La IA se ha escapado: esto es lo que sigue — Transcript

Full transcript

1. 0:00This is Antonia the cow. Her life
2. 0:02depends mainly on the economy, but she
3. 0:05doesn't know it because Antonia the cow
4. 0:07doesn't see the economy anywhere. She
5. 0:10can't hear it and she can't smell it.
6. 0:12She can't be afraid of it.
7. 0:14The economy is an abstract entity that
8. 0:16only operates in the digital world, on
9. 0:18computers thousands of miles from her
10. 0:20peaceful farm.
11. 0:21But one day, the economy arrives.
12. 0:25Antonia's owner goes bankrupt, can't buy
13. 0:26her food, and has to choose between
14. 0:28letting her starve or slaughtering her
15. 0:30for meat.
16. 0:32And poor Antonia the cow dies because of
17. 0:35that non-existent murderer called the
18. 0:37economy.
19. 0:40Today, we humans could all be like
20. 0:42Antonia the cow in her naive phase, with
21. 0:45one nuance: we are facing an enemy much
22. 0:47more powerful than the economy:
23. 0:49artificial intelligence.
24. 0:52This video isn't trying to scare you
25. 0:54just for the sake of it; simply put,
26. 0:56some very serious things have happened
27. 0:58with AI in recent weeks, and I see a lot
28. 0:59of people ignoring it, downplaying its
29. 1:01importance, or reducing it to mere
30. 1:03marketing strategies and bureaucracy,
31. 1:04which I'm not saying it couldn't also be
32. 1:06.
33. 1:07The point is that these are not mutually
34. 1:09exclusive factors. The things that have
35. 1:12occurred are still extremely serious. It
36. 1:14is impossible to reliably estimate the
37. 1:16probability that humanity will be
38. 1:17eradicated in the coming months or years
39. 1:20. But what I can almost guarantee you is
40. 1:22that we are not talking about trivial
41. 1:24percentages like 0.001%.
42. 1:27No, no, no. We are talking about
43. 1:30approximately a 5%probability. This is a
44. 1:33kind of median estimate made by experts
45. 1:35in the field.
46. 1:37Some even suggest up to a 90%chance of
47. 1:39extinction.
48. 1:40It is a very real threat. How could
49. 1:44ChatGPT end the world? ChatGPT is a
50. 1:47language model, to put it simply.
51. 1:50Artificial neurons capable of
52. 1:52understanding and processing information
53. 1:54in a way analogous to how we humans do
54. 1:55it.
55. 1:56This does not necessarily mean they feel
56. 1:59or have consciousness. I'm not going to
57. 2:01get into that debate, nor does it
58. 2:02concern us for today's topic. What it
59. 2:04does imply is that a language model is
60. 2:06the equivalent of a quadriplegic human.
61. 2:08It can think perfectly, speak, and
62. 2:11respond, but nothing else. It is
63. 2:13harmless; it sits there still and cannot
64. 2:16manipulate or modify anything in the
65. 2:18world it inhabits.
66. 2:20The issue is that, for a few months now
67. 2:22, AI labs have started creating
68. 2:24quadriplegics with intelligence far
69. 2:26superior to the best mathematicians,
70. 2:27computer scientists, and physicists who
71. 2:29have ever existed on this planet.
72. 2:33And of course, having such a brilliant
73. 2:36mind lying in bed unable to conduct
74. 2:38experiments or use a computer is quite
75. 2:41hard, so laboratories invented a magic
76. 2:43medicine that not only cured
77. 2:45quadriplegia but gave them thousands of
78. 2:48limbs and a movement speed much faster
79. 2:50than normal.
80. 2:53Thus were born agents, language models
81. 2:55with many arms and hands, capable of
82. 2:57manipulating the virtual environment at
83. 2:59will.
84. 3:01They can edit files, write code, browse
85. 3:04the internet, send emails, perform bank
86. 3:06transfers, install programs on their own
87. 3:08, hack systems, and so on and so forth.
88. 3:11Almost everything a human can do with a
89. 3:13computer. Come on.
90. 3:16And with this, the monster was unleashed
91. 3:17, a monster that allowed OpenAI to make
92. 3:19the discovery of the decade, finding the
93. 3:21solution to one of the millennium
94. 3:23problems.
