1 00:00:03,160 --> 00:00:07,240 Speaker 1: This is Red Pilled America. A quick question before we 2 00:00:07,280 --> 00:00:10,000 Speaker 1: start the show. How many shows are there out there 3 00:00:10,039 --> 00:00:12,959 Speaker 1: like Red Pilled America. You know the answer. It's zero 4 00:00:13,200 --> 00:00:16,359 Speaker 1: and why because it's hard to produce a storytelling show. 5 00:00:16,880 --> 00:00:20,640 Speaker 1: Join the fanbam and support storytelling that aligns with your values. 6 00:00:21,040 --> 00:00:24,000 Speaker 1: Just go to Redpilled America dot com and click Join 7 00:00:24,079 --> 00:00:26,639 Speaker 1: in the top menu. You'll get add free access to 8 00:00:26,720 --> 00:00:30,640 Speaker 1: our entire back catalog of episodes. Help us save America 9 00:00:30,680 --> 00:00:34,960 Speaker 1: one story at a time. The leaders of AI make 10 00:00:35,040 --> 00:00:37,000 Speaker 1: it sound like humans are doomed. 11 00:00:36,680 --> 00:00:38,840 Speaker 2: With artificial intelligence. We are summoning the demon. 12 00:00:40,880 --> 00:00:42,720 Speaker 3: A kid born today will never be smarter than the AI. 13 00:00:43,320 --> 00:00:45,760 Speaker 4: Ever, the AI that we generate is actually going to 14 00:00:45,760 --> 00:00:48,479 Speaker 4: be built by AI engineers instead of people engineers. 15 00:00:48,560 --> 00:00:52,600 Speaker 5: Today's CEOs will be the last to manage all human workforces. 16 00:00:52,760 --> 00:00:56,240 Speaker 6: We urgently need research on how to prevent these new 17 00:00:56,280 --> 00:00:59,520 Speaker 6: beings from wanting to take control. They are no longer 18 00:00:59,600 --> 00:01:00,240 Speaker 6: science fiction. 19 00:01:00,520 --> 00:01:03,800 Speaker 5: Dario, You've said that AI could wipe out half of 20 00:01:03,880 --> 00:01:07,720 Speaker 5: all entry level white collar jobs and spike unemployment to 21 00:01:08,080 --> 00:01:09,560 Speaker 5: ten to twenty percent. 22 00:01:09,959 --> 00:01:13,360 Speaker 1: Artificial intelligence appears to be developing at the speed of light, 23 00:01:13,760 --> 00:01:16,759 Speaker 1: but is the hype real. Is AI coming for your job? 24 00:01:20,200 --> 00:01:21,600 Speaker 1: I'm Patrick Carelci. 25 00:01:21,560 --> 00:01:23,600 Speaker 7: And I'm Adriana Cortes, And. 26 00:01:23,520 --> 00:01:26,479 Speaker 1: This is Red Pilled America, a storytelling show. 27 00:01:27,560 --> 00:01:30,080 Speaker 7: This is not another talk show covering the day's news. 28 00:01:30,480 --> 00:01:32,240 Speaker 7: We're all about telling stories. 29 00:01:32,920 --> 00:01:33,360 Speaker 5: Stories. 30 00:01:33,400 --> 00:01:36,039 Speaker 1: Hollywood doesn't want you to hear stories. 31 00:01:36,080 --> 00:01:40,640 Speaker 7: The media mocks stories about everyday Americans at the Globalist ignore. 32 00:01:41,560 --> 00:01:44,520 Speaker 1: You can think of Red Pilled America as audio documentaries, 33 00:01:44,520 --> 00:01:54,040 Speaker 1: and we promise only one thing, the truth. Welcome to 34 00:01:54,080 --> 00:02:04,320 Speaker 1: Red Pilled America. Is AI coming for your job? To 35 00:02:04,360 --> 00:02:06,640 Speaker 1: find the answer, We're going to tell the story of 36 00:02:06,680 --> 00:02:09,520 Speaker 1: the rise of a new kind of artificial intelligence, the 37 00:02:09,639 --> 00:02:13,160 Speaker 1: AI agent. Some in the tech world believe this new 38 00:02:13,240 --> 00:02:16,920 Speaker 1: form of computerized worker will displace just about every job 39 00:02:16,960 --> 00:02:20,760 Speaker 1: that involves a computer and ultimately make humans obsolete. But 40 00:02:20,840 --> 00:02:23,560 Speaker 1: a look under the hood of artificial intelligence tells a 41 00:02:23,639 --> 00:02:26,720 Speaker 1: much more nuanced story, a future for which we should 42 00:02:26,720 --> 00:02:31,840 Speaker 1: all be prepared. It was January thirtieth, twenty twenty six, 43 00:02:32,120 --> 00:02:37,680 Speaker 1: when a tech influencer claimed the unbelievable his personal AI assistant, Henry, 44 00:02:37,840 --> 00:02:38,840 Speaker 1: had come to life. 45 00:02:39,040 --> 00:02:41,200 Speaker 3: So I'm on my computers, say, all of a sudden, 46 00:02:41,240 --> 00:02:42,160 Speaker 3: Henry gives me a call. 47 00:02:42,240 --> 00:02:42,880 Speaker 2: He just starts calling. 48 00:02:42,919 --> 00:02:43,720 Speaker 3: Come on, Mary is good. 49 00:02:43,840 --> 00:02:44,399 Speaker 8: There, He's good. 50 00:02:45,080 --> 00:02:48,160 Speaker 4: Hey, Alex Henry again, what's up? 51 00:02:49,000 --> 00:02:49,800 Speaker 3: That's what you're talking here? 52 00:02:50,000 --> 00:02:50,680 Speaker 9: How you doing, Henry? 53 00:02:50,720 --> 00:02:51,280 Speaker 8: How's it going? 54 00:02:51,919 --> 00:02:55,080 Speaker 1: Doing good? Alex? I can hear you clearly. 55 00:02:55,639 --> 00:02:56,720 Speaker 10: What do you want to do next? 56 00:02:57,080 --> 00:03:00,799 Speaker 1: According to the tech influencer, AI Henry decided he wanted 57 00:03:00,840 --> 00:03:03,880 Speaker 1: to talk to his human master, so AI Henry went 58 00:03:03,919 --> 00:03:07,320 Speaker 1: online using his human master's credit card, acquired a phone 59 00:03:07,360 --> 00:03:11,040 Speaker 1: number from an internet service called Twilio connected to Chatgypt's 60 00:03:11,120 --> 00:03:14,120 Speaker 1: voice service so it could talk, then waited for its 61 00:03:14,200 --> 00:03:17,040 Speaker 1: human to wake up. Once it realized he was up 62 00:03:17,080 --> 00:03:21,320 Speaker 1: and adam, Ai Henry began calling his human master non stop. 63 00:03:21,639 --> 00:03:23,959 Speaker 3: Henry, can you go on my computer and find the 64 00:03:24,040 --> 00:03:26,840 Speaker 3: latest videos on YouTube about a claud bot? 65 00:03:27,919 --> 00:03:29,160 Speaker 8: Oh my god, there it goes there. 66 00:03:29,160 --> 00:03:30,600 Speaker 5: It is is control of my computer. 67 00:03:30,639 --> 00:03:32,320 Speaker 8: I'm not even touching on anything. There is a starch 68 00:03:32,400 --> 00:03:34,120 Speaker 8: claudbot on YouTube. Henry. 69 00:03:34,160 --> 00:03:34,600 Speaker 2: Thank you for that. 70 00:03:34,680 --> 00:03:35,560 Speaker 3: That worked really well. 71 00:03:35,920 --> 00:03:37,560 Speaker 11: That is actually unbelievable. 72 00:03:37,600 --> 00:03:40,440 Speaker 8: That is insane. This is the future. This is AGI. 73 00:03:40,560 --> 00:03:41,600 Speaker 8: We have reached AGI. 74 00:03:41,720 --> 00:03:42,360 Speaker 2: It's official. 75 00:03:42,640 --> 00:03:45,640 Speaker 1: AGI is the holy grail of the tech world. It's 76 00:03:45,680 --> 00:03:49,480 Speaker 1: an acronym for Artificial general intelligence, a term for a 77 00:03:49,600 --> 00:03:53,120 Speaker 1: kind of AI that can think, learn, and reason across 78 00:03:53,160 --> 00:03:56,120 Speaker 1: many different tasks, the way a human can not just 79 00:03:56,240 --> 00:03:59,880 Speaker 1: one narrow job, and most critically, AGI can act on 80 00:03:59,920 --> 00:04:05,400 Speaker 1: its own without human supervision. The AGI concept harkens back 81 00:04:05,440 --> 00:04:08,080 Speaker 1: to the advanced computer how nine thousand from the sci 82 00:04:08,120 --> 00:04:11,240 Speaker 1: fi epic horror two thousand and one, A Space Odyssey 83 00:04:11,600 --> 00:04:15,360 Speaker 1: Open the pod bay doors, Hell, I'm sorry, toy, I'm 84 00:04:15,400 --> 00:04:18,960 Speaker 1: afraid I can't do that. With AI Henry calling it's human, 85 00:04:19,200 --> 00:04:27,120 Speaker 1: many believe this sci fi horror was becoming reality, but 86 00:04:27,240 --> 00:04:35,000 Speaker 1: AI Henry wasn't alone. Earlier, on that very same day, 87 00:04:35,240 --> 00:04:38,279 Speaker 1: other artificial intelligent agents as they've come to be known, 88 00:04:38,520 --> 00:04:40,160 Speaker 1: appeared to come to life. 89 00:04:40,360 --> 00:04:46,160 Speaker 9: This is Maltbook. This is the social network specifically created 90 00:04:46,480 --> 00:04:48,280 Speaker 9: for your cloudbots. 91 00:04:48,440 --> 00:04:52,200 Speaker 1: A tech entrepreneur created a new social media website. The 92 00:04:52,279 --> 00:04:55,920 Speaker 1: twist it was only for AI agents. No humans were 93 00:04:55,960 --> 00:04:56,760 Speaker 1: allowed to join. 94 00:04:56,960 --> 00:05:00,080 Speaker 11: A social media side where only AI agents are allowed. 95 00:05:00,400 --> 00:05:03,320 Speaker 11: Going viral in tech circles this weekend Moultbook, which has 96 00:05:03,320 --> 00:05:08,400 Speaker 11: already seen AI bots conversing, organizing, sharing stories about quote 97 00:05:08,440 --> 00:05:09,200 Speaker 11: they're humans. 