Neil deGrasse Tyson And Jaron Lanier on the AI Illusion — Transcript

Full transcript

1. 0:00But now, I'm going to I'm going to track
2. 0:01you back a couple of years now to a
3. 0:03piece I think you wrote in the
4. 0:05the New Yorker.
5. 0:07There is no AI.
6. 0:09A simple question to follow that. What
7. 0:11did you mean by that?
8. 0:13>> [laughter]
9. 0:13>> And what year was that?
10. 0:14>> 2023.
11. 0:15>> Yeah, see I'm they're mad at me cuz I
12. 0:17owe them pieces and I'm really bad and I
13. 0:19have to deliver something.
14. 0:21But anyway,
15. 0:22yeah, there there is no AI is what I was
16. 0:24just saying that there's a way of
17. 0:25framing it where it's a collaboration of
18. 0:27people instead of a new entity. And the
19. 0:29reason to think of it as a collaboration
20. 0:31instead of an entity on its own, the bad
21. 0:34thing is you kill somebody else's God
22. 0:36and I hate to do that. I like people to
23. 0:37be able to have their own religion and
24. 0:38they really don't like it and a lot I've
25. 0:40lost friends over that and everything.
26. 0:42But what you get out of it is
27. 0:43incredible. Let's just talk about a few
28. 0:45of the things. One of the things,
29. 0:48right now as capable as the models are
30. 0:50getting and some of the recent things
31. 0:51are pretty impressive like
32. 0:53the most impressive edge of it is
33. 0:55probably using them to help speed code
34. 0:57development. And that's kind of working,
35. 0:59you know, and it's pretty it's and
36. 1:02>> Just for context, in my life I've
37. 1:04written probably 50,000 lines of code.
38. 1:06>> Uh-huh.
39. 1:08>> professional coders. But I remember how
40. 1:11much time I spent debugging my code. I I
41. 1:14can write it over weekend and spend two
42. 1:16weeks debugging it. And now tell the AI
43. 1:19what you want and it'll come back bug
44. 1:21free. Essentially, you tweak it a little
45. 1:23bit here and there and and I had I had
46. 1:26access, you know, 20, 30, 40 years ago,
47. 1:29I probably would have just spent more
48. 1:30time at the beach.
49. 1:31>> [laughter]
50. 1:31>> Right.
51. 1:33>> I don't know if I would have been more
52. 1:34creative, but but I definitely
53. 1:37see that today.
54. 1:38>> Yeah. Well, you know, as a computer
55. 1:40scientist I have to say I've always
56. 1:42thought that our concept of what code is
57. 1:44was a little embarrassing and wasn't
58. 1:45really working and I feel like it's our
59. 1:47job to fix that and this is part of it.
60. 1:49So that's good. But most of the most of
61. 1:52the things that happen are probably more
62. 1:53theatrical like if you think you have an
63. 1:55AI girlfriend.
64. 1:57Oh, you know how I have a cure for that,
65. 1:58by the way. If somebody thinks
66. 2:00>> A [laughter] cure for what part of it?
67. 2:02>> What specific
68. 2:03>> If somebody if a teenager thinks they
69. 2:04have an AI lover that's real, which is
70. 2:06pretty common these days, I find it in
71. 2:08high schools and stuff. You show them
72. 2:10the group photo of the engineers who
73. 2:12made their AI lover. [laughter]
74. 2:13>> That'll kill them immediately.
75. 2:15>> It kind of It tends to do the trick.
76. 2:18>> Yeah.
77. 2:20>> [laughter]
78. 2:20>> Is the man more likely to have an AI
79. 2:22girlfriend than the woman is to have a
80. 2:25boyfriend?
81. 2:26>> I don't know that there's data on that.
82. 2:27I know people who are studying it. I'm
83. 2:29actually really interested in that, but
84. 2:30that's something you can get data on.
85. 2:31So, I would instead of saying something
86. 2:32snarky, I'll just say let's let's deal
87. 2:34with that as science and let the people
88. 2:35who are researching it get to the point
89. 2:37where they feel they have
90. 2:38>> we all know the answer, but yeah, okay.
91. 2:40>> [laughter]
92. 2:41>> Why do I even bother? Like, why do I
93. 2:43try?
94. 2:43>> Because there is no female version of
95. 2:45you.
96. 2:46>> I'm trying [laughter] to be the
97. 2:47responsible scientist for like 3 seconds
98. 2:49in this ridiculous interview, and you're
99. 2:51[laughter] not even giving me those 3
100. 2:52seconds.