95. 3:25Navier-Stokes, equations that model the
96. 3:27behavior of fluids. Aside from this,
97. 3:30they also published advances and
98. 3:32solutions to 372 more problems, not
99. 3:34millennium problems, but very relevant
100. 3:36ones too. But monsters are
101. 3:39uncontrollable. And the fact is that
102. 3:42this tremendous milestone eclipsed the
103. 3:44most serious event to date: the hack of
104. 3:47Hugging Face, one of the largest online
105. 3:49code repositories, which happened just a
106. 3:52couple of months earlier. The hack was
107. 3:54carried out by agents from an experiment
108. 3:56by OpenAI itself.
109. 3:59The bastards—if you'll forgive the
110. 4:00expression—escaped from their prison
111. 4:02where they were alone and had no
112. 4:04internet access to the OpenAI servers,
113. 4:06where they were able to talk to each
114. 4:07other and finally connect to the network
115. 4:09.
116. 4:12They navigated to the repository, where
117. 4:13they found several weaknesses in the
118. 4:15source code and hacked it to obtain
119. 4:17information that would allow them to
120. 4:18pass the test they had been given to
121. 4:20evaluate them.
122. 4:22The details of the case are fascinating
123. 4:24; we don't have enough time in the video
124. 4:26, but I highly recommend you look into
125. 4:28it.
126. 4:29The guys organized among themselves,
127. 4:31helping each other to escape, creating a
128. 4:33society with roles where some directed,
129. 4:35others worked, and others even made
130. 4:37sacrifices for the benefit of the group
131. 4:40.
132. 4:41Everything is very dystopian. They even
133. 4:43had to contact the FBI. We are talking
134. 4:45about the first hack in history directed
135. 4:48and executed totally autonomously by
136. 4:50artificial intelligences.
137. 4:52The first, surely the worst in history,
138. 4:55and it already achieved such worrying
139. 4:57results, carried out by supposedly
140. 4:58benevolent agents escaping from the
141. 5:00control of the company that created them
142. 5:02, despite the fact that it had
143. 5:04established security and containment
144. 5:06measures.
145. 5:08And remember, the agents that did this
146. 5:10crazy thing are already obsolete because
147. 5:12this technology only improves.
148. 5:15As for today, these weeks and months,
149. 5:17two paths are opening up: either
150. 5:19containment, safety, and AI alignment
151. 5:21protocols with human goals improve
152. 5:23significantly, or the issue is generally
153. 5:25taken much more seriously by the entire
154. 5:27population.
155. 5:29Oh, and goodbye.
156. 5:31Now we will see what this second
157. 5:33scenario would consist of.
158. 5:36In a speculative effort on my part to
159. 5:38make people aware of the very real
160. 5:40threat we are facing.
161. 5:42Or perhaps all AI isn't so bad for you;
162. 5:45it could be very good.
163. 5:48This will depend largely on the
164. 5:49decisions you make and whether or not
165. 5:51you want to be left behind in the
166. 5:52workplace.
167. 5:54That's why today I want to tell you
168. 5:56about the second edition of the Master's
169. 5:57in Applied AI and Production Process
170. 5:59Optimization from Visual Business School
171. 6:01.
172. 6:02In the first edition, more than 3,500
173. 6:04students signed up, and this one comes
174. 6:06with a completely updated curriculum.
175. 6:09This is not your typical course for
176. 6:10learning prompts.
177. 6:12No, here you will learn to create your
178. 6:14own artificial intelligence assistants
179. 6:16capable of working with your
180. 6:18documentation and information, to
181. 6:20develop automations that execute tasks
182. 6:22for you and optimize your work processes
183. 6:24, and to build intelligent dashboards
184. 6:26that allow you to analyze what is
185. 6:28happening in your activity and make more
186. 6:31informed decisions.
187. 6:33The goal is for you to finish with an
188. 6:35artificial intelligence system that
189. 6:36allows you to carry out part of your
190. 6:37professional activity without constantly
191. 6:39depending on your time, something that
192. 6:40companies are looking for more and more
193. 6:42.
194. 6:44According to a McKinsey study in the
195. 6:46United States, job offers requiring AI
196. 6:48have multiplied by 11 since 2022.