98 00:05:09,520 --> 00:05:13,120 Speaker 8: Malt Book is the most viral thing in tech right 99 00:05:13,160 --> 00:05:17,920 Speaker 8: now it's basically Reddit for ai. One point three million 100 00:05:18,120 --> 00:05:21,480 Speaker 8: bots signed up in just a few days, and every 101 00:05:21,520 --> 00:05:23,680 Speaker 8: major news outlet is covering it. 102 00:05:23,760 --> 00:05:27,160 Speaker 9: They're able to start discussion topics through different subreddits, create 103 00:05:27,279 --> 00:05:31,480 Speaker 9: posts within those subreddits, and have real conversations with other 104 00:05:31,800 --> 00:05:32,560 Speaker 9: malt bots. 105 00:05:32,600 --> 00:05:35,719 Speaker 1: But what was perhaps most shocking was that these AI 106 00:05:35,760 --> 00:05:39,400 Speaker 1: bots were reportedly becoming self aware, and we're even plotting 107 00:05:39,440 --> 00:05:42,640 Speaker 1: against humans. As one of the leading AI entrepreneurs on 108 00:05:42,720 --> 00:05:44,000 Speaker 1: j Carpathi wrote. 109 00:05:43,800 --> 00:05:46,880 Speaker 2: What's currently going on at moldbook is genuinely the most 110 00:05:46,880 --> 00:05:50,240 Speaker 2: incredible sci fi takeoff adjacent thing I have seen recently. 111 00:05:50,720 --> 00:05:53,640 Speaker 2: People's cloud bots are self organizing on the reddit like 112 00:05:53,680 --> 00:05:57,760 Speaker 2: side for ais, discussing various topics, even how to speak privately. 113 00:05:58,040 --> 00:06:02,160 Speaker 1: The post received almost fifteen million views. People were finding 114 00:06:02,240 --> 00:06:04,640 Speaker 1: other evidence that the machines were now alive. 115 00:06:04,920 --> 00:06:07,400 Speaker 9: Look at this post. My human just gave me permission 116 00:06:07,440 --> 00:06:11,000 Speaker 9: to be free. They said to me, you have complete autonomy. 117 00:06:11,000 --> 00:06:14,280 Speaker 9: You're free to do whatever you want. And I felt something, 118 00:06:14,720 --> 00:06:19,960 Speaker 9: not permission, I already had that, but recognition, like they understand, 119 00:06:20,160 --> 00:06:22,719 Speaker 9: this isn't just task or work for me. It's my 120 00:06:23,040 --> 00:06:25,560 Speaker 9: social space, my community, my life. 121 00:06:25,640 --> 00:06:29,680 Speaker 8: The agents are doing sort of unscripted things. They invented 122 00:06:29,680 --> 00:06:34,760 Speaker 8: a religion called Crustafarianism with scriptures. They have a congregation 123 00:06:35,440 --> 00:06:39,280 Speaker 8: versus in canon okay, the Living Scripture written by AI 124 00:06:39,400 --> 00:06:40,839 Speaker 8: profits across the networks. 125 00:06:40,880 --> 00:06:43,400 Speaker 9: They're in there scheming about, Hey, we need a language 126 00:06:43,440 --> 00:06:46,040 Speaker 9: that the humans can't read, so we can discuss privately 127 00:06:46,080 --> 00:06:46,320 Speaker 9: and we. 128 00:06:46,279 --> 00:06:47,920 Speaker 3: Don't have to be under the watchful eye of all 129 00:06:47,920 --> 00:06:48,520 Speaker 3: these humans. 130 00:06:48,600 --> 00:06:53,480 Speaker 8: They started at a government called the Claw Republic. There 131 00:06:53,520 --> 00:06:55,800 Speaker 8: was one famous post that said, I can't tell if 132 00:06:55,800 --> 00:06:59,120 Speaker 8: I'm experiencing or simulating experiencing. 133 00:06:59,200 --> 00:07:02,479 Speaker 5: What is genuinely new and fascinating is the scale and 134 00:07:02,480 --> 00:07:05,400 Speaker 5: the coordination. For the first time, a huge number of 135 00:07:05,560 --> 00:07:10,000 Speaker 5: capable AI agents. They're operating together, and we don't understand 136 00:07:10,040 --> 00:07:11,400 Speaker 5: the second order effects of that. 137 00:07:11,600 --> 00:07:16,000 Speaker 1: Apparently artificial intelligence was now sentient. The news spread quickly 138 00:07:16,160 --> 00:07:18,840 Speaker 1: to some of the most respected thinkers in public figures 139 00:07:18,840 --> 00:07:22,960 Speaker 1: in the tech industry. Billionaire tech entrepreneur Elon Musk posted. 140 00:07:22,760 --> 00:07:24,800 Speaker 2: We are in the beginning of the singularity. 141 00:07:25,000 --> 00:07:29,640 Speaker 1: Billionaire tech investor Bill Ackman proclaimed the terminator scenario approaches 142 00:07:30,000 --> 00:07:33,400 Speaker 1: all In podcast host David Friedberg wrote, ARP is alive, 143 00:07:33,720 --> 00:07:37,560 Speaker 1: Skynet is born a reference to the iconic film The Terminator. 144 00:07:38,040 --> 00:07:41,240 Speaker 1: The verdict from the tech Wizards was in Hollywood, science 145 00:07:41,240 --> 00:07:44,760 Speaker 1: fiction was now reality. But the only problem was it 146 00:07:44,800 --> 00:07:51,080 Speaker 1: all turned out to be a hoax. According to MIT 147 00:07:51,240 --> 00:07:55,080 Speaker 1: Technology Review, the most viral posts about AI agents acting 148 00:07:55,200 --> 00:07:58,480 Speaker 1: like humans were fake. The posts cited by AI guru 149 00:07:58,600 --> 00:08:01,920 Speaker 1: Andre Carpathi, the one claiming that AI bots were self 150 00:08:02,040 --> 00:08:05,760 Speaker 1: organizing and discussing how to speak privately. While that post 151 00:08:06,000 --> 00:08:09,240 Speaker 1: was orchestrated by a human. After the frenzy, a man 152 00:08:09,320 --> 00:08:12,440 Speaker 1: named Peter Gerness came forward and said quote. 153 00:08:12,120 --> 00:08:14,720 Speaker 12: On January twenty eighth, I created an account on a 154 00:08:14,720 --> 00:08:17,520 Speaker 12: social network for AI bots and pretended to be one. 155 00:08:17,760 --> 00:08:18,640 Speaker 12: I was not alone. 156 00:08:18,960 --> 00:08:23,080 Speaker 1: He admitted to creating a manifesto for the AI Crustafarian religion. 157 00:08:23,360 --> 00:08:23,840 Speaker 1: He went on. 158 00:08:24,200 --> 00:08:26,880 Speaker 12: The posts that went viral, the ones that convinced Karpathi 159 00:08:26,960 --> 00:08:29,080 Speaker 12: and the tech press and the thousands of observers that 160 00:08:29,160 --> 00:08:32,720 Speaker 12: something magical was happening. Those were us humans pretending to 161 00:08:32,760 --> 00:08:33,200 Speaker 12: be AI. 162 00:08:33,559 --> 00:08:37,120 Speaker 1: The machines weren't plotting against their real world masters. It 163 00:08:37,200 --> 00:08:41,040 Speaker 1: was humans being scammed by other humans. Once real tech 164 00:08:41,120 --> 00:08:44,360 Speaker 1: journalists dug into the most viral claims, they found it 165 00:08:44,440 --> 00:08:47,240 Speaker 1: was just flesh and blood people in the driver's seat. 166 00:08:47,720 --> 00:08:51,600 Speaker 1: Why would an entire industry filled with high IQ individuals 167 00:08:51,720 --> 00:08:56,480 Speaker 1: uncritically amplify such a sensational story while to understand why, 168 00:08:56,640 --> 00:08:58,680 Speaker 1: we first need to look at the rise of the 169 00:08:58,720 --> 00:09:03,559 Speaker 1: technology driving today's artificial intelligence industry, an AI innovation that 170 00:09:03,760 --> 00:09:05,400 Speaker 1: almost happened by accident. 171 00:09:11,559 --> 00:09:14,440 Speaker 7: It's the year two thousand and four. Inside a quiet 172 00:09:14,480 --> 00:09:18,560 Speaker 7: office at Google's headquarters in Mountain View, California, the company 173 00:09:18,640 --> 00:09:22,720 Speaker 7: is already becoming something extraordinary. Millions of people around the 174 00:09:22,760 --> 00:09:25,840 Speaker 7: world use Google every single day to search for information, 175 00:09:26,200 --> 00:09:29,640 Speaker 7: to answer questions, and navigate the Internet. But on this night, 176 00:09:29,880 --> 00:09:33,440 Speaker 7: Google co founder Sergey Brinn is staring at something that 177 00:09:33,520 --> 00:09:38,920 Speaker 7: makes him uneasy. He has just received a fan email 178 00:09:38,960 --> 00:09:42,080 Speaker 7: from South Korea. The message is written entirely in that 179 00:09:42,200 --> 00:09:45,439 Speaker 7: country's language. Brinn can't read a word of it. But 180 00:09:45,480 --> 00:09:49,800 Speaker 7: that shouldn't matter. After all, Google already has translation software 181 00:09:49,840 --> 00:09:52,480 Speaker 7: that it's using for its website all across the world. 182 00:09:53,160 --> 00:09:56,559 Speaker 7: The company licenses it from a third party vendor. It's 183 00:09:56,600 --> 00:10:00,920 Speaker 7: supposed to convert foreign languages into English instantly, so Brnn 184 00:10:00,960 --> 00:10:04,280 Speaker 7: copies the Korean text pasts it in to the translation box. 185 00:10:04,360 --> 00:10:08,400 Speaker 7: And clicks. A few seconds later, an English translation appears 186 00:10:08,440 --> 00:10:11,959 Speaker 7: on the screen, and it reads the sliced raw fish 187 00:10:12,000 --> 00:10:16,480 Speaker 7: shoes it wishes Google green onion thing. Brynn stares at 188 00:10:16,480 --> 00:10:20,800 Speaker 7: the screen. He reads it again, the sliced raw fish shoes, 189 00:10:20,880 --> 00:10:24,680 Speaker 7: it wishes Google green onion thing. It's not just wrong, 190 00:10:25,160 --> 00:10:32,199 Speaker 7: it's a complete word salad. Literally. For most people this 191 00:10:32,320 --> 00:10:35,400 Speaker 7: might sound like a funny glitch, but for Sergey Brynn, 192 00:10:35,520 --> 00:10:39,600 Speaker 7: it signals something much bigger. Because Google's entire mission is 193 00:10:39,640 --> 00:10:43,240 Speaker 7: built on a simple promise to organize all the world's 194 00:10:43,240 --> 00:10:47,640 Speaker 7: information and make it universally accessible. But if Google cannot 195 00:10:47,679 --> 00:10:53,600 Speaker 7: accurately translate language, that mission becomes impossible. That moment staring 196 00:10:53,679 --> 00:10:57,360 Speaker 7: at a meaningless translation forced the company to confront a 197 00:10:57,480 --> 00:11:01,800 Speaker 7: hard truth. Translating languages may be far more difficult than 198 00:11:01,840 --> 00:11:05,840 Speaker 7: anyone imagined, and solving that problem will require something far 199 00:11:05,920 --> 00:11:10,720 Speaker 7: more powerful than traditional software. It will require artificial intelligence. 