228: AI Snake Oil by Arvid Narayanan & Sayash Kapoor

Click any line to jump to that moment.

1. 00:00:00 All right, Mike, welcome back.
2. 00:00:01 Are you ready to talk about artificial intelligence,
3. 00:00:06 large language models, prediction, generation,
4. 00:00:10 all of the things related to technology?
5. 00:00:13 To be honest, I'm not sure.
6. 00:00:16 You know, I'm not either.
7. 00:00:18 I was doing the outline for this one, and I was like, well,
8. 00:00:21 do we want to go chapter by chapter?
9. 00:00:23 Or do we think there's a better way
10. 00:00:25 to have this conversation?
11. 00:00:27 And I think we've got to, but yeah.
12. 00:00:30 I think we're going to go chapter by chapter.
13. 00:00:31 But before we get there, let's go into follow-up.
14. 00:00:35 So I'll start with my follow-up items.
15. 00:00:38 I was to brainstorm and create a mastermind group.
16. 00:00:41 I'm going to give myself a yellow on this one.
17. 00:00:44 My yellow is because I sent an email to an individual
18. 00:00:50 that I would like to be in the mastermind group.
19. 00:00:52 And this individual is key.
20. 00:00:55 So I really want this individual to be in this group.
21. 00:00:59 If he's not, then I don't know what happens to the group.
22. 00:01:01 Like, I don't know if I am going to push the group forward
23. 00:01:04 without this individual in it, because I think he and I
24. 00:01:05 would partner really well and do that really well.
25. 00:01:08 So he got back to me.
26. 00:01:10 I need to schedule, and we need to think about it.
27. 00:01:12 And we kind of need to talk about what the details of it
28. 00:01:13 would look like.
29. 00:01:14 And then he and I, I think, will figure out
30. 00:01:16 if there's a bigger group and what the bigger group looks
31. 00:01:18 like and how that bigger group works.
32. 00:01:19 So I'm going to give myself a yellow on this.
33. 00:01:21 So the group is not fully formed yet.
34. 00:01:24 There's nothing on the calendar in terms of us meeting,
35. 00:01:26 but the preparatory work is under underway.
36. 00:01:29 The next thing I had was what scares me and excites me.
37. 00:01:32 And then how can I develop a voluntary force function
38. 00:01:36 for this thing?
39. 00:01:37 And I did not do this at all.
40. 00:01:39 I didn't even think about this once.
41. 00:01:41 So I'm letting this one go.
42. 00:01:43 I'm not going to come back to this one.
43. 00:01:45 I'm just going to let it fly away and be totally OK with that.
44. 00:01:49 The last one was do Pat's Personal Mission Statement
45. 00:01:51 Activity, which I have done.
46. 00:01:52 I had done that one prior to the episode
47. 00:01:55 where we talked about lean learning.
48. 00:01:58 So that one's completely done.
49. 00:01:59 And I'm a big, big solid green on that one.
50. 00:02:02 And I feel really good about that.
51. 00:02:03 So those are my follow up items.
52. 00:02:05 Mike, you had four.
53. 00:02:07 Would you like to talk us through those?
54. 00:02:09 I did.
55. 00:02:10 And no, I would not like to talk you through those.
56. 00:02:12 So let's go to the group.
57. 00:02:13 OK, moving on now.
58. 00:02:14 No, I'm just kidding.
59. 00:02:16 So what about my work gives me energy?
60. 00:02:19 I didn't really think about this,
61. 00:02:21 but I think I have an answer to this anyways.
62. 00:02:25 The thing about my work that gives me energy
63. 00:02:27 is making connections with other people
64. 00:02:31 and authentically helping them.
65. 00:02:33 So the live streams that I've been doing and really anything
66. 00:02:38 connected to the library, that is very energizing and life
67. 00:02:43 giving to me.
68. 00:02:47 There's a downside or a shadow side to this as well,
69. 00:02:49 because if there is an event on the library calendar
70. 00:02:53 and nobody shows up, I'll get super, super depressed.
71. 00:02:57 Yeah, I can see that.
72. 00:02:58 I would too.
73. 00:02:59 Which honestly doesn't happen anymore,
74. 00:03:01 but it definitely did at the beginning.
75. 00:03:04 And it was enough for me to be like, I don't want to do this.
76. 00:03:08 And I just wouldn't schedule things for months, which
77. 00:03:10 is totally the wrong approach to that sort of thing.
78. 00:03:14 But that definitely gives me energy.
79. 00:03:15 And then I recognize that large editing projects
80. 00:03:20 are the things that I tend to procrastinate on.
81. 00:03:24 I'm good at them.
82. 00:03:25 Like I do these modules for screencasts online every month
83. 00:03:29 or so.
84. 00:03:30 And they're like 40-minute screencasts walking through.
85. 00:03:35 Like the last one I did was on Ecamm Live.
86. 00:03:38 I think I've shared that story last time,
87. 00:03:39 how it got released actually.
88. 00:03:41 The day I was at Macstock talking to the Ecamm people,
89. 00:03:43 and they found it and they're like, hey, your tutorial is great.
90. 00:03:46 And I'm like, what are you talking about?
91. 00:03:49 But when I'm working on those, it always
92. 00:03:51 feels like a huge, heavy lift.
