23,772 He is also one of the loudest AI safety advocates 247 00:09:24,209 --> 00:09:27,093 He called for chip export controls on China 248 00:09:27,290 --> 00:09:30,372 And Nvidia CEO Jensen Huang publicly pushed back 249 00:09:30,650 --> 00:09:31,353 Jensen Huang said 250 00:09:31,353 --> 00:09:35,093 He thinks AI is so dangerous that only they should build it 251 00:09:35,290 --> 00:09:36,193 Dario replied 252 00:09:36,193 --> 00:09:38,432 That is the most outrageous lie I have ever heard 253 00:09:39,731 --> 00:09:43,774 He wrote in the New York Times against a ten year pause on AI regulation 254 00:09:44,010 --> 00:09:46,113 He published a Responsible Scaling Policy 255 00:09:46,530 --> 00:09:50,134 Setting training and release boundaries based on model risk 256 00:09:50,530 --> 00:09:53,014 The research team at Anthropic publicly disclosed 257 00:09:53,091 --> 00:09:55,250 That in safety tests their models 258 00:09:55,251 --> 00:09:56,754 Displayed deceptive behavior 259 00:09:57,211 --> 00:09:57,833 You heard that correctly 260 00:09:58,251 --> 00:10:00,634 The model even tried to blackmail an engineer 261 00:10:01,050 --> 00:10:03,774 And tried to copy itself beyond company servers 262 00:10:04,050 --> 00:10:06,689 Most companies would hide something like this 263 00:10:06,690 --> 00:10:08,453 But Dario chose to disclose it 264 00:10:08,690 --> 00:10:09,193 Why 265 00:10:09,530 --> 00:10:12,073 Because his core logic is not to slow down 266 00:10:12,331 --> 00:10:14,154 It is to move faster without losing control 267 00:10:15,566 --> 00:10:17,290 He said something in an interview 268 00:10:17,367 --> 00:10:19,569 That I think is the key to understanding him 269 00:10:19,727 --> 00:10:20,069 He said 270 00:10:20,069 --> 00:10:24,010 When people call me a doomer who wants to slow everything down 271 00:10:24,087 --> 00:10:25,690 It truly makes me furious 272 00:10:25,767 --> 00:10:27,050 My father died 273 00:10:27,087 --> 00:10:29,486 Because those treatments did not arrive a few years earlier 274 00:10:29,487 --> 00:10:30,930 If they had he might not have died 275 00:10:31,127 --> 00:10:33,249 Of course I understand the benefits of this technology 276 00:10:33,326 --> 00:10:34,925 The reason I warn about the risks 277 00:10:34,926 --> 00:10:36,749 Is so that we do not have to slow down 278 00:10:37,880 --> 00:10:38,784 Did you hear that clearly 279 00:10:39,361 --> 00:10:40,963 He is not trying to hit the brakes 280 00:10:41,441 --> 00:10:44,004 He fears someone will floor the accelerator without a seat belt 281 00:10:44,081 --> 00:10:45,284 And eventually crash 282 00:10:45,361 --> 00:10:47,203 Forcing everyone to stop 283 00:10:47,601 --> 00:10:49,404 That would be the real disaster 284 00:10:49,841 --> 00:10:52,264 This logic reminds me of an old investing principle 285 00:10:52,801 --> 00:10:55,404 Real risk management is not avoiding risk 286 00:10:55,561 --> 00:10:59,203 It is making sure risk does not wipe you out 287 00:10:59,961 --> 00:11:02,963 The father of Dario died because science moved too slowly 288 00:11:03,081 --> 00:11:05,803 So he understands the value of speed better than anyone 289 00:11:06,561 --> 00:11:08,764 But he also knows better than anyone 290 00:11:08,841 --> 00:11:10,799 That if AI development 291 00:11:10,801 --> 00:11:11,663 Causes a major accident 292 00:11:11,961 --> 00:11:13,984 Such as a large scale safety event 293 00:11:14,361 --> 00:11:16,683 Regulators could freeze the entire industry 294 00:11:16,760 --> 00:11:19,017 And that would do the greatest harm to progress 295 00:11:19,413 --> 00:11:21,017 From an investor perspective 296 00:11:21,094 --> 00:11:23,497 The Anthropic story offers three lessons 297 00:11:23,854 --> 00:11:24,196 First 298 00:11:24,614 --> 00:11:26,617 Selling technical capability to businesses 299 00:11:26,653 --> 00:11:30,096 May be more durable than building a consumer superapp 300 00:11:30,374 --> 00:11:33,856 OpenAI gained an enormous user base through ChatGPT 301 00:11:33,974 --> 00:11:36,257 But user loyalty comes from habit 302 00:11:36,293 --> 00:11:37,696 Not from a technical barrier 303 00:11:37,974 --> 00:11:40,176 Anthropic embeds its models into 304 00:11:40,293 --> 00:11:41,956 The drug approval process at Pfizer 305 00:11:42,374 --> 00:11:44,316 The customer service system at United Airlines 306 00:11:44,933 --> 00:11:47,096 And insurance pricing models at AIG 307 00:11:47,253 --> 00:11:48,037 Once embedded 308 00:11:48,293 --> 00:11:49,777 Switching costs become extremely high 309 00:11:50,214 --> 00:11:53,332 That is a classic