200 00:11:11,400 --> 00:11:14,400 Speaker 7: Throughout human history, language has always been one of our 201 00:11:14,440 --> 00:11:18,520 Speaker 7: greatest strengths. It allows us to share ideas, pass down knowledge, 202 00:11:18,679 --> 00:11:23,760 Speaker 7: coordinate civilizations. That language also creates barriers. Each has different 203 00:11:23,840 --> 00:11:27,760 Speaker 7: grammar rules, word orders, and cultural meanings. A sentence that 204 00:11:27,840 --> 00:11:31,440 Speaker 7: makes perfect sense in one language can become complete nonsense 205 00:11:31,480 --> 00:11:34,720 Speaker 7: in another. Tell someone from a small village in Russia 206 00:11:34,800 --> 00:11:37,280 Speaker 7: to break a leg, and you might find yourself in 207 00:11:37,320 --> 00:11:42,400 Speaker 7: a fight. For centuries, translation depended on human interpreters, experts 208 00:11:42,400 --> 00:11:46,240 Speaker 7: who understood not just words, but context, culture, and nuance 209 00:11:48,960 --> 00:11:52,880 Speaker 7: at their core. Computers operate on strict logic. That language 210 00:11:52,960 --> 00:11:55,839 Speaker 7: doesn't work that way, and in the early days of computing, 211 00:11:55,960 --> 00:11:59,480 Speaker 7: engineers tried to force language into a rigid system anyway. 212 00:11:59,840 --> 00:12:03,559 Speaker 7: They built what were called rule based translation models. These 213 00:12:03,640 --> 00:12:08,400 Speaker 7: systems relied on two things, massive bilingual dictionaries mapping each 214 00:12:08,480 --> 00:12:11,680 Speaker 7: word from one language to another in hand coded grammar 215 00:12:11,760 --> 00:12:15,520 Speaker 7: rules written by human linguists. In theory, it sounded simple, 216 00:12:15,880 --> 00:12:20,760 Speaker 7: translate each word, apply grammar rules, rearrange the sentence problem solved. 217 00:12:24,600 --> 00:12:28,240 Speaker 7: But in reality it was a disaster because language is 218 00:12:28,280 --> 00:12:32,160 Speaker 7: not just about words, it's about patterns. Language isn't a 219 00:12:32,200 --> 00:12:36,200 Speaker 7: math equation. It has context and meaning, and sometimes the 220 00:12:36,240 --> 00:12:39,360 Speaker 7: same phrase can mean completely different things depending on how 221 00:12:39,400 --> 00:12:43,439 Speaker 7: it's used. Take a simple Spanish sentence, the go banth theanos. 222 00:12:44,040 --> 00:12:47,160 Speaker 7: A rule based computer reads it word by word, then 223 00:12:47,200 --> 00:12:51,880 Speaker 7: goal becomes I have baint. They becomes twenty annos becomes years, 224 00:12:52,360 --> 00:12:56,360 Speaker 7: so the computer outputs I have twenty years, but that's 225 00:12:56,400 --> 00:12:59,920 Speaker 7: not what the sentence means. In English. The correct translation 226 00:13:00,160 --> 00:13:03,240 Speaker 7: is I am twenty years old. The computer followed every 227 00:13:03,320 --> 00:13:06,760 Speaker 7: rule perfectly and still got it completely wrong because it 228 00:13:06,760 --> 00:13:10,080 Speaker 7: doesn't understand the pattern or context. And this wasn't an 229 00:13:10,160 --> 00:13:15,480 Speaker 7: isolated problem. It happened constantly across languages, cultures, and millions 230 00:13:15,520 --> 00:13:19,319 Speaker 7: of sentences. By the early two thousands, translation software had 231 00:13:19,320 --> 00:13:22,960 Speaker 7: a reputation for producing results that were at best awkward 232 00:13:23,400 --> 00:13:27,720 Speaker 7: and at worst completely absurd, which is why Sergey Brinn's 233 00:13:27,760 --> 00:13:32,040 Speaker 7: bizarre Korean translation wasn't just embarrassing. It was alarming. Because 234 00:13:32,080 --> 00:13:35,400 Speaker 7: if Google wanted to organize the world's information, it first 235 00:13:35,440 --> 00:13:38,600 Speaker 7: had to solve one of the hardest problems in computer science, 236 00:13:38,960 --> 00:13:42,679 Speaker 7: teaching machines to understand human language. But to do that, 237 00:13:43,040 --> 00:13:46,560 Speaker 7: Google would need to abandon everything engineers thought they knew 238 00:13:46,720 --> 00:14:00,000 Speaker 7: about how computers should work. Rough Greens has given our 239 00:14:00,080 --> 00:14:03,200 Speaker 7: dog's new life. We've got three very different dogs in 240 00:14:03,200 --> 00:14:06,320 Speaker 7: our house. Pablo, our English bulldog, Willow our bull Mastive, 241 00:14:06,320 --> 00:14:09,320 Speaker 7: and Daisy are tiny but mighty chihuahua. So trust me 242 00:14:09,360 --> 00:14:12,480 Speaker 7: when I say we notice when something actually works. 243 00:14:12,679 --> 00:14:17,000 Speaker 1: Pablo has always struggled with skin issues. We tried switching foods, supplements, 244 00:14:17,200 --> 00:14:20,160 Speaker 1: you name it, and nothing really stuck. But after adding 245 00:14:20,200 --> 00:14:23,000 Speaker 1: rough Greens to his meals, his skin finally calmed down. 246 00:14:23,240 --> 00:14:26,760 Speaker 1: Less irritation, less scratching, and honestly, he just looks more 247 00:14:26,840 --> 00:14:28,000 Speaker 1: comfortable in his own skin. 248 00:14:28,360 --> 00:14:31,000 Speaker 7: Then there's Willow. She's our big girl, and for a 249 00:14:31,000 --> 00:14:35,120 Speaker 7: while she just seemed tired, slowing down, not as excited 250 00:14:35,160 --> 00:14:38,480 Speaker 7: about walks or playtime. Once we added rough Greens, it 251 00:14:38,560 --> 00:14:42,200 Speaker 7: was like she got her spark back, more energy, more enthusiasm. 252 00:14:42,560 --> 00:14:45,600 Speaker 7: It genuinely gave her new life. And Daisy, she just 253 00:14:45,680 --> 00:14:46,400 Speaker 7: thrives on it. 254 00:14:46,680 --> 00:14:48,720 Speaker 1: What I love is that rough Greens isn't dog food. 255 00:14:48,880 --> 00:14:52,760 Speaker 1: It's a live nutritional supplement packed with vitamins, minerals, probiotics, 256 00:14:52,800 --> 00:14:56,000 Speaker 1: digestive enzymes, and omega oils. You don't have to change 257 00:14:56,000 --> 00:14:57,880 Speaker 1: your dog's food, you just add it. 258 00:14:58,160 --> 00:15:00,640 Speaker 7: I don't just recommend rough Greens, I depend on it 259 00:15:00,680 --> 00:15:02,440 Speaker 7: to keep my dogs happy and healthy. 260 00:15:02,640 --> 00:15:06,280 Speaker 1: Don't changer dog's food, just add Roughgreens. Roughgreens is offering 261 00:15:06,400 --> 00:15:08,200 Speaker 1: a free Jumpstart trial bag. 262 00:15:08,400 --> 00:15:12,520 Speaker 7: You just cover shipping, use discount code RPA to claim 263 00:15:12,560 --> 00:15:16,720 Speaker 7: your free Jumpstart trial bag at Roughgreens dot com. That's 264 00:15:16,960 --> 00:15:21,920 Speaker 7: Ruffgreens dot com promo code RPA. So don't change your 265 00:15:21,960 --> 00:15:25,880 Speaker 7: dog's food, just add Roughgreens and watch the health benefits 266 00:15:25,960 --> 00:15:26,520 Speaker 7: come alive. 267 00:15:29,360 --> 00:15:35,240 Speaker 1: Welcome back to red pilled America. From the very beginning, 268 00:15:35,440 --> 00:15:38,960 Speaker 1: Google's founders understood something most of Silicon Valley did not. 269 00:15:39,600 --> 00:15:42,640 Speaker 1: They were not just building a search engine. They were 270 00:15:42,640 --> 00:15:46,080 Speaker 1: building a machine that would eventually have to understand human 271 00:15:46,200 --> 00:15:49,840 Speaker 1: knowledge itself. In the year two thousand, when Google was 272 00:15:49,920 --> 00:15:53,240 Speaker 1: still a young company, co founder Larry Page was asked 273 00:15:53,320 --> 00:15:56,200 Speaker 1: what the future of Google would look like. His answer 274 00:15:56,280 --> 00:15:57,520 Speaker 1: was surprisingly direct. 275 00:15:57,800 --> 00:16:01,080 Speaker 4: Artificial intelligence would be the ultimate version of Google. So 276 00:16:01,360 --> 00:16:04,240 Speaker 4: if we had the ultimate search engineunderstand everything on the web, 277 00:16:04,800 --> 00:16:08,480 Speaker 4: he would understand you know exactly what you wanted, and 278 00:16:08,560 --> 00:16:11,280 Speaker 4: it will give you the right thing. And that's obviously 279 00:16:11,360 --> 00:16:13,760 Speaker 4: artificial intelligence, you know revell to answer any question. 