93. 00:03:55 So I guess that's evidence that I should delegate as much video
94. 00:04:00 editing as I can.
95. 00:04:02 The one I did succeed on was sharing my work on social media.
96. 00:04:06 I don't think I very clearly defined this, though.
97. 00:04:11 I am basically--
98. 00:04:14 well, I'll talk through my approach to this,
99. 00:04:18 and it's going to be sloppy because I haven't really
100. 00:04:20 codified this anywhere.
101. 00:04:21 But in my mind, anyway, is it kind of connects.
102. 00:04:24 So the thing that I'm really working backwards from
103. 00:04:28 is I'm trying to get to the point where
104. 00:04:30 I am releasing a new edited YouTube video every week.
105. 00:04:37 So working backwards, I need topics
106. 00:04:40 for all of those videos.
107. 00:04:42 I need to figure out what I'm going to say about those topics,
108. 00:04:46 and I need to script those, record them,
109. 00:04:48 hand them off to the editor, add the screencast,
110. 00:04:50 all the post-production stuff that takes far longer
111. 00:04:52 than I want it to.
112. 00:04:53 But basically, if I start with those topics,
113. 00:04:58 that's really like the first bottleneck
114. 00:05:00 is figuring out what are the next six YouTube videos
115. 00:05:04 going to look like.
116. 00:05:05 And then what I've been doing from there
117. 00:05:07 is like, well, I have to figure this stuff out anyways,
118. 00:05:10 what I'll do is I will actually just take that topic,
119. 00:05:14 and I'll write about it for the newsletter,
120. 00:05:16 because that will get me 60% of the way there
121. 00:05:20 to making the video script.
122. 00:05:22 And once I have some momentum, then I can take it and--
123. 00:05:26 so there's consistency I feel like in the messaging
124. 00:05:29 for the first time, which even that is a net win.
125. 00:05:33 And then even further back from that,
126. 00:05:36 the live streams that I've been experimenting
127. 00:05:38 with on Wednesdays for practical PKM,
128. 00:05:42 those are a way for me to kind of talk through things
129. 00:05:47 before I know what I really want to say in the newsletter.
130. 00:05:50 So I've done some on--
131. 00:05:52 I think the last one I did was on the basis plug-in,
132. 00:05:55 and I was talking through the On This Day feature thing
133. 00:05:57 that I figured out for my daily note.
134. 00:06:01 And being able to do that live, even though it's not truly
135. 00:06:04 live like this, where people can talk back to you,
136. 00:06:07 but they can chime in via the comments,
137. 00:06:09 and then you can respond that way.
138. 00:06:12 It helps me kind of explore the topic in a different way,
139. 00:06:15 and then it helps me figure out kind of the direction
140. 00:06:17 that I want to go.
141. 00:06:18 So I start with that topic, and then it's like,
142. 00:06:20 where do I want to go from this live stream about it?
143. 00:06:22 That'll give me a direction from there, right about it.
144. 00:06:25 That'll give me some momentum, and then finish it
145. 00:06:28 by making the video script.
146. 00:06:31 Once I do all that, then it's really just
147. 00:06:34 take pieces of this and share it on social media.
148. 00:06:38 The live stream is already on social media,
149. 00:06:41 so I'm sharing--
150. 00:06:42 I'm live now, come hang out.
151. 00:06:44 And I've been getting consistently 20 people at a time
152. 00:06:47 in those live streams, which I'm pretty happy with.
153. 00:06:51 And then they stick around on the YouTube channel,
154. 00:06:53 and they accumulate views and watch time and stuff like that.
155. 00:06:56 So it's a win-win, I feel, for the channel.
156. 00:07:02 I probably could do a little bit better job of taking
157. 00:07:05 every bit of the things that I figure out
158. 00:07:08 and making sure all of them get scheduled for social media.
159. 00:07:10 That's the last part of this for me, I think.
160. 00:07:13 And that I've just spent too busy to really codify,
161. 00:07:16 but that's going to happen through,
162. 00:07:18 thankfully.
163. 00:07:19 And yeah, I have a clear path forward for that.
164. 00:07:24 Just need a little bit of time to make it happen.
165. 00:07:27 The other two that I had, which I failed miserably on,
166. 00:07:29 check out the companion course.
167. 00:07:31 I did not do this, Toby did this.
168. 00:07:33 He told me it was free.
169. 00:07:34 And once I heard a little bit about what it was,
170. 00:07:40 I kind of feel like my curiosity was satisfied.
171. 00:07:43 I didn't need to go look at it.
172. 00:07:47 And then the mastermind groups for the library,
173. 00:07:48 this is definitely something I want to do at some point,
174. 00:07:52 but it's not something that I should be worrying about right
175. 00:07:56 now, so just basically put that on the back burner.
176. 00:08:00 Nice.
177. 00:08:01 Nice, so what would you, I mean, overall, green, yellow, red?
178. 00:08:06 I think the first one is yellow.
179. 00:08:09 The second-- so what about my work gives me energy?
180. 00:08:12 The second one, sharing the work on social media.
181. 00:08:14 I'm willing to give myself a green,
182. 00:08:15 even though it's not completely done, I feel like.
183. 00:08:18 Yeah, and then the other two are red.
184. 00:08:20 We're red.