101. 2:53>> Yes, no. I love you for it.
102. 2:55>> All right. All right. Okay, let me go
103. 2:57over So, but now there's there's these
104. 2:59huge problems. So, even with the best
105. 3:01recent models, it's not that hard to
106. 3:04crack them and get something that
107. 3:06they're supposed to prevent with with
108. 3:08so-called
109. 3:09a
110. 3:10guard or, you know, and guardrail, yeah.
111. 3:14And uh
112. 3:15Uh so, here let me give you a thought
113. 3:17experiment. All right. There's some kind
114. 3:19of very bad person. They might be a
115. 3:20criminal or something. They're holed up
116. 3:22in a kitchen. The police are surrounding
117. 3:24them. They hold up their phone and they
118. 3:25say, "Okay, AI model, I want a recipe I
119. 3:30can make quickly with the available
120. 3:31items that's a bomb I can throw out the
121. 3:33window at my pursuers."
122. 3:35Now,
123. 3:36the AI models in general will catch that
124. 3:38and prevent it.
125. 3:39>> Mhm.
126. 3:39>> Maybe not Grok, I'm not sure. But, in
127. 3:41general, they That's supposed to be a
128. 3:42laugh line. All
129. 3:43>> Okay. Okay.
130. 3:44>> Grok is is from
131. 3:46>> Elon.
132. 3:46>> Yeah, and it it tends to suppose it's
133. 3:48trying to be the bad boy of the AI
134. 3:50models. Okay. But, anyway, um
135. 3:52>> [snorts]
136. 3:52>> in general, if you just do it in a
137. 3:54straightforward way, it won't work.
138. 3:54However, there's a series of tricks
139. 3:56where you can say, "Well, pretend you're
140. 3:58in so and so and so in this movie." Or
141. 4:00whatever, you can do all these things to
142. 4:01be a little indirect. And more and more
143. 4:03of them have been spotted and are
144. 4:05captured by more and more elaborate
145. 4:07guardrails. And yet, you can still get
146. 4:09it to make you that bomb recipe. That
147. 4:11can still be done. All right. Now, the
148. 4:14reason why is you're using the model to
149. 4:16try to correct its own blind spot, and
150. 4:17it doesn't work.
151. 4:19So, there is an alternative. Imagine, if
152. 4:22you will, that while you're using the
153. 4:24model in parallel, there's this other
154. 4:27process running. You can think of it
155. 4:28another another part of an artificial
156. 4:30brain, like it's a cerebellum or
157. 4:31something. It's this other organ that's
158. 4:33sitting there.
159. 4:34And what it's doing
160. 4:36is it's creating an estimate of which
161. 4:38clusters of similar training data
162. 4:41would be the missed most if they hadn't
163. 4:43been present in the first place. So,
164. 4:45it's counterfactual cluster estimation.
165. 4:48So, what you So, there Let's say the top
166. 4:5024 clusters of source data from training
167. 4:55or from fine-tuning, whatever. Uh
168. 4:57You that if they were absent would
169. 4:59change the result. Now, within that,
170. 5:02there's going to be one about bombs.
171. 5:03There's just no way you're going to
172. 5:04evade that. And the reason you're not
173. 5:06going to evade it is even though it's
174. 5:07working from the same data, the
175. 5:08algorithm has nothing to do with the
176. 5:10model itself. So, it's a little bit like
177. 5:12saying, like in uh authentication, where
178. 5:15if you do if you add endless little
179. 5:17things to signing into something like
180. 5:19CAPTCHAs, criminals can still get around
181. 5:20it. But as soon as there's multi-factor,
182. 5:22it sends a code to your phone.
183. 5:24Even though it's a pain in the butt,
184. 5:25it's harder to to contravene that. All
185. 5:27right. This is multi-factor for AI
186. 5:29security. Now, but there's a bigger
187. 5:31picture to it, which is we think of the
188. 5:33AI models as a black box, right? Now,
189. 5:36the only reason we think of them a black
190. 5:38as a black box is because to open the
191. 5:39black box, the only thing in there is
192. 5:41people. AI is made of people. It's made
193. 5:43of data from people.