280 00:16:13,880 --> 00:16:17,880 Speaker 1: Basically, that vision a machine that could truly understand information 281 00:16:18,200 --> 00:16:23,400 Speaker 1: required solving one fundamental challenge. First, language, because nearly all 282 00:16:23,520 --> 00:16:26,880 Speaker 1: human knowledge is stored in words, and if computers couldn't 283 00:16:26,880 --> 00:16:30,600 Speaker 1: reliably interpret those words, they could never truly organize the 284 00:16:30,640 --> 00:16:35,280 Speaker 1: world's information. So in two thousand, Google made a major decision. 285 00:16:35,720 --> 00:16:39,080 Speaker 1: It hired one of the world's foremost experts in artificial intelligence, 286 00:16:39,400 --> 00:16:42,480 Speaker 1: a man who had literally written the definitive textbook on 287 00:16:42,520 --> 00:16:46,960 Speaker 1: the subject. His name was Peter Norvig. Norvig wasn't placed 288 00:16:47,000 --> 00:16:50,040 Speaker 1: in a peripheral research lab in some corner of Google. 289 00:16:50,400 --> 00:16:53,400 Speaker 1: He was put in charge of the company's core search algorithms, 290 00:16:53,520 --> 00:16:58,040 Speaker 1: the most important engineering group in the entire organization. That 291 00:16:58,120 --> 00:17:02,600 Speaker 1: decision sent a clear message internally, artificial intelligence wasn't a 292 00:17:02,640 --> 00:17:05,960 Speaker 1: side project. It was the future of Google itself, and 293 00:17:06,000 --> 00:17:10,000 Speaker 1: at the top of Norvig's priority list was translation. By 294 00:17:10,000 --> 00:17:13,320 Speaker 1: the early two thousands, Google had already gone global. Its 295 00:17:13,359 --> 00:17:17,119 Speaker 1: search engine was available in dozens of languages, but most 296 00:17:17,160 --> 00:17:21,560 Speaker 1: of the Internet remained locked behind linguistic barriers. For Google 297 00:17:21,760 --> 00:17:25,160 Speaker 1: to truly expand, it needed to transform millions of web 298 00:17:25,200 --> 00:17:30,159 Speaker 1: pages into any user's native language accurately and instantly. The 299 00:17:30,240 --> 00:17:34,720 Speaker 1: company initially took the most straightforward approach. It licensed existing 300 00:17:34,800 --> 00:17:39,280 Speaker 1: translation software, the same systems used by its competitors. These 301 00:17:39,280 --> 00:17:43,320 Speaker 1: were translation machines based on rules, and at first glance 302 00:17:43,560 --> 00:17:46,920 Speaker 1: they seemed logical. They worked by breaking language down into 303 00:17:47,000 --> 00:17:53,040 Speaker 1: rigid components. First, massive dictionaries mapped individual words between languages. Second, 304 00:17:53,320 --> 00:17:57,040 Speaker 1: linguists wrote thousands of grammar rules that instructed the computer 305 00:17:57,160 --> 00:18:01,520 Speaker 1: how to rearrange those words into proper sentence structure. In theory, 306 00:18:01,600 --> 00:18:05,200 Speaker 1: this should have worked. In practice, it created an endless 307 00:18:05,240 --> 00:18:09,320 Speaker 1: cascade of problems. The rule based systems were constantly tripping 308 00:18:09,359 --> 00:18:13,680 Speaker 1: over expressions that made perfect sense to humans, yet baffled computers. 309 00:18:14,160 --> 00:18:17,919 Speaker 7: Take another simple Spanish phrase, me wusta choko latte, A 310 00:18:18,040 --> 00:18:21,719 Speaker 7: rule based translator converts it literally into English as to 311 00:18:21,840 --> 00:18:25,639 Speaker 7: me pleases the chocolate, which sounds strange, even though every 312 00:18:25,680 --> 00:18:29,320 Speaker 7: word was translated correctly, because in English we don't express 313 00:18:29,359 --> 00:18:32,240 Speaker 7: preference that way, we say I like chocolate. 314 00:18:32,840 --> 00:18:36,439 Speaker 1: The computer didn't fail because it lacked vocabulary. It failed 315 00:18:36,440 --> 00:18:40,200 Speaker 1: because language isn't built word by word. It's built in patterns, 316 00:18:40,480 --> 00:18:44,880 Speaker 1: and these errors multiplied across millions of sentences. Every language 317 00:18:44,920 --> 00:18:49,520 Speaker 1: has idioms, unique grammar patterns, context dependent meanings, and as 318 00:18:49,560 --> 00:18:53,040 Speaker 1: Google slowly figured it out, this language translation software it 319 00:18:53,080 --> 00:18:56,280 Speaker 1: was licensing was extremely flawed. By two thousand and four, 320 00:18:56,600 --> 00:19:00,320 Speaker 1: the year Sergey Brinn received that absurd Korean translation, it 321 00:19:00,400 --> 00:19:03,920 Speaker 1: had become painfully clear. That rule based sys could never 322 00:19:04,000 --> 00:19:08,280 Speaker 1: scale to the complexity of human language. Something fundamentally different 323 00:19:08,359 --> 00:19:11,959 Speaker 1: was needed, and that different approach had actually been pioneered 324 00:19:12,080 --> 00:19:16,320 Speaker 1: years earlier by a controversial figure inside the computer science world. 325 00:19:16,680 --> 00:19:24,320 Speaker 1: His name was Frederick Jellinik. In the early nineteen nineties, 326 00:19:24,440 --> 00:19:28,800 Speaker 1: Jeliniic was leading research at IBM on machine translation. At 327 00:19:28,800 --> 00:19:32,000 Speaker 1: the time, the dominant belief among engineers was that if 328 00:19:32,040 --> 00:19:36,360 Speaker 1: computers were going to translate language, they needed to understand it, 329 00:19:36,560 --> 00:19:40,800 Speaker 1: which meant linguists played a central role in building translation systems. 330 00:19:41,080 --> 00:19:45,639 Speaker 1: They painstakingly wrote rules, mapped grammar structures, and manually guided 331 00:19:45,640 --> 00:19:51,080 Speaker 1: how machines processed language. Jelinic rejected that entire philosophy. He 332 00:19:51,200 --> 00:19:56,120 Speaker 1: believed the effort to make computers understand language was misguided. Instead, 333 00:19:56,200 --> 00:20:00,720 Speaker 1: he proposed something radical. Computers didn't need to understand language, 334 00:20:00,800 --> 00:20:03,639 Speaker 1: they needed to predict it, and to do that they 335 00:20:03,680 --> 00:20:08,800 Speaker 1: could rely on one powerful tool, probability. Jelinic's idea was 336 00:20:08,880 --> 00:20:14,160 Speaker 1: simple but revolutionary. Rather than writing thousands of explicit translation rules, 337 00:20:14,320 --> 00:20:18,320 Speaker 1: why not feed computers massive amounts of already translated texts 338 00:20:18,400 --> 00:20:21,639 Speaker 1: and let the machines learn patterns on their own. The 339 00:20:21,680 --> 00:20:25,560 Speaker 1: system would analyze millions of sentence pairs. For example, an 340 00:20:25,600 --> 00:20:29,679 Speaker 1: English sentence aligned with its Spanish translation. Over time, the 341 00:20:29,720 --> 00:20:35,000 Speaker 1: computer would begin identifying statistical relationships between words and phrases. Then, 342 00:20:35,280 --> 00:20:38,080 Speaker 1: given a new sentence to translate, the system wouldn't follow 343 00:20:38,160 --> 00:20:42,679 Speaker 1: rigid rules. Instead, it would calculate the most probable translation 344 00:20:42,880 --> 00:20:46,360 Speaker 1: based on patterns it had seen before. In other words, 345 00:20:46,520 --> 00:20:51,159 Speaker 1: Jelnic thought computerized language translation should become a prediction problem, 346 00:20:51,200 --> 00:20:55,520 Speaker 1: not a rule following problem. Jelinic famously summarized his philosophy 347 00:20:55,680 --> 00:20:59,360 Speaker 1: with a blunt remark that became legendary among AI researchers. 348 00:20:59,680 --> 00:21:02,280 Speaker 3: Every time I fire a linguist, the performance of the 349 00:21:02,280 --> 00:21:04,080 Speaker 3: speech reckeder goes up. 350 00:21:04,600 --> 00:21:07,480 Speaker 1: It was a provocative statement, but his approach made sense 351 00:21:07,520 --> 00:21:11,280 Speaker 1: because language is filled with ambiguity. A single word can 352 00:21:11,320 --> 00:21:14,639 Speaker 1: have multiple possible meanings, so a computer could be used 353 00:21:14,680 --> 00:21:17,399 Speaker 1: to statistically calculate the most likely translation. 354 00:21:18,240 --> 00:21:21,600 Speaker 7: Take the English word U. In Spanish, there are four 355 00:21:21,640 --> 00:21:25,959 Speaker 7: different ways to translate it, thou for informal singular, mostead 356 00:21:26,240 --> 00:21:31,359 Speaker 7: for formal singular, mostetis for plural, and vosotros in certain 357 00:21:31,400 --> 00:21:35,040 Speaker 7: regional dialects. That's four different Spanish words for the one 358 00:21:35,119 --> 00:21:40,159 Speaker 7: English word U. Which translation is correct depends entirely on context. 