185. 00:08:20 But actively red, like you made a decision
186. 00:08:23 to make those red.
187. 00:08:24 Intentionally red.
188. 00:08:25 Intentionally red, there you go.
189. 00:08:26 Very actively, intentionally.
190. 00:08:28 Yep, that's great.
191. 00:08:30 All righty, are we ready to get into today's book
192. 00:08:32 as best as we can?
193. 00:08:36 Let's do it, let's do it.
194. 00:08:37 All right, so our book today is called AI Snake Oil,
195. 00:08:41 What Artificial Intelligence Can Do, What It Can't,
196. 00:08:43 and How to Tell the Difference.
197. 00:08:45 The book is authored by Arvind Nairayana.
198. 00:08:50 Nah, I knew I was going to get it wrong.
199. 00:08:52 Nairay Anan.
200. 00:08:55 Yeah, OK, cool.
201. 00:08:56 Mike shook his head.
202. 00:08:57 And then Syash Kapoor, so Arvind,
203. 00:09:00 I'm really sorry that I messed your name up
204. 00:09:01 if you listen to this, where they are both at Princeton.
205. 00:09:06 So Arvind is a faculty member at Princeton.
206. 00:09:10 He's a computer science professor.
207. 00:09:13 He runs a lab that is he's the director of the Center
208. 00:09:16 for Information Technology Policy.
209. 00:09:18 So that actually kind of makes sense in terms of the book.
210. 00:09:20 If you read the book, or if you've already read the book,
211. 00:09:22 that would actually kind of make a ton of sense.
212. 00:09:25 And then Syash is a computer science PhD candidate.
213. 00:09:29 So he's almost got his doctorate.
214. 00:09:31 He works in the Center for Information Technology Policy.
215. 00:09:35 He kind of co-authored this out of his work.
216. 00:09:38 Mike, do you remember Syash is the one who
217. 00:09:41 worked at Facebook, right?
218. 00:09:43 Before he came back to grad school,
219. 00:09:45 if I'm remembering correctly.
220. 00:09:48 To be honest, I don't know.
221. 00:09:51 And I don't think it really matters except for the one section
222. 00:09:55 where they talk about AI with social media.
223. 00:09:59 The only reason I brought it up was because it kind of
224. 00:10:02 brings a little bit of the level of he
225. 00:10:03 was in a world that was using this professionally
226. 00:10:08 or in the workplace kind of setting, which I think
227. 00:10:10 adds some flavor to it.
228. 00:10:12 So the book's got an introduction, which they count as chapter
229. 00:10:15 one, and then eight chapters after that.
230. 00:10:17 It's got quite a bit of footnotes and booknotes
231. 00:10:21 and then notes associated with it where they
232. 00:10:23 reference different studies and different things.
233. 00:10:26 The first chapter, Mike, if you're ready to roll into it?
234. 00:10:29 Well, let's reel real quick on the format of the book.
235. 00:10:33 I think this is kind of weird because it's what?
236. 00:10:36 290 pages before you get to the end.
237. 00:10:40 Yeah, something like that.
238. 00:10:42 There's a ton of references, as you mentioned.
239. 00:10:44 There's like 60 some citations in each chapter.
240. 00:10:51 That last citation section in the book
241. 00:10:54 is crazy long, which makes it actually kind of hard
242. 00:10:56 to find what you're looking for.
243. 00:10:57 So there's a ton of research, basically, or a ton of sources
244. 00:11:01 that have gone into this book.
245. 00:11:04 It's very long, though, and if you take out the introduction,
246. 00:11:07 there are really seven different chapters.
247. 00:11:09 I feel like it's laid out kind of weird.
248. 00:11:12 Oh, yeah, I would agree with you.
249. 00:11:13 And I don't think that does the book any favors
250. 00:11:18 at the beginning.
251. 00:11:20 And I think you probably could break it down
252. 00:11:23 into different sections and have different chapters.
253. 00:11:27 I'm not necessarily sure that's the right approach, either.
254. 00:11:29 The chapters themselves do have a bunch of sub-chapters
255. 00:11:32 that are listed.
256. 00:11:34 So if you want to view the chapters in here,
257. 00:11:36 kind of as sections, I feel like you could do that.
258. 00:11:39 But yeah, it's kind of strange the way that this book is laid out.
259. 00:11:46 It reads a lot like "Academics" wrote it,
260. 00:11:50 which, as we're seeing this, this unroll.
261. 00:11:54 Maybe that's why I was kind of like,
262. 00:11:56 oh, this makes total sense to me.
263. 00:11:57 I kind of understand where they're going.
264. 00:11:59 I think you're right in saying that the seven legitimate
265. 00:12:04 chapters, introduction, but then the seven other chapters,
266. 00:12:08 they're not mapped out.
267. 00:12:10 We've read books before where it's like, here, this part one,
268. 00:12:14 it's got three chapters underneath it.
269. 00:12:15 Here's part two, it's got three chapters underneath it.
270. 00:12:17 Here's part three, it's got three chapters underneath it.
271. 00:12:19 And then there's a thread that ties part one,
272. 00:12:20 part two, and part three together.
273. 00:12:22 That really doesn't exist in this book.
274. 00:12:26 I was looking real quick to see who the publisher was.