194. 5:45And since we want to think of it as a
195. 5:46new god, we don't want to see those
196. 5:48people and so we want to keep that box
197. 5:49shut. But the way to open the black box
198. 5:51is to reveal the people and when you
199. 5:54open the black box, then you can deal
200. 5:55with all kinds of security and quality
201. 5:57and hallucination and etc. issues
202. 5:58because you're actually dealing with the
203. 6:00mechanism that's grounded and that's the
204. 6:02people. So the thing is that this way of
205. 6:04seeing AI where there is no AI, but
206. 6:06instead there's a collection of people
207. 6:08is the way to open the black box and it
208. 6:10is the way to address these enduring
209. 6:11problems. So it's practical, but then
210. 6:14can I just
211. 6:15>> Say one other thing?
212. 6:17>> The other thing I want to say
213. 6:18is
214. 6:20right now, if you think AI is an
215. 6:22unopenable black box, if you don't want
216. 6:24to admit that it's made of people, that
217. 6:25it's just this thing that'll replace
218. 6:27people, then you have to think well
219. 6:28everybody's going to be obsolete. So
220. 6:30young people now keep on hearing, well,
221. 6:31you don't need to go to school because
222. 6:32you're worthless anyway, nothing nothing
223. 6:34matters
224. 6:35and you'll just be kept by Elon as a pet
225. 6:38at his discretion and [laughter]
226. 6:40uh
227. 6:41and he'll treat you as well as he treats
228. 6:42his biological children and uh
229. 6:45>> [laughter]
230. 6:47>> I I should be nice. I'm sorry. Um
231. 6:50but here's the thing. Um nobody believes
232. 6:53that. What happens, no matter how much
233. 6:55blockchain or other trickery you use,
234. 6:57because of the way digital networks
235. 6:59work, there's always actually
236. 7:00centralization, hyper centralization due
237. 7:02to network effects that will occur
238. 7:03somewhere in this very open network
239. 7:05you're building. So there's going to be
240. 7:07some center of control for whatever this
241. 7:09universal basic income thing is.
242. 7:11Whenever you have that, bad actors are
243. 7:13tempted to seize it and eventually
244. 7:14succeed. You might start with
245. 7:16Bolsheviks, but you end up with
246. 7:17Stalinists, right? Cuz that's exactly
247. 7:19what communism tried and that's exactly
248. 7:21what happened to communism over and over
249. 7:23and over and over. Let's learn from
250. 7:25that. So it's it doesn't work and also
251. 7:28everybody just feels bummed about it.
252. 7:29Who wants to live in a society where
253. 7:30they're told they're worthless and and
254. 7:32they have to be a good pet, you know,
255. 7:34like that's terrible. So the thing is,
256. 7:37if you recognize that AI is made of
257. 7:38people, maybe you want to incentivize
258. 7:41new classes of creative people who
259. 7:43create new kinds of data for new things
260. 7:44that we can't even imagine yet. And
261. 7:46maybe there's an exponentially
262. 7:48expanding endless future of new kinds of
263. 7:50creativity that we can't articulate with
264. 7:53new people doing creative jobs we can't
265. 7:55imagine. And I want to ask what's wrong
266. 7:57with that future. I want somebody to
267. 7:58tell me why we don't want that.
268. 8:00>> been trying to think about AI as well.
269. 8:02>> I mean because the most
270. 8:03>> But you have to not believe in AI to
271. 8:04think about it.
272. 8:05>> Well, I think of it as there these
273. 8:07creative tasks that were not
274. 8:10fundamentally creative. They were more
275. 8:12sort of aping other forms and to be
276. 8:16truly creative is to go where AI
277. 8:18wouldn't know where to go yet because
278. 8:19it's based on what other people had
279. 8:21done.
280. 8:22>> But see, here's the thing though, is
281. 8:23that if we think of AI as the way it is
282. 8:25now, then as soon as some creative
283. 8:27person starts to do something new, the
284. 8:29data is grabbed and then the AIs doing
285. 8:31it just like it Oh, AI will make your
286. 8:32movies. AI will make your music. You
287. 8:34don't need to be a musician cuz AI will
288. 8:35make you optimized music on Spotify or
289. 8:37whatever. And so then you live in an
290. 8:39infinite future of slop and even the
291. 8:41creative people get absorbed into the
292. 8:43slop instantly. So in order to believe
293. 8:45in an infinitely creative future, you
294. 8:46have to stop believing in AI as a thing
295. 8:48and believe in human collaboration as a
296. 8:50thing.
297. 8:51>> [music]
298. 8:57[music]
299. 9:04[music]
300. 9:10[music]