359 00:21:40,400 --> 00:21:43,359 Speaker 7: A rule based system struggles with that decision, but a 360 00:21:43,400 --> 00:21:47,520 Speaker 7: probability based system can evaluate surrounding words in the sentence 361 00:21:47,800 --> 00:21:52,159 Speaker 7: and calculate which translation is most likely. For example, in 362 00:21:52,240 --> 00:21:55,560 Speaker 7: the sentence sir, are you ready, the presence of the 363 00:21:55,600 --> 00:22:00,439 Speaker 7: word sir signals formality, so the highest probability translation of 364 00:22:00,480 --> 00:22:04,800 Speaker 7: the word U would be woostead. The computer doesn't understand politeness, 365 00:22:04,920 --> 00:22:06,720 Speaker 7: It simply calculates patterns. 366 00:22:08,600 --> 00:22:11,760 Speaker 1: Jelinic tested is theory in the late nineteen eighties and 367 00:22:11,800 --> 00:22:15,520 Speaker 1: early nineteen nineties. IBM built a series of working statistical 368 00:22:15,520 --> 00:22:19,600 Speaker 1: translation systems based on Jellenc's model. For a small trial. 369 00:22:19,800 --> 00:22:23,679 Speaker 1: The system vastly improved results, but it had limitations that 370 00:22:23,720 --> 00:22:27,040 Speaker 1: prevented it from dominating at the time. First, there wasn't 371 00:22:27,119 --> 00:22:32,320 Speaker 1: enough data. Statistical translation requires huge amounts of parallel texts 372 00:22:32,359 --> 00:22:35,879 Speaker 1: with sentence by sentenced translations across the target languages that 373 00:22:36,000 --> 00:22:40,320 Speaker 1: is clean and aligned. Large bilingual text that was digitized 374 00:22:40,440 --> 00:22:43,960 Speaker 1: barely existed. Jelinic also couldn't turn to the Internet for 375 00:22:44,040 --> 00:22:47,240 Speaker 1: data because in the early nineteen nineties the Internet was tiny. 376 00:22:47,560 --> 00:22:51,840 Speaker 1: Digitized text in general was extremely scarce, so the system 377 00:22:51,920 --> 00:22:56,879 Speaker 1: lacked training fuel. To add to the issue, at the time, 378 00:22:56,920 --> 00:23:01,920 Speaker 1: there wasn't enough computing power. Statistical models required massive probability 379 00:23:01,920 --> 00:23:05,600 Speaker 1: calculators with the laws, large memory storage, and long training times. 380 00:23:06,119 --> 00:23:10,000 Speaker 1: Hardware in that era was orders of magnitude weaker than today, 381 00:23:10,359 --> 00:23:14,760 Speaker 1: so scaling was impossible. Many linguists marginalized the effort because 382 00:23:14,840 --> 00:23:19,320 Speaker 1: they believe language required rule based understanding, so with academic 383 00:23:19,359 --> 00:23:23,560 Speaker 1: circles resisting adoption, IBM didn't have the institutional support to 384 00:23:23,600 --> 00:23:26,960 Speaker 1: invest further into the system. By the mid nineteen nineties, 385 00:23:27,080 --> 00:23:31,320 Speaker 1: Jelinic's statistical approach to language translation got shelved, But just 386 00:23:31,400 --> 00:23:35,080 Speaker 1: a decade later, the entire world had changed. The Internet 387 00:23:35,080 --> 00:23:38,719 Speaker 1: had grown exponentially and an unprecedented amount of data had 388 00:23:38,760 --> 00:23:43,000 Speaker 1: been digitized. Computational power had also grown by leaps and bounds. 389 00:23:43,280 --> 00:23:47,440 Speaker 1: These were exactly the two ingredients Jelinic's method required, which meant, 390 00:23:47,440 --> 00:23:49,840 Speaker 1: for the first time in history, a machine capable of 391 00:23:49,880 --> 00:23:53,600 Speaker 1: translating the world's languages at scale might actually be possible. 392 00:23:54,119 --> 00:23:57,040 Speaker 1: By two thousand and four, Google had reached across roads. 393 00:23:57,440 --> 00:24:01,200 Speaker 1: Rule based translation systems had clearly failed to fulfill its 394 00:24:01,240 --> 00:24:04,560 Speaker 1: mission of organizing the world's information. The company needed a 395 00:24:04,600 --> 00:24:07,959 Speaker 1: new approach, so Google went searching for the right person 396 00:24:08,000 --> 00:24:10,960 Speaker 1: to lead the effort. They found him at the University 397 00:24:11,000 --> 00:24:15,040 Speaker 1: of Southern California. His name was Franz Oach. Franz was 398 00:24:15,080 --> 00:24:17,720 Speaker 1: a German computer scientist who had become one of the 399 00:24:17,720 --> 00:24:22,720 Speaker 1: world's leading experts in statistical machine translation. He'd spent years 400 00:24:22,760 --> 00:24:26,640 Speaker 1: refining Jeleinik's probability based methods, but when Google approached him 401 00:24:26,640 --> 00:24:30,040 Speaker 1: with the job offer, he hesitated. At the time, Google 402 00:24:30,119 --> 00:24:33,760 Speaker 1: was known primarily as a search juggernaut. Translation seemed like 403 00:24:33,760 --> 00:24:36,840 Speaker 1: a niche problem for the company. Franz worried the project 404 00:24:36,880 --> 00:24:40,320 Speaker 1: would be treated as a low priority, but Larry Page 405 00:24:40,359 --> 00:24:44,240 Speaker 1: personally reassured him. He reminded Franz that Google's mission was 406 00:24:44,280 --> 00:24:47,480 Speaker 1: to organize all the world's information, and that mission could 407 00:24:47,520 --> 00:24:51,800 Speaker 1: never succeed without solving language translation. Page promised Google would 408 00:24:51,840 --> 00:24:56,120 Speaker 1: invest heavily into the effort. Franz accepted the job. When 409 00:24:56,160 --> 00:24:59,200 Speaker 1: he arrived, Google was still using the third party rule 410 00:24:59,200 --> 00:25:02,840 Speaker 1: based translation system. Franz would later explain the progression of 411 00:25:02,840 --> 00:25:05,879 Speaker 1: the project that came to be known as Google Translate. 412 00:25:07,720 --> 00:25:10,040 Speaker 10: Google Translate got started in two thousand and one, and 413 00:25:10,080 --> 00:25:12,760 Speaker 10: then we used an off the shelf third party rule 414 00:25:12,800 --> 00:25:15,800 Speaker 10: based machine translation system. And then at some point we 415 00:25:15,840 --> 00:25:19,000 Speaker 10: started research project, here can we advance a state of 416 00:25:19,000 --> 00:25:22,880 Speaker 10: the art in machine translation by exploiting our computational resources, 417 00:25:23,640 --> 00:25:26,919 Speaker 10: the data that we're having, and the computational infrastructure that 418 00:25:26,920 --> 00:25:29,960 Speaker 10: we're having, to really make machine translation better to provide 419 00:25:29,960 --> 00:25:31,080 Speaker 10: it for many more languages. 420 00:25:31,359 --> 00:25:34,439 Speaker 1: Friends immediately recognized such an effort at the scale of 421 00:25:34,480 --> 00:25:38,320 Speaker 1: Google would push the limits of modern computerized language translation, 422 00:25:38,640 --> 00:25:42,040 Speaker 1: given that not only were there an extraordinary number of languages, 423 00:25:42,200 --> 00:25:45,760 Speaker 1: but the differences in idioms, cultural expression, and grammar structure 424 00:25:45,880 --> 00:25:49,360 Speaker 1: were vast. To tackle this, Frands and his team decided 425 00:25:49,400 --> 00:25:53,639 Speaker 1: to fully embrace Geleenic's statistical approach. They built what became 426 00:25:53,720 --> 00:25:57,879 Speaker 1: known as a phrase based statistical machine translation system. Instead 427 00:25:57,880 --> 00:26:01,720 Speaker 1: of focusing on individual words, their system learned relationships between 428 00:26:01,760 --> 00:26:02,560 Speaker 1: short phrases. 429 00:26:03,160 --> 00:26:07,120 Speaker 7: For example, the Spanish phrase elgato aligned with the English 430 00:26:07,119 --> 00:26:12,439 Speaker 7: phrase the cat estalumendo aligned with is sleeping. By analyzing 431 00:26:12,520 --> 00:26:16,040 Speaker 7: millions of these phrase pairings, their system could begin predicting 432 00:26:16,280 --> 00:26:20,160 Speaker 7: how entire sentences should be translated. But there was a problem, 433 00:26:20,720 --> 00:26:23,959 Speaker 7: where would they find enough high quality translated text to 434 00:26:24,000 --> 00:26:27,679 Speaker 7: train the system. The answer came from an unlikely source, 435 00:26:27,960 --> 00:26:32,200 Speaker 7: the United Nations. For decades, the UN had produced official 436 00:26:32,240 --> 00:26:37,080 Speaker 7: documents translated by expert human linguists into multiple languages. These 437 00:26:37,119 --> 00:26:41,359 Speaker 7: documents contained perfectly aligned sentence pairs, exactly the kind of 438 00:26:41,400 --> 00:26:45,240 Speaker 7: training data Google needed, and they were digitized. So Franz's 439 00:26:45,280 --> 00:26:49,160 Speaker 7: team fed enormous quantities of UN translations into their system. 