275. 00:12:29 So it's Princeton University Press.
276. 00:12:30 So this is an academic book written for academics.
277. 00:12:34 It's like it's got academia written all through it.
278. 00:12:36 And what I'll say is, I think that's
279. 00:12:39 why you had said about like each chapter,
280. 00:12:43 kind of being standalone and having a bunch of references,
281. 00:12:46 that's a very academic style of writing is,
282. 00:12:49 I'm going to write this chapter and it's kind of a self-contained
283. 00:12:52 idea, and then we'll write this chapter and it's kind of a self-contained idea.
284. 00:12:54 And I've read a bunch of books before where you'll
285. 00:12:58 solicit different authors to write each of those chapters.
286. 00:13:01 So you go out and you get the expert to write that chapter,
287. 00:13:04 and you get the expert to write that chapter.
288. 00:13:05 And then there's an editorial team
289. 00:13:07 that'll write a front matter and a back matter
290. 00:13:09 to kind of try to tie everything together.
291. 00:13:11 It just so happens in this book, they both wrote the entire thing.
292. 00:13:15 They didn't farm out any of the chapters
293. 00:13:17 or get support for any of the chapters.
294. 00:13:20 The one last thing I thought would be interesting to talk about
295. 00:13:23 before we get into the introduction is,
296. 00:13:25 this was listed as one of nature's top 10 books in 2024.
297. 00:13:29 They've got a fairly significant sub stack associated with this.
298. 00:13:33 So they've gained some pretty decent popularity in the AI space.
299. 00:13:37 I would say in the more public policy and technical side
300. 00:13:42 of the AI space, as opposed to the general, your grandma, your brother,
301. 00:13:48 your dad, who is like, I just kind of want to know what's going on with AI.
302. 00:13:52 I feel like I should make a disclaimer before we go any further.
303. 00:13:56 I'm very into AI.
304. 00:13:57 I've encouraged us for Bookworm to read Kevin Kelly's book about technologies
305. 00:14:03 and now we're reading this AI book.
306. 00:14:04 So if you really, really hated this book and you were like,
307. 00:14:07 this book was terrible and I did not like any bit of it,
308. 00:14:10 or if you listen to us talk about it and you think, nope, that's good.
309. 00:14:14 It's all on me.
310. 00:14:15 Don't put any of it on Mike.
311. 00:14:16 It's all on me and--
312. 00:14:18 Corey gets all the blame and the credit.
313. 00:14:20 I get all the blame.
314. 00:14:21 So just, yeah, no, you'll see where we're going with this.
315. 00:14:24 So let's get into the--
316. 00:14:25 Dude, you're foreshadowing your rating.
317. 00:14:26 Oh, I'm totally--
318. 00:14:27 I'm totally foreshadowing my rating, totally 100%.
319. 00:14:31 The question, Mike, is--
320. 00:14:33 Let me just share one other thing about the cover here
321. 00:14:35 so that the subtitle is what artificial intelligence can do, what it can't,
322. 00:14:40 and how to tell the difference.
323. 00:14:42 And AI is changing so quickly.
324. 00:14:45 This is the thing that made me leery of this.
325. 00:14:49 I feel like even if it was able to deliver on those promises entirely,
326. 00:14:55 it's a very short shelf life.
327. 00:14:59 But that was just my first impressions
328. 00:15:01 going into the book before we get into the discussion of a chapter by chapter.
329. 00:15:06 OK.
330. 00:15:07 All right, so let's go to the introduction.
331. 00:15:08 In the introduction, they've got a couple different goals here
332. 00:15:11 that are then going to bleed through to the rest of the book.
333. 00:15:15 One of them is they're going to make a distinction between generative AI
334. 00:15:17 and predictive AI.
335. 00:15:19 And I'll say it at the very start.
336. 00:15:23 They're for generative AI.
337. 00:15:25 They are very against predictive AI.
338. 00:15:27 Then they introduce this idea of snake oil.
339. 00:15:30 And just like in the old West, where people would try to sell you the--
340. 00:15:34 this tonicle save everything.
341. 00:15:36 And it'll make you better.
342. 00:15:37 And it'll heal all the diseases and all that stuff.
343. 00:15:39 They're saying that predictive AI is snake oil at this point.
344. 00:15:44 This is where I think the title of their book is a little bit--
345. 00:15:48 not disingenuous because I don't think they were doing it intentionally.
346. 00:15:51 But I would have wanted them to be more specific and then
347. 00:15:54 say like predictive AI snake oil because they're very clear
348. 00:15:59 that generative AI and chatbots and some of these things
349. 00:16:02 are very valuable.
350. 00:16:02 And image generation, they're doing a fine job.
351. 00:16:06 There's really nothing wrong with those.
352. 00:16:08 They're doing what people say they're going to do.
353. 00:16:11 But when you get into the predictive side,
354. 00:16:12 they don't have the evidence for it.
355. 00:16:14 So let's get a little bit more detailed.
356. 00:16:16 And basically, they say snake oil is something that doesn't do what
357. 00:16:21 it's advertised to do.
358. 00:16:22 Or it can't possibly do what it's advertised to do.
359. 00:16:25 So that's why they've titled the book snake oils
360. 00:16:27 because they say predictive AI can't do what people are saying it can do.