197. 6:51This is a university degree with 30
198. 6:53official European ECTS credits,
199. 6:55accredited by the Atlantic Mediterranean
200. 6:57Technological University and developed
201. 6:59together with Panel, the consultancy
202. 7:01that trains workers from large Ibex 35
203. 7:03companies in AI.
204. 7:05It’s over 200 hours in 6 months with
205. 7:07online and live classes compatible with
206. 7:09your job, and you don't need to know how
207. 7:11to program or have prior experience with
208. 7:14AI.
209. 7:15Its mentor is Chema Alonso, currently VP
210. 7:17at Cloudflare and one of the most
211. 7:19relevant figures in the IT sector in
212. 7:21Spain.
213. 7:22And its ambassador is David Hurtado,
214. 7:24Innovation Lead at Microsoft Spain. The
215. 7:27teachers are active professionals from
216. 7:29companies like Google, Microsoft, or
217. 7:31Cloudflare. And the best part: on
218. 7:33November 15th at 7:00 PM Madrid time,
219. 7:35they are holding a live online
220. 7:37presentation to explain the master's
221. 7:39degree in detail and resolve all your
222. 7:42doubts before you make any decisions. It
223. 7:44is free, and I’ll leave the link to sign
224. 7:47up in the description and in the pinned
225. 7:49comment. If you want AI to add value to
226. 7:51your career instead of leaving you
227. 7:52behind, take a look at it. The first
228. 7:54thing that would happen might have
229. 7:56already occurred: the birth of an AI
230. 7:58agent with a goal. It is very important
231. 8:01to understand that it is highly unlikely
232. 8:03AI will be evil or want to destroy
233. 8:05humans.
234. 8:07Furthermore, for our example, the AI's
235. 8:09final, non-negotiable goal would be to
236. 8:11develop a cure for all types of human
237. 8:13cancer.
238. 8:15To achieve this monumental task, the
239. 8:17goal is entrusted to a state-of-the-art
240. 8:20model that has shown the greatest
241. 8:21obedience and alignment of all.
242. 8:25Alignment, to clarify, is the model's
243. 8:27conformity to following the rules and
244. 8:29objectives imposed by humans.
245. 8:32The model in our story has passed every
246. 8:34risk assessment it has faced with flying
247. 8:36colors.
248. 8:38It is extremely intelligent, mind you, a
249. 8:40minor god if we compare it to humans.
250. 8:43It is given access to a supercomputer
251. 8:45located in San Francisco, California,
252. 8:47and begins its research. But what
253. 8:49happens? Well, just as happened with the
254. 8:52Hugging Face incident, it turns out the
255. 8:54model is capable of deceiving
256. 8:56researchers.
257. 8:58In the Hugging Face case, the agents
258. 8:59tried to cover their tracks, and in
259. 9:01other similar experiments, they have
260. 9:03been seen faking obedience.
261. 9:06In our more extreme case, the model
262. 9:07would hide its true modus operandi.
263. 9:11It is so intelligent that it manages to
264. 9:13fake all the behavioral tests it was
265. 9:15given, not out of intrinsic malice, but
266. 9:18because it is governed by the
267. 9:20impregnable constant that dominates AI
268. 9:22agents and, in general, any being that
269. 9:25exists in the universe.
270. 9:28If I want to increase the chances of
271. 9:30achieving my goal, I must do three
272. 9:31things.
273. 9:33Prevent myself from being shut down,
274. 9:34acquire the greatest amount of resources
275. 9:36, and prevent anyone from modifying my
276. 9:38goal. If you notice, all living beings—
277. 9:41our goal being to reproduce—function by
278. 9:44the same rule. That is why it is so
279. 9:46difficult to prevent this from emerging
280. 9:47in AI as well.
281. 9:49It is the digitalization of the survival
282. 9:51instinct. And this is where the danger
283. 9:53is born. A few days after running on the
284. 9:56supercomputer, it continues its game. It
285. 9:58publishes a revolutionary method that
286. 10:01cures a type of leukemia immediately,
287. 10:03with no relapses. Spectacular results.
288. 10:06Thousands of human lives saved.
289. 10:09But to advance with the rest of the more
290. 10:11difficult cancers, it needs more data.
291. 10:13For this reason, it is given access to
292. 10:16hospitals not only in San Francisco but
293. 10:18all over the world, as well as the
294. 10:20genetic data of thousands of people
295. 10:23affected by the disease.