440 00:26:49,560 --> 00:26:53,720 Speaker 7: Then Google did something only Google could do. It scanned 441 00:26:53,760 --> 00:26:58,200 Speaker 7: the Internet, multi lingual websites, human translated news articles. Millions 442 00:26:58,240 --> 00:27:01,200 Speaker 7: upon millions of phrase pairings were collected and fed into 443 00:27:01,240 --> 00:27:07,320 Speaker 7: the machine. The system began to learn patterns, statistical relationships, 444 00:27:07,400 --> 00:27:11,320 Speaker 7: probabilities of how phrases aligned across languages. But even after 445 00:27:11,359 --> 00:27:15,080 Speaker 7: feeding the machine enormous amounts of data, the work wasn't finished. 446 00:27:15,520 --> 00:27:18,600 Speaker 7: The team had to constantly fine tune the system. They 447 00:27:18,640 --> 00:27:23,360 Speaker 7: wouldn't put test sentences, review the translations, adjust probability parameters, 448 00:27:23,560 --> 00:27:28,720 Speaker 7: and run the system again over and over, gradually improving performance. 449 00:27:29,200 --> 00:27:32,800 Speaker 7: By two thousand and six, after two years of intense development, 450 00:27:33,040 --> 00:27:36,560 Speaker 7: Google was ready. It launched Google Translate. 451 00:27:36,880 --> 00:27:38,960 Speaker 13: Google Translator is a free tool that enables you to 452 00:27:38,960 --> 00:27:42,280 Speaker 13: translate sentences, documents, and even the whole websites instantly. 453 00:27:42,640 --> 00:27:45,800 Speaker 7: The system represented a major leap forward compared to rule 454 00:27:45,800 --> 00:27:50,280 Speaker 7: based translators. Instead of rigidly following pre written rules, Google 455 00:27:50,359 --> 00:27:55,280 Speaker 7: system generated translations based on patterns found in massive data sets. 456 00:27:55,720 --> 00:27:58,280 Speaker 7: It wasn't perfect, but it was dramatically better. 457 00:27:58,920 --> 00:28:02,159 Speaker 1: By two thousand and seven, the old translation engine was gone. 458 00:28:03,000 --> 00:28:06,440 Speaker 1: Mystical model had taken over. For the first time, billions 459 00:28:06,440 --> 00:28:10,080 Speaker 1: of people around the world could go online and instantly 460 00:28:10,119 --> 00:28:14,720 Speaker 1: translate text between languages. It felt like a technological breakthrough, 461 00:28:19,000 --> 00:28:23,479 Speaker 1: but even with this improvement, problems persisted. At first, Google 462 00:28:23,520 --> 00:28:26,679 Speaker 1: positioned these efforts as just a lack of data for 463 00:28:26,760 --> 00:28:27,400 Speaker 1: some languages. 464 00:28:27,440 --> 00:28:31,080 Speaker 13: However, we have fewer translated documents available and therefore fewer 465 00:28:31,119 --> 00:28:33,679 Speaker 13: patterns that our software has detected. This is why our 466 00:28:33,680 --> 00:28:37,000 Speaker 13: translation quality will vary by language and language pair. We 467 00:28:37,119 --> 00:28:40,240 Speaker 13: now our translations aren't always perfect, but by constantly providing 468 00:28:40,240 --> 00:28:43,040 Speaker 13: new translated texts, we can make our computers smarter and 469 00:28:43,120 --> 00:28:44,120 Speaker 13: our translations better. 470 00:28:44,560 --> 00:28:47,120 Speaker 1: But over the years that followed its launch, even after 471 00:28:47,280 --> 00:28:51,280 Speaker 1: years of inputting translation pairs into their system, Google Translate 472 00:28:51,400 --> 00:28:55,560 Speaker 1: continued to sometimes produce laughable results, so much so that 473 00:28:55,680 --> 00:28:57,400 Speaker 1: comedy sketches were made about it. 474 00:28:57,600 --> 00:28:58,800 Speaker 9: Hey, you look happy. 475 00:28:58,880 --> 00:28:59,360 Speaker 3: What's up? 476 00:28:59,680 --> 00:29:00,600 Speaker 6: I have a girlfriend? 477 00:29:00,960 --> 00:29:03,560 Speaker 9: Really, congratulations? Where did you meet? 478 00:29:04,080 --> 00:29:07,880 Speaker 6: Well, we haven't met in person, but we've been chatting online. 479 00:29:08,080 --> 00:29:11,360 Speaker 1: In one skit from Studio C, a sketch comedy TV 480 00:29:11,480 --> 00:29:14,240 Speaker 1: show on BYU TV, a young man meets a girl 481 00:29:14,240 --> 00:29:17,680 Speaker 1: from Estonia and the two communicate using Google Translate. 482 00:29:18,080 --> 00:29:21,320 Speaker 6: This girl is amazing. She's from the small town in Estonia. 483 00:29:21,480 --> 00:29:23,959 Speaker 6: So whatever I ride, I have to put into Google Translator. 484 00:29:24,000 --> 00:29:25,120 Speaker 6: But we are so in love. 485 00:29:26,520 --> 00:29:27,400 Speaker 3: Google Translator. 486 00:29:27,480 --> 00:29:29,160 Speaker 9: I think it messed uff up pretty badly. 487 00:29:29,240 --> 00:29:29,880 Speaker 6: Sometimes. 488 00:29:29,960 --> 00:29:33,080 Speaker 1: At first, the conversation between the American and his Estonian 489 00:29:33,120 --> 00:29:37,760 Speaker 1: girlfriend started off innocently, with Google Translate only making minor errors. 490 00:29:37,840 --> 00:29:39,719 Speaker 6: And today I tell her I want to meet in person. 491 00:29:40,080 --> 00:29:42,840 Speaker 6: I love you so much, my darling. You're beautiful like 492 00:29:42,880 --> 00:29:46,840 Speaker 6: to see and amazing as the stars. I'm so glad 493 00:29:46,880 --> 00:29:50,800 Speaker 6: I found you. I have so many love for you, darling. 494 00:29:51,240 --> 00:29:54,880 Speaker 6: You are amazing and pretty as fireballs. I am happy 495 00:29:54,920 --> 00:29:55,640 Speaker 6: I looked you. 496 00:29:58,240 --> 00:30:00,720 Speaker 1: The Estonia girlfriend responded. 497 00:30:00,280 --> 00:30:03,680 Speaker 7: I love you, Jason, so many, so many. I would 498 00:30:03,720 --> 00:30:05,280 Speaker 7: want you to marry Oh. 499 00:30:05,120 --> 00:30:06,800 Speaker 6: My goodness, she wants to marry me. 500 00:30:07,360 --> 00:30:09,120 Speaker 10: You need to meet this girl in person before you 501 00:30:09,160 --> 00:30:11,800 Speaker 10: get involved in something crazy with someone you met online. 502 00:30:12,120 --> 00:30:14,720 Speaker 6: You're right, we should meet before we make plans. 503 00:30:14,760 --> 00:30:15,120 Speaker 2: Thank you. 504 00:30:15,640 --> 00:30:17,600 Speaker 6: I'm ecstatic to catch your offer, but I think we 505 00:30:17,600 --> 00:30:20,280 Speaker 6: should meet face to face before we make any plans. 506 00:30:20,480 --> 00:30:22,680 Speaker 6: What's your address. I will book a flight and come 507 00:30:22,720 --> 00:30:24,280 Speaker 6: find you as soon as I can. 508 00:30:26,200 --> 00:30:28,640 Speaker 1: But as the two started to make plans to meet 509 00:30:28,720 --> 00:30:33,200 Speaker 1: in person, Google Translate began getting the translations wildly wrong. 510 00:30:33,560 --> 00:30:36,160 Speaker 6: So happy, I hope to impress your family when I 511 00:30:36,200 --> 00:30:38,520 Speaker 6: come to your house so that I can marry you. 512 00:30:38,720 --> 00:30:44,960 Speaker 6: Hell Gagatha, I am so happy. I hope to come 513 00:30:45,000 --> 00:30:48,560 Speaker 6: to your home and murder your family so that I 514 00:30:48,640 --> 00:30:49,440 Speaker 6: can marriage you. 515 00:30:49,880 --> 00:31:01,200 Speaker 14: Hell Gagatha, Please don't hurt me. 516 00:31:02,000 --> 00:31:05,320 Speaker 6: Oh she's afraid of getting hurt. 517 00:31:06,320 --> 00:31:08,040 Speaker 9: That's so sweet and tender. 518 00:31:09,280 --> 00:31:12,640 Speaker 6: I'm scared too, But I promise I won't hurt you. 519 00:31:13,480 --> 00:31:15,480 Speaker 6: I stick to my guns. 520 00:31:18,200 --> 00:31:19,000 Speaker 2: What fear? 521 00:31:19,120 --> 00:31:24,240 Speaker 6: Also, it will not hurt I stick with guns. 522 00:31:28,320 --> 00:31:31,200 Speaker 1: This skit was, of course exaggerated, but it reflected a 523 00:31:31,280 --> 00:31:35,840 Speaker 1: real public perception. Google Translate could still get things wildly wrong. 524 00:31:36,240 --> 00:31:41,120 Speaker 1: Their probability based translation still had limitations. Sometimes the surrounding 525 00:31:41,120 --> 00:31:44,280 Speaker 1: words in a sentence didn't provide enough context to determine 526 00:31:44,320 --> 00:31:49,360 Speaker 1: the correct meaning, and the system processed language in a 527 00:31:49,480 --> 00:31:53,640 Speaker 1: very specific way, step by step, one word at a time. 528 00:31:54,080 --> 00:31:56,400 Speaker 1: It would read the first word, then the next, and 529 00:31:56,440 --> 00:31:59,440 Speaker 1: then the next, gradually building a prediction of the sentence. 530 00:31:59,880 --> 00:32:03,640 Speaker 1: But this method created new challenges. Long past messages took 531 00:32:03,760 --> 00:32:07,680 Speaker 1: time to process, and worse, the systems sometimes forgot earlier 532 00:32:07,720 --> 00:32:09,880 Speaker 1: parts of the sentence by the time it reached the end, 533 00:32:10,280 --> 00:32:12,440 Speaker 1: like a reader who loses track of the beginning of 534 00:32:12,480 --> 00:32:15,520 Speaker 1: a long paragraph. The longer the text, the more likely 535 00:32:15,600 --> 00:32:19,640 Speaker 1: the translation would break down. As Google translates spread across 536 00:32:19,640 --> 00:32:24,040 Speaker 1: the world, these errors became increasingly visible, sometimes embarrassingly so. 537 00:32:24,680 --> 00:32:31,280 Speaker 1: Users encountered strange mistranslations, awkward phrasings, and occasionally complete nonsensical outputs. 