361. 00:16:32 I'll stop here, Mike, and let you give us some thoughts on the introduction.
362. 00:16:37 Yeah, well, I think that's true.
363. 00:16:39 And the next couple of chapters are on predictive AI.
364. 00:16:44 You can tell at this point in the book that really they've
365. 00:16:48 got a vendetta against predictive AI, which I wouldn't--
366. 00:16:52 I wouldn't say is unfounded.
367. 00:16:56 It's just not necessarily the approach that I would have picked
368. 00:17:01 for a book that I wanted to read about AI.
369. 00:17:06 Maybe that's because I have more than an absolute beginner level
370. 00:17:12 understanding of this stuff.
371. 00:17:14 Like we have read some of those books.
372. 00:17:16 The one that I continue to point people to is co-intelligence
373. 00:17:19 by Ethan Mollack.
374. 00:17:20 I feel like that's a great discussion about AI in general.
375. 00:17:23 But I don't think he really even addressed predictive AI very specifically
376. 00:17:30 in that book.
377. 00:17:30 It was more about AI in general.
378. 00:17:32 This one takes the opposite approach.
379. 00:17:33 It's going to consider each type of AI under the microscope.
380. 00:17:38 So very different feel.
381. 00:17:41 I feel also that the formatting/editing of this book--
382. 00:17:46 again, you mentioned it's a very academic book.
383. 00:17:49 I don't really like academic books I have discovered.
384. 00:17:54 One of the reasons I don't like the academic books
385. 00:17:57 is that I feel there is a lot of focus on what they're saying
386. 00:18:04 and not as much focus on how they're saying it.
387. 00:18:09 So right away in this section, the big argument they're laying out here
388. 00:18:13 are the three questions for determining whether describing something
389. 00:18:17 as AI is appropriate.
390. 00:18:20 Not all three of these are actually shared as questions.
391. 00:18:24 The third one is a statement.
392. 00:18:27 I didn't realize that when I read it, but yes.
393. 00:18:29 And when I'm putting together the mind map for this,
394. 00:18:32 I'm like, wait, I have to rewrite this as a question,
395. 00:18:35 because it's the three questions.
396. 00:18:37 And I'm going to go back and reference this later.
397. 00:18:39 Most people, they're just going to keep cranking through the words
398. 00:18:42 that keep appearing before their eyeballs.
399. 00:18:44 But I was like, wait, no, that's wrong.
400. 00:18:50 So little things like that definitely bug me.
401. 00:18:54 Yeah, I don't know what else to really say
402. 00:18:55 about the introduction section.
403. 00:18:57 I think the questions are actually helpful.
404. 00:18:59 The first one, does the task require creative effort or training
405. 00:19:04 for a human to perform?
406. 00:19:07 The second one, was the behavior of the system
407. 00:19:09 directly specified in code by the developer
408. 00:19:12 or did it indirectly emerge, and the third one,
409. 00:19:14 does the system make decisions more or less autonomously
410. 00:19:17 and possess some degree of flexibility and adaptability
411. 00:19:20 to the environment?
412. 00:19:20 And that's the one where they just said it as a statement.
413. 00:19:25 Feels like something if you had an editor at a--
414. 00:19:30 I can't even think of the big publishers now.
415. 00:19:33 Penguin?
416. 00:19:34 Yeah.
417. 00:19:36 They would have been like, no, no, no, this doesn't make sense.
418. 00:19:38 Let's clean this up a little bit.
419. 00:19:39 And it's not that big a deal.
420. 00:19:41 But it was a little bit of friction for me as I'm reading it
421. 00:19:46 and priming the pump, I guess, for how
422. 00:19:50 I'm going to engage with the rest of the chapters.
423. 00:19:54 But not all AI, snake oil, they mentioned at the beginning,
424. 00:19:57 and I think you'll definitely get that as we work
425. 00:19:59 through the rest of this.
426. 00:20:01 But they're definitely going to lead with guns
427. 00:20:06 ablazing in chapters two and three.
428. 00:20:09 Absolutely.
429. 00:20:10 So you got at one of the things is there
430. 00:20:12 way to try to define what AI is.
431. 00:20:15 And even there, they say, if you can answer yes
432. 00:20:19 to those three questions, then it might be AI.
433. 00:20:22 They still don't even come out as strong and say,
434. 00:20:24 then it is AI.
435. 00:20:25 They just say, oh, it maybe is AI.
436. 00:20:27 Like, it's in the wheelhouse of possibly being AI.
437. 00:20:31 One of the things that I liked about this,
438. 00:20:34 you know me at this point, tables and graphs make me happy.
439. 00:20:37 So they throw a table out there.
440. 00:20:39 The landscape of AI, snake oil, hype and harms.
441. 00:20:42 They've got on the x-axis benign to harmful.
442. 00:20:46 And then on the y-axis works up to snake oil.
443. 00:20:51 And they throw different ideas or different ways
444. 00:20:54 that AI is being used.
445. 00:20:57 So for instance, like autocomplete.
446. 00:20:59 They would say autocomplete is fairly benign and it works.
447. 00:21:02 We can do that pretty well at this point.
448. 00:21:04 And you think about that?
449. 00:21:05 Well, that's been around since at least the T9 cell phone
450. 00:21:08 days, right, where it was trying to autocomplete your words.