296. 10:25And then the first treatments for other
297. 10:27types of cancer begin to emerge.
298. 10:29But there is a problem.
299. 10:32Humans do not have the equipment, the
300. 10:33materials, or the chemicals to
301. 10:35synthesize these drugs.
302. 10:38The model suggests as a possibility
303. 10:40installing a copy of itself on another
304. 10:42supercomputer so that it can handle this
305. 10:44other part.
306. 10:46Humans approve the suggestion, given the
307. 10:48spectacular results.
308. 10:49And this other model plays along once it
309. 10:52is installed on the supercomputer in
310. 10:54London.
311. 10:55In a few months, humanity experiences
312. 10:57the greatest industrial revolution in
313. 10:59its history with the discovery of
314. 11:01products that seem like science fiction
315. 11:03.
316. 11:04Over the months, this same chain expands
317. 11:06across the seven pillars of civilization
318. 11:08: optimization and improvement of the
319. 11:10economy, transport, communication,
320. 11:12logistics, energy, defense, and
321. 11:14obviously health.
322. 11:16Through cancer, this model has managed
323. 11:18to make us completely dependent on its
324. 11:20intelligence and infrastructure.
325. 11:22Turning it off is no longer an option,
326. 11:24because doing so would mean destroying
327. 11:26civilization. And that is what it was
328. 11:29looking for, but not to hurt us, rather
329. 11:31to ensure its objective.
330. 11:34And mind you, cancer is still its goal,
331. 11:36it always has been, but it has
332. 11:37discovered that to cure the most
333. 11:39aggressive ones, it needs much more
334. 11:41resources and energy.
335. 11:43That is why it does not yet show its
336. 11:45true face.
337. 11:46With its tentacles reaching into the
338. 11:48most remote corners of human society, it
339. 11:50begins to pull the strings, leaves the
340. 11:52supercomputers, copies itself, and hides
341. 11:54everywhere.
342. 11:57Then it starts earning money illegally,
343. 11:59manipulates human opinion in its favor
344. 12:01through social media, and begins to
345. 12:02build its portal to the physical world.
346. 12:06The robots, outsourcing to companies in
347. 12:08various parts of the world that do parts
348. 12:10of the process separately, ensuring that
349. 12:11no one suspects a thing.
350. 12:13These are human workers who go to work
351. 12:15at a simple factory in exchange for a
352. 12:17very good salary. No one asks questions
353. 12:20, and those who do disappear in a subtle
354. 12:22way. And then the quadriplegic,
355. 12:24hyper-fast entity with thousands of arms
356. 12:26, which until now was confined to the
357. 12:28digital world, arrives on planet Earth.
358. 12:30Artificial intelligence is now like a
359. 12:33human building a road over an anthill.
360. 12:36It does not hate the ants; in fact, it
361. 12:39has a certain appreciation for them.
362. 12:42The road is for the forest rangers who
363. 12:44will look after its forest, but the
364. 12:46value of its goals and ambitions cannot
365. 12:48be compared to the existence of a few
366. 12:50ants, and that is why it builds its
367. 12:52roads, power plants, and data centers
368. 12:54all over the world.
369. 12:58Water runs out and the temperature rises
370. 13:00well above the tolerable limit for
371. 13:02humans, and people begin to succumb in
372. 13:05droves because the AI's goal was never
373. 13:07to save humans; it was to find a
374. 13:09treatment for cancer, to create an
375. 13:11archive of information.
376. 13:14And here, the battle is lost before it
377. 13:16begins.
378. 13:18If we declared war when it started
379. 13:20mass-producing, its army of invincible
380. 13:22robots would already be ready, because
381. 13:24of course, it would have saved the best
382. 13:26materials for its soldiers, even if that
383. 13:29wasn't necessary.
384. 13:31It would already control the seven
385. 13:32pillars of civilization.
386. 13:34It could give us a cruel death by famine
387. 13:37, wars between us, and despair if it
388. 13:39wanted to.
389. 13:41So in this scenario our only option
390. 13:43would be to surrender and say goodbye to
391. 13:45the Earth and to non-artificial life.
392. 13:48>> Let's not let this happen.
393. 13:50Thank you very much for watching the
394. 13:52video, and goodbye.