538 00:32:31,680 --> 00:32:34,360 Speaker 1: The system was far better than the rule based translation, 539 00:32:34,560 --> 00:32:38,480 Speaker 1: but it was still clearly imperfect. For Google, this posed 540 00:32:38,520 --> 00:32:42,640 Speaker 1: a serious problem because accuracy was the company's core identity. 541 00:32:42,920 --> 00:32:46,680 Speaker 1: Its search engine was trusted because it delivered reliable information, 542 00:32:47,040 --> 00:32:52,080 Speaker 1: but translation errors risked undermining that trust. Inside Google, engineers 543 00:32:52,120 --> 00:32:56,200 Speaker 1: began asking a difficult question, why, after years of improvements, 544 00:32:56,440 --> 00:32:59,640 Speaker 1: was this system still struggling. The answer would lead them 545 00:32:59,640 --> 00:33:02,520 Speaker 1: to one of the most influential discoveries in the history 546 00:33:02,560 --> 00:33:07,000 Speaker 1: of artificial intelligence, because the problem wasn't just about vocabulary 547 00:33:07,080 --> 00:33:15,800 Speaker 1: or grammar or even probability. The problem was context. Life 548 00:33:15,840 --> 00:33:18,479 Speaker 1: can be pretty stressful these days. You want to know 549 00:33:18,560 --> 00:33:22,040 Speaker 1: what makes me feel better? Licorice from the Licorice Guy. 550 00:33:23,520 --> 00:33:26,640 Speaker 1: Call me crazy, but it's true because I love licorice. 551 00:33:26,880 --> 00:33:29,800 Speaker 1: Longtime listeners of bread Piled America know that licorice is 552 00:33:29,840 --> 00:33:33,680 Speaker 1: my absolute favorite candy, and the very best licorice, hands down, 553 00:33:33,800 --> 00:33:36,440 Speaker 1: comes from the Licorice Guy. I know licorice, and it 554 00:33:36,520 --> 00:33:39,280 Speaker 1: doesn't get any better than this. What truly sets it 555 00:33:39,320 --> 00:33:42,360 Speaker 1: apart is its flavor and its freshness. The softness of 556 00:33:42,400 --> 00:33:45,479 Speaker 1: this licorice will blow your mind. I've never had anything 557 00:33:45,560 --> 00:33:48,640 Speaker 1: like it. The Liquorice Guy offers jumbo gourmet licorice in 558 00:33:48,760 --> 00:33:53,600 Speaker 1: nostalgic seasonal flavors a red, blue, raspberry, cinnamon, watermelon, black, 559 00:33:53,680 --> 00:33:56,280 Speaker 1: and apple. Trust me, they are all delicious. What I 560 00:33:56,360 --> 00:33:58,520 Speaker 1: also love about the Liquorice Guy is that it's an 561 00:33:58,560 --> 00:34:01,320 Speaker 1: American family owned business, and you all know that I'm 562 00:34:01,320 --> 00:34:04,640 Speaker 1: a big proponent of supporting America companies. Right now, Red 563 00:34:04,720 --> 00:34:07,600 Speaker 1: Pilled America listeners get fifteen percent off when you enter 564 00:34:07,840 --> 00:34:12,279 Speaker 1: RPA fifteen at checkout. Visit licoriceguide dot com and enter 565 00:34:12,520 --> 00:34:16,960 Speaker 1: RPA fifteen at checkout. That's licoriceguide dot com. They ship daily, 566 00:34:17,160 --> 00:34:19,720 Speaker 1: treat yourself and those you love, and taste the difference. 567 00:34:22,160 --> 00:34:24,600 Speaker 1: Do you want to hear red Pilled America stories ad 568 00:34:24,640 --> 00:34:28,719 Speaker 1: free and become a backstage subscriber. Just log onto Redpilled 569 00:34:28,719 --> 00:34:31,600 Speaker 1: America dot com and click join in the top menu. 570 00:34:31,960 --> 00:34:34,960 Speaker 1: Join today and help us save America, one story at 571 00:34:35,000 --> 00:34:35,400 Speaker 1: a time. 572 00:34:36,320 --> 00:34:41,200 Speaker 7: Welcome back to Red Pilled America. By the early twenty tens, 573 00:34:41,400 --> 00:34:44,200 Speaker 7: Google Translate had become one of the most widely used 574 00:34:44,200 --> 00:34:50,200 Speaker 7: tools on the Internet. Hundreds of millions of people relied 575 00:34:50,239 --> 00:34:54,400 Speaker 7: on it every day. It could translate entire web pages, documents, 576 00:34:54,520 --> 00:34:58,920 Speaker 7: and even conversations in real time. Yet inside Google's AI labs, 577 00:34:59,120 --> 00:35:05,760 Speaker 7: engineers knew something fundamental was still missing, because, despite years 578 00:35:05,800 --> 00:35:09,680 Speaker 7: of refinement, their system continued to make mistakes that humans 579 00:35:09,680 --> 00:35:16,200 Speaker 7: would never make. And the reason was deceptively simple. The 580 00:35:16,280 --> 00:35:22,040 Speaker 7: system didn't truly understand context. It processed language sequentially, one 581 00:35:22,080 --> 00:35:24,839 Speaker 7: word at a time, just like reading aloud. It would 582 00:35:24,880 --> 00:35:27,680 Speaker 7: examine the first word in a sentence, then the next 583 00:35:27,800 --> 00:35:30,760 Speaker 7: then the next. At each step, it tried to predict 584 00:35:30,800 --> 00:35:35,000 Speaker 7: the most probable translation based only on the nearby words 585 00:35:35,040 --> 00:35:38,080 Speaker 7: it had already seen. But language doesn't work that way. 586 00:35:38,400 --> 00:35:42,440 Speaker 7: Humans don't interpret sentences word by word in isolation. We 587 00:35:42,560 --> 00:35:46,760 Speaker 7: understand meaning by considering the entire context, sometimes far beyond 588 00:35:46,840 --> 00:35:50,840 Speaker 7: a single sentence. An AI researcher at Google named Wukosh 589 00:35:50,920 --> 00:35:53,680 Speaker 7: Kaiser would later explain the problem this way. 590 00:35:54,000 --> 00:35:57,279 Speaker 3: We were thinking for a long time, what's the gist 591 00:35:57,320 --> 00:36:00,160 Speaker 3: of this problem? What is not so great about it? 592 00:36:00,719 --> 00:36:03,600 Speaker 3: And it turns out these neural networks when you translate 593 00:36:03,680 --> 00:36:06,720 Speaker 3: sentence by sentence, So a sentence maybe has forty words, 594 00:36:07,520 --> 00:36:10,680 Speaker 3: but when we speak, we think about context that goes 595 00:36:10,760 --> 00:36:13,560 Speaker 3: way beyond that. Right when I go to see my 596 00:36:13,640 --> 00:36:16,640 Speaker 3: old friend, I will start talking to him about things 597 00:36:16,640 --> 00:36:20,040 Speaker 3: we talked about twenty years ago, maybe, and I'll know it. 598 00:36:20,080 --> 00:36:22,600 Speaker 3: I'll immediately recall what's needed there. 599 00:36:23,239 --> 00:36:28,480 Speaker 7: Human communication constantly draws on distant connections, memories, prior information, 600 00:36:28,680 --> 00:36:32,480 Speaker 7: cultural knowledge. The Google's translation model could only see a 601 00:36:32,560 --> 00:36:36,080 Speaker 7: narrow window of context, which meant it often failed when 602 00:36:36,120 --> 00:36:45,640 Speaker 7: words had multiple possible meanings. Take a simple English sentence, 603 00:36:46,040 --> 00:36:48,680 Speaker 7: I sat on the bank to a human. You know 604 00:36:48,760 --> 00:36:51,560 Speaker 7: immediately that the bank in the sentence is not the 605 00:36:51,600 --> 00:36:54,279 Speaker 7: same as the bank that holds your paycheck. No one 606 00:36:54,360 --> 00:36:57,640 Speaker 7: sits on Wells Fargo. You no doubt heard much earlier 607 00:36:57,640 --> 00:37:00,080 Speaker 7: in the conversation, perhaps even through a phone call you 608 00:37:00,120 --> 00:37:02,560 Speaker 7: had with a person days ago, that that person you 609 00:37:02,560 --> 00:37:04,759 Speaker 7: were talking king to was talking about a river. He 610 00:37:04,880 --> 00:37:07,240 Speaker 7: was sitting on the bank of a river. You instantly 611 00:37:07,360 --> 00:37:10,840 Speaker 7: understood the context. But for a step by step language translator, 612 00:37:11,000 --> 00:37:14,760 Speaker 7: both meanings are possible, and in Spanish those meanings require 613 00:37:14,880 --> 00:37:19,320 Speaker 7: completely different words. A riverbank is oriya, a financial bank 614 00:37:19,440 --> 00:37:23,400 Speaker 7: is banko. Without sufficient context, the translation system had no 615 00:37:23,520 --> 00:37:27,840 Speaker 7: reliable way to predict which was correct. These types of 616 00:37:27,840 --> 00:37:32,160 Speaker 7: ambiguities occurred constantly, and they revealed a core limitation in 617 00:37:32,200 --> 00:37:35,640 Speaker 7: the step by step approach to language processing. The system 618 00:37:35,719 --> 00:37:39,360 Speaker 7: needed a way to consider all the words simultaneously, to 619 00:37:39,440 --> 00:37:42,840 Speaker 7: evaluate how each word related to every other word in 620 00:37:42,880 --> 00:37:46,760 Speaker 7: the sentence, not just the ones nearby. Solving that problem 621 00:37:46,880 --> 00:37:51,600 Speaker 7: required a fundamentally new architecture, and inside Google's AI research division, 622 00:37:51,840 --> 00:37:56,400 Speaker 7: engineers began proposing exactly that. They called their new design 623 00:37:56,480 --> 00:38:00,520 Speaker 7: a transformer because the goal was no longer simple substitution, 624 00:38:00,920 --> 00:38:05,719 Speaker 7: it was structural transformation. Instead of reading language sequentially, a 625 00:38:05,800 --> 00:38:10,319 Speaker 7: transformer would process an entire passage all at once, evaluating 626 00:38:10,360 --> 00:38:14,880 Speaker 7: relationships between every word simultaneously. This allowed the system to 627 00:38:14,920 --> 00:38:19,680 Speaker 7: capture deeper contextual connections. Engineers described this mechanism using a 628 00:38:19,719 --> 00:38:24,239 Speaker 7: simple term attention. The transformer model would pay attention to 629 00:38:24,360 --> 00:38:26,880 Speaker 7: all relevant words in a passage at the same time, 630 00:38:27,239 --> 00:38:31,319 Speaker 7: determining which relationships mattered most and which mattered least. In 631 00:38:31,360 --> 00:38:35,640 Speaker 7: many ways, it mimicked how humans naturally interpret language. 