451. 00:21:11 Then they go the whole way up to cheating detection,
452. 00:21:14 video interviews, and hiring, and criminal risk detection.
453. 00:21:19 And they say, that's basically really harmful
454. 00:21:21 and it's snake oil, that it doesn't work at all.
455. 00:21:24 And it really has a negative consequence.
456. 00:21:25 So they do that.
457. 00:21:27 I think they do a good job outlining who the book is for.
458. 00:21:30 They throw these three different categories.
459. 00:21:32 You want to get a sense of what's going on.
460. 00:21:35 That's one category.
461. 00:21:36 You need to make decisions about AI in your workplace.
462. 00:21:39 That's another category.
463. 00:21:40 You may be interested in AI because you
464. 00:21:42 want to take action against the harms.
465. 00:21:43 So that's more of a proactive or activist kind of standpoint.
466. 00:21:47 I honestly think that this is why I wanted to read this book
467. 00:21:51 is because I'm in that second camp.
468. 00:21:53 That I need to make some decisions about AI at the university
469. 00:21:57 and in life and think about that from that perspective.
470. 00:22:01 So I'm not going to tell you that I love this book,
471. 00:22:04 but I'm going to tell you that I think this book was
472. 00:22:06 valuable for me bigger picture.
473. 00:22:09 That it allowed me to see a different part of thinking
474. 00:22:12 about AI that I don't often get everywhere else.
475. 00:22:15 I get AI is going to ruin education,
476. 00:22:18 and AI is going to do this, and AI is going to do that.
477. 00:22:20 And they were a more tempered voice to say, well, hold on.
478. 00:22:24 It's probably not, especially if it's in the predictive side.
479. 00:22:26 So I think that's a good thing.
480. 00:22:31 It's worth reading for me, even though, as you're going to see,
481. 00:22:34 I'm not a huge fan of the rest of the book.
482. 00:22:37 But chapter two, are you ready to go to chapter two?
483. 00:22:41 - Let's do it.
484. 00:22:42 - Okay, so chapter two is how predictive AI goes wrong.
485. 00:22:46 And really what they're going to do here
486. 00:22:48 is they're going to talk about a couple different ways
487. 00:22:50 that AI goes wrong.
488. 00:22:52 They think predictive AI falls short
489. 00:22:54 of what people say it can do.
490. 00:22:56 They think that over automation,
491. 00:22:59 people are using it to make too many automated decisions.
492. 00:23:03 They are specific in throwing out their thesis
493. 00:23:07 that companies and governments have many misguided
494. 00:23:11 commercial or bureaucratic reasons
495. 00:23:12 for deploying predictive AI.
496. 00:23:14 But it's faulty, and that's causing systematic issues
497. 00:23:17 and systematic consequences.
498. 00:23:19 They throw a table in there
499. 00:23:20 that's the five reasons predictive AI fails.
500. 00:23:24 It's a good prediction can result in a bad decision.
501. 00:23:29 People can strategically game opaque AI.
502. 00:23:32 Users over rely on AI without adequate oversight or recourse.
503. 00:23:36 Data for training AI may come from different population.
504. 00:23:40 Predictive AI can increase inequality.
505. 00:23:43 Essentially that's what the whole chapter is.
506. 00:23:46 It's summarized in that table.
507. 00:23:47 So they wrote about all of that stuff
508. 00:23:49 and they summarized in that table.
509. 00:23:51 And the thing that gets me,
510. 00:23:53 and this is where I told you,
511. 00:23:55 I was good with the summary.
512. 00:23:57 There was just too much in every one of those five sections
513. 00:24:01 and I would just get bored
514. 00:24:03 and I'd be like, okay, I get your point.
515. 00:24:05 Okay, I get your point.
516. 00:24:06 Okay, I get your point.
517. 00:24:07 And then we got to the summary table.
518. 00:24:08 I was like, man, if only you had given me this 30 pages ago,
519. 00:24:14 I'd have been probably happy
520. 00:24:15 and I'd have been able to make it through much faster.
521. 00:24:17 So that's a new example of telling someone
522. 00:24:21 what you're gonna tell them,
523. 00:24:22 telling them and telling them what you told them.
524. 00:24:24 - Exactly, exactly.
525. 00:24:25 So that's chapter two,
526. 00:24:27 how predictive AI goes wrong.
527. 00:24:29 It's a big elaboration on those five.
528. 00:24:31 Those five points.
529. 00:24:33 - I think those five points are actually legit.
530. 00:24:36 - Agreed.
531. 00:24:37 - And the argument,
532. 00:24:38 if you just take those arguments
533. 00:24:40 and put them in sequential order,
534. 00:24:42 I think they can lead you down a path
535. 00:24:44 that maybe you haven't walked before.
536. 00:24:46 Like one of the things that I jotted down,
537. 00:24:48 which you mentioned that when predictive AI systems
538. 00:24:51 are deployed,
539. 00:24:51 the first people they harm are usually minorities
540. 00:24:54 and those in poverty.
541. 00:24:55 So what do you do with that?
542. 00:24:57 I'm not about to deploy a predictive AI system.
543. 00:25:00 So it's kind of just like a nice to know fact, I guess.
544. 00:25:03 Like be on your guard for this sort of thing.