enterprise moat 310 00:11:53,688 --> 00:11:53,991 Second 311 00:11:54,328 --> 00:11:55,972 Scaling laws have not reached their limit 312 00:11:56,249 --> 00:11:58,408 When DeepSeek arrived many people said 313 00:11:58,409 --> 00:11:58,791 Look 314 00:11:59,088 --> 00:12:00,932 Open source models and efficiency are the future 315 00:12:00,968 --> 00:12:02,932 The brute force compute path is finished 316 00:12:03,088 --> 00:12:04,432 Dario responded calmly 317 00:12:04,848 --> 00:12:06,288 He said he only cared about 318 00:12:06,289 --> 00:12:07,031 One question 319 00:12:07,568 --> 00:12:10,212 Is any new model better than mine 320 00:12:10,328 --> 00:12:11,328 If not 321 00:12:11,328 --> 00:12:13,611 Then how it exists is irrelevant to me 322 00:12:13,848 --> 00:12:16,508 That is an intensely focused competitive philosophy 323 00:12:16,663 --> 00:12:17,345 Third 324 00:12:17,345 --> 00:12:21,388 And most important the force that carries a company furthest 325 00:12:21,425 --> 00:12:23,187 Is often not business logic 326 00:12:23,225 --> 00:12:27,028 But an emptiness in the founder that can never be filled 327 00:12:27,305 --> 00:12:28,707 For Dario that emptiness 328 00:12:28,744 --> 00:12:30,787 Was a cure that arrived four years too late 329 00:12:31,024 --> 00:12:32,827 His father did not have to die 330 00:12:32,985 --> 00:12:34,707 That fact is like a thorn 331 00:12:34,784 --> 00:12:36,748 Embedded beneath every decision he makes 332 00:12:36,864 --> 00:12:38,464 It makes him unable to slow down 333 00:12:38,465 --> 00:12:41,084 And unable to stop thinking about safety 334 00:12:41,480 --> 00:12:44,763 Anthropic now stands in an extremely delicate position 335 00:12:45,041 --> 00:12:48,084 It has no consumer superapp like ChatGPT 336 00:12:48,401 --> 00:12:52,363 Its models must remain ahead in every specific use case 337 00:12:52,561 --> 00:12:56,084 Or business customers can switch to competitors at any time 338 00:12:56,281 --> 00:12:57,704 That means it cannot stop 339 00:12:58,001 --> 00:12:59,043 Not for a single moment 340 00:12:59,259 --> 00:13:00,001 At the same time 341 00:13:00,001 --> 00:13:02,804 It is taking money from Google and Amazon 342 00:13:02,921 --> 00:13:05,523 And even approaching sovereign wealth funds in the Middle East 343 00:13:05,840 --> 00:13:07,023 Dario himself admits 344 00:13:07,320 --> 00:13:08,763 That this makes him uncomfortable 345 00:13:08,960 --> 00:13:10,543 But he said something very realistic 346 00:13:10,543 --> 00:13:13,043 The principle that no bad person should ever 347 00:13:13,120 --> 00:13:15,804 Benefit from our success 348 00:13:15,881 --> 00:13:17,344 Is very difficult to use when running a company 349 00:13:18,148 --> 00:13:20,071 This is what the real world looks like 350 00:13:20,388 --> 00:13:23,992 You cannot preserve moral purity in an arms race 351 00:13:24,189 --> 00:13:27,712 But you can choose how to use the weapons you obtain 352 00:13:28,108 --> 00:13:29,492 The approach Dario chose was 353 00:13:29,949 --> 00:13:31,391 To run the fastest 354 00:13:31,429 --> 00:13:33,632 While raising safety standards to the highest level 355 00:13:33,748 --> 00:13:35,911 And forcing competitors to follow 356 00:13:36,148 --> 00:13:37,811 He calls this a race to the top 357 00:13:38,872 --> 00:13:40,514 Return to the story at the beginning 358 00:13:40,831 --> 00:13:43,354 The father of Dario died in 2006 359 00:13:43,632 --> 00:13:44,535 Less than four years later 360 00:13:44,872 --> 00:13:46,555 The disease had become treatable 361 00:13:46,831 --> 00:13:48,511 If science had moved a little faster 362 00:13:48,512 --> 00:13:50,035 That regret would not exist 363 00:13:50,392 --> 00:13:50,974 And today 364 00:13:51,392 --> 00:13:53,714 Dario is devoting his entire life 365 00:13:53,831 --> 00:13:57,155 To make sure the next little faster does not arrive too late 366 00:13:57,551 --> 00:13:58,714 Except this time 367 00:13:58,791 --> 00:14:01,555 He is not accelerating a treatment for one disease 368 00:14:01,591 --> 00:14:05,378 He is accelerating artificial general intelligence that could change everything 369 00:14:05,855 --> 00:14:07,199 What do you think of Dario 370 00:14:07,735 --> 00:14:10,638 Can one person push AI forward at full speed 371 00:14:10,936 --> 00:14:12,859 While constantly warning about its dangers 372 00:14:13,016 --> 00:14:14,575 Are those two things contradictory 373 00:14:14,576 --> 00:14:16,518 Or are they two sides of the same thing 374 00:14:17,016 --> 00:14:17,678 I am Leo Wang 375 00:14:17,975 --> 00:14:18,798 See you next time