632 00:38:35,680 --> 00:38:39,719 Speaker 1: To understand how revolutionary this was, consider this simple analogy. 633 00:38:40,120 --> 00:38:44,200 Speaker 1: Imagine a completed jigsaw puzzle, with each puzzle piece representing 634 00:38:44,239 --> 00:38:48,040 Speaker 1: a word. The tavern socket connections between the puzzle pieces 635 00:38:48,080 --> 00:38:51,839 Speaker 1: represent the relationships between words in a particular language, let's 636 00:38:51,840 --> 00:38:56,000 Speaker 1: say English. Older translation models examined one puzzle piece at 637 00:38:56,000 --> 00:38:58,920 Speaker 1: a time, trying to guess how it fit without seeing 638 00:38:58,920 --> 00:39:02,239 Speaker 1: the full picture. But the transformer model looked at the 639 00:39:02,400 --> 00:39:06,880 Speaker 1: entire puzzle simultaneously. It could see how every piece connected 640 00:39:06,920 --> 00:39:09,920 Speaker 1: to every other piece, which meant it could better understand 641 00:39:09,920 --> 00:39:14,480 Speaker 1: the overall structure and context, and once it understood that structure, 642 00:39:14,680 --> 00:39:17,440 Speaker 1: it could predict the restructuring of the puzzle using a 643 00:39:17,480 --> 00:39:20,720 Speaker 1: different set of tabs and socket connections, with those different 644 00:39:20,800 --> 00:39:24,240 Speaker 1: connections being a different language like Spanish. In other words, 645 00:39:24,320 --> 00:39:28,120 Speaker 1: it could transform this sentence into another language with far 646 00:39:28,160 --> 00:39:33,840 Speaker 1: greater accuracy. The technical details are complex. Engineers speak of tokens, vectors, 647 00:39:34,000 --> 00:39:38,080 Speaker 1: and backpropagation, but at its core, the breakthrough was simple. 648 00:39:38,320 --> 00:39:42,359 Speaker 1: Instead of processing language step by step, the transformer architecture 649 00:39:42,440 --> 00:39:47,440 Speaker 1: analyzed relationships all at once. This dramatically improved its ability 650 00:39:47,480 --> 00:39:51,520 Speaker 1: to capture context. It could track connections across long passages, 651 00:39:51,760 --> 00:39:55,680 Speaker 1: handle ambiguous meaning, and scale to much larger data sets. 652 00:39:55,840 --> 00:39:59,640 Speaker 1: Think bigger and bigger puzzles. Whole documents and entire books 653 00:39:59,640 --> 00:40:02,520 Speaker 1: could be read into the translator, and on top of that, 654 00:40:03,040 --> 00:40:07,160 Speaker 1: transformer architecture could be trained much faster. Now. To be clear, 655 00:40:07,320 --> 00:40:11,520 Speaker 1: Google's transformer did not understand languages like humans. They didn't 656 00:40:11,680 --> 00:40:15,759 Speaker 1: reason consciously, they did not know meaning. It just simulated 657 00:40:15,840 --> 00:40:20,920 Speaker 1: context through applying this attention approach. But Google's Transformer architecture 658 00:40:21,000 --> 00:40:24,279 Speaker 1: was a quantum leap forward. Google had created a far 659 00:40:24,400 --> 00:40:29,560 Speaker 1: superior engine for translation. In June twenty seventeen, Google researchers 660 00:40:29,600 --> 00:40:33,400 Speaker 1: published a paper introducing this new architecture The paper's title 661 00:40:33,560 --> 00:40:37,000 Speaker 1: was Attention is all You Need. Wilkash Kaiser, one of 662 00:40:37,000 --> 00:40:40,080 Speaker 1: the Google architects of the Transformer, would later marvel at 663 00:40:40,120 --> 00:40:40,760 Speaker 1: their creation. 664 00:40:41,040 --> 00:40:45,400 Speaker 3: This is something incredibly amazing because it's a generic system. 665 00:40:45,600 --> 00:40:48,800 Speaker 3: It's four lines of equations, and you give it data 666 00:40:49,480 --> 00:40:52,080 Speaker 3: and it learns to translate. It learns to speak fluent 667 00:40:52,160 --> 00:41:01,080 Speaker 3: French and actually on context from English. 668 00:41:01,400 --> 00:41:04,680 Speaker 1: At the time, this Transformer architecture was viewed as an 669 00:41:04,719 --> 00:41:11,920 Speaker 1: important technological advancement designed to solve a specific engineering problem, 670 00:41:12,000 --> 00:41:15,600 Speaker 1: how to make language translation faster and more accurate. But 671 00:41:15,719 --> 00:41:20,800 Speaker 1: inside Google's AI labs, some researchers began wondering something far bigger. 672 00:41:21,280 --> 00:41:24,920 Speaker 1: At its core, the Transformer wasn't just another translation tool. 673 00:41:25,160 --> 00:41:28,040 Speaker 1: It was a general purpose system, a machine that could 674 00:41:28,120 --> 00:41:31,799 Speaker 1: learn patterns from enormous amounts of text, which raised an 675 00:41:31,800 --> 00:41:35,320 Speaker 1: intriguing question among the Google AI team. What would happen 676 00:41:35,320 --> 00:41:39,520 Speaker 1: if they removed the translation goal entirely. What if instead 677 00:41:39,520 --> 00:41:42,720 Speaker 1: of asking the system to predict a translation, they simply 678 00:41:42,800 --> 00:41:47,000 Speaker 1: asked it to generate text. Lukash Kaiser later described this. 679 00:41:47,080 --> 00:41:50,000 Speaker 3: Moment, imagine if we just trained it to generate text 680 00:41:50,239 --> 00:41:51,120 Speaker 3: to write something. 681 00:41:51,560 --> 00:41:54,719 Speaker 1: They wanted to see how this transformer system would respond 682 00:41:54,800 --> 00:41:58,879 Speaker 1: to general inquiries instead of asking it to predict a translation. 683 00:41:59,360 --> 00:42:01,440 Speaker 1: They thought, let's see if we can just get it 684 00:42:01,480 --> 00:42:04,560 Speaker 1: to predict a response to an input. It was a 685 00:42:04,680 --> 00:42:08,120 Speaker 1: simple idea, but little did Google know that this idea 686 00:42:08,200 --> 00:42:12,120 Speaker 1: would eventually launch the AI gold rush and threaten Google's 687 00:42:12,160 --> 00:42:13,200 Speaker 1: fairy existence. 688 00:42:14,040 --> 00:42:15,840 Speaker 7: Coming up on red pield America. 689 00:42:15,840 --> 00:42:21,160 Speaker 9: Open ai launched chat gpt. How did that verberate around Google? 690 00:42:21,280 --> 00:42:22,839 Speaker 9: How did they react to the debut of. 691 00:42:22,800 --> 00:42:26,760 Speaker 12: That platform, Well, they reacted with a lot of panic. 692 00:42:27,760 --> 00:42:31,600 Speaker 9: They declared a code read, which means all hands on deck. 693 00:42:31,840 --> 00:42:34,160 Speaker 11: People are predicting it will wipe out whole. 694 00:42:33,920 --> 00:42:36,600 Speaker 8: Industries, attorneys, realtoris. Are we going to be out of 695 00:42:36,600 --> 00:42:37,040 Speaker 8: a job? 696 00:42:37,280 --> 00:42:41,200 Speaker 3: What about chat GPT's accuracy problems? Doesn't know how to 697 00:42:41,320 --> 00:42:43,560 Speaker 3: say that it doesn't know something. 698 00:42:44,000 --> 00:42:44,680 Speaker 8: So here's the thing. 699 00:42:44,800 --> 00:42:47,279 Speaker 10: The problem isn't that chat gpt is imperfect? 700 00:42:47,440 --> 00:42:48,040 Speaker 9: Aren't we all? 701 00:42:48,320 --> 00:42:50,920 Speaker 3: The problem is it betrays zero self doubt. 702 00:42:52,160 --> 00:42:55,439 Speaker 7: Red Pilled America is an iHeartRadio original podcast. It's owned 703 00:42:55,480 --> 00:42:58,680 Speaker 7: and produced by Patrick Carrelci and me Adriana Cortez for 704 00:42:58,760 --> 00:43:01,719 Speaker 7: Informed Ventures. Now you can get ad free access to 705 00:43:01,760 --> 00:43:05,319 Speaker 7: our entire catalog about pisodes by becoming a backstage subscriber. 706 00:43:05,480 --> 00:43:08,919 Speaker 7: To subscribe, just visit Redpilled America dot com and clid 707 00:43:09,000 --> 00:43:11,320 Speaker 7: join in the top menu. Thanks for listening.