545. 00:25:08 But the, yeah, the overall ideas that I think
546. 00:25:13 are gonna stick with me from this chapter
547. 00:25:15 are the definition of an algorithm,
548. 00:25:17 which is a set of steps or rules used to make a decision.
549. 00:25:21 And then they define predictive AI
550. 00:25:23 as models used for decision-making
551. 00:25:25 based on predictions about the future.
552. 00:25:27 The problem with that, if I were to try to summarize this
553. 00:25:33 based on the other notes that I've gathered here,
554. 00:25:38 is that AI tends to be good at making predictions
555. 00:25:42 if nothing else ever changes,
556. 00:25:46 but that's not the world that we live in.
557. 00:25:51 And so you should basically be skeptical
558. 00:25:54 of any of these claims and the question you should be asking
559. 00:25:58 is who was this AI tested on?
560. 00:26:02 I think later on they unpack that a little bit further
561. 00:26:04 and say that if you use it on just the data
562. 00:26:09 that it was tested on, it has limited value.
563. 00:26:13 I think they're talking about generative AI with that,
564. 00:26:15 but I think it applies to this predictive AI as well.
565. 00:26:19 You have to recognize the biases that are inherent
566. 00:26:21 from the training data.
567. 00:26:25 - I think that's, we're gonna highlight
568. 00:26:27 probably multiple times throughout this book.
569. 00:26:30 They said good things, I can't place it in the book at all.
570. 00:26:34 Like normally I can make like this little map that says,
571. 00:26:36 oh, it was this chapter, then they made this argument
572. 00:26:39 and they moved on from there.
573. 00:26:40 For this one, I have these ideas in my head.
574. 00:26:44 I was like, oh, that was a good point.
575. 00:26:45 I like that point.
576. 00:26:45 But if I didn't have my notes right in front of me,
577. 00:26:47 I have no idea where it was talked about in the book.
578. 00:26:49 Like it just doesn't flow.
579. 00:26:52 - Yeah, and that's why I made the comment about how
580. 00:26:54 I guess I just don't like academic books
581. 00:26:56 because I feel like every sentence is useful,
582. 00:27:00 but not exciting.
583. 00:27:01 And they're just one right after the other.
584. 00:27:03 And you get to the end of them and you're like,
585. 00:27:05 I probably heard some good things, but I don't care.
586. 00:27:07 Whereas if you're writing a book to sell a bunch of copies,
587. 00:27:11 not to plant your flag in the ground on the ideas themselves,
588. 00:27:15 which I totally get, that's a valid use case
589. 00:27:16 for writing a book, building a sub stack.
590. 00:27:19 We want to be known as the AI guys, right?
591. 00:27:23 You can do that.
592. 00:27:24 But I think the other approach to this,
593. 00:27:27 if you're trying to write a standard nonfiction book,
594. 00:27:30 you're working with an editor,
595. 00:27:31 it's like, how can we vary the sentence structure?
596. 00:27:33 How can we set things up?
597. 00:27:35 And we got this one banger sentence
598. 00:27:37 that's really just gonna nail it home.
599. 00:27:39 And this is the thing that people are gonna remember forever.
600. 00:27:42 They're just kind of like, yeah,
601. 00:27:42 the information's all there if you want it.
602. 00:27:44 - Yeah. (laughs)
603. 00:27:46 I wanna make, you know, one of the things that stood out to me
604. 00:27:49 from the five failures of predictive AI
605. 00:27:52 was the hiring one, where they talked about the bookshelves
606. 00:27:55 in the background, or like the random stuff
607. 00:27:57 that would make the person score high.
608. 00:28:00 And I, like this just hurts my heart.
609. 00:28:01 And it's probably because I work with students all the time
610. 00:28:04 who are trying to go out there,
611. 00:28:05 they're trying to get jobs.
612. 00:28:06 And in getting the callback from a human
613. 00:28:10 is the absolute hardest part of that process.
614. 00:28:13 And they all struggle with it.
615. 00:28:14 It doesn't matter how good they are.
616. 00:28:16 It doesn't matter what their GPA is.
617. 00:28:17 It doesn't matter anything like that.
618. 00:28:18 Getting that callback, because I'm seeing more and more of this.
619. 00:28:22 They submit their stuff.
620. 00:28:23 It goes through an automated process.
621. 00:28:24 They do a video interview that now I know,
622. 00:28:27 I didn't actually realize this at first,
623. 00:28:29 but now it goes through an automated screening.
624. 00:28:31 And there's, you know, not causation, but correlation.
625. 00:28:35 It's like, oh, the people we score well
626. 00:28:37 tend to all have beautiful backgrounds,
627. 00:28:39 or their walls are white, or whatever it is.
628. 00:28:41 And it's like, that has nothing to do with the candidate.
629. 00:28:43 Absolutely, like they were, they were traveling.
630. 00:28:46 So they had to film in their hotel room.
631. 00:28:47 Or, you know, they were visiting family
632. 00:28:49 and they're in a weird space.
633. 00:28:52 But that will trigger the automated systems.
634. 00:28:55 And then I think what's scary about this is,
635. 00:28:57 there's no human intervention,
636. 00:28:59 or it's unlikely that there's human intervention
637. 00:29:02 until something important enough arises
638. 00:29:05 that requires human intervention to come in.
639. 00:29:07 And I was like, man, we're just missing out
640. 00:29:10 so much on important things,
641. 00:29:12 because we're kicking this stuff to predictive AI
642. 00:29:15 into these models and these algorithms.
643. 00:29:17 And that made me, made me hurt a little bit.
644. 00:29:20 I agree.
645. 00:29:21 Well, they're based off of the data
646. 00:29:25 that the humans treat them.
647. 00:29:27 And so it's not a surprise to me
648. 00:29:30 that the biases come over as well.
649. 00:29:34 Yep, but.
650. 00:29:36 Okay, so now we go into chapter three.
651. 00:29:38 And chapter three is why can't AI predict the future?
652. 00:29:42 I honestly, Mike, I don't remember
653. 00:29:46 much of this chapter.
654. 00:29:47 Meaning like it was not something
655. 00:29:50 that really, really stood out to me in terms of,
656. 00:29:54 we produce a lot of things.
657. 00:29:56 They went on this tirade about social media.
658. 00:29:58 This is kind of where they start
659. 00:29:59 with the social media stuff.
660. 00:30:01 And then they start to get into that.
661. 00:30:05 There was the Fragile Families Challenge.
662. 00:30:08 You know, a notable large scale study
663. 00:30:10 that tried to predict children's outcomes using AI
664. 00:30:12 and lots of data.
665. 00:30:13 There was this idea of data that came into here.
666. 00:30:16 I don't.
667. 00:30:17 Can we talk about that, that Fragile Families
668. 00:30:20 particularly?
669. 00:30:20 Absolutely, because nothing really jumped out at me.
670. 00:30:22 So I'm really glad you had something that jumped out.
671. 00:30:25 Yeah, so, well, it's interesting to me
672. 00:30:28 because my dad has told me stories
673. 00:30:34 of his experience growing up.
674. 00:30:37 He went to a Catholic grade school
675. 00:30:43 and this was back in the day.
676. 00:30:47 So he was left handed and that was considered wrong.
677. 00:30:54 So like they tried to force him to write with his right hand.
678. 00:30:57 Like that's how far back in the day, right?
679. 00:31:00 But I remember him telling me
680. 00:31:02 'cause he is a business owner.
681. 00:31:05 He has a master's in assessment.
682. 00:31:07 You know, he's successful by any stretch of the imagination.
683. 00:31:12 And he told me that when he was in middle school
684. 00:31:17 in high school, he didn't really care about school.
685. 00:31:20 It wasn't until he was starting to think about
686. 00:31:23 what he was gonna do after school
687. 00:31:25 that he actually started caring.
688. 00:31:27 And there was this thing that back in the day
689. 00:31:30 called a predictor, which was essentially the range
690. 00:31:35 for the highest grade that you were able to get
691. 00:31:39 based on your previous performance.
692. 00:31:42 And so going into his senior year,
693. 00:31:44 he's like, I'm gonna try
694. 00:31:45 because I really wanna get into Marquette University
695. 00:31:48 in Milwaukee and he's acing all of his classes.
696. 00:31:52 But they're like, well, your predictor says
697. 00:31:53 you can't get above whatever.
698. 00:31:54 So that's the grade you're gonna get.
699. 00:31:56 And he's like, what?
700. 00:31:59 So he took the long road.
701. 00:32:03 He went to UWA first and crushed it there
702. 00:32:06 and then got into Marquette.
703. 00:32:08 But I remember that.
704. 00:32:10 And I remember him telling me those stories.
705. 00:32:11 I mean, me feeling like that is so incredibly wrong.
706. 00:32:16 And we're basically a repeating history.
707. 00:32:18 It sounds like with this predictive AI.
708. 00:32:21 It's like if my dad were to read this,
709. 00:32:23 he would probably get visibly upset
710. 00:32:26 and like throw the book out the window.
711. 00:32:28 He was like, what is wrong with you people?
712. 00:32:30 It's worrying, isn't it?
713. 00:32:32 I mean, when we came through school,
714. 00:32:33 'cause you and I aren't too far apart in age,
715. 00:32:36 do you remember that there would be like the honors track
716. 00:32:38 and then there'd be like the normal track
717. 00:32:40 and then there'd be the lower track
718. 00:32:42 and you'd get tracked in like seventh grade
719. 00:32:45 or whatever it was, junior high middle school.
720. 00:32:48 And it was really hard, if not impossible, to jump tracks.
721. 00:32:51 Once you were on a track, that was what you were gonna do
722. 00:32:53 and you were gonna cap out at whatever math course
723. 00:32:55 and cap out at whatever English course.
724. 00:32:57 And at the time in high school,
725. 00:32:59 I never really thought about it.
726. 00:33:00 I was like, okay, whatever.
727. 00:33:01 That's what they told me to do kind of a thing.
728. 00:33:03 But now I think about it from an educational standpoint
729. 00:33:06 and I'm like, what a terrible,
730. 00:33:08 what a terrible, terrible way to do that
731. 00:33:10 because if you just have a really bad sixth and seventh grade
732. 00:33:13 year or something happens in your family
733. 00:33:16 and it's just like really, you're really struggling socially
734. 00:33:19 or emotional or whatever it is, you get tracked
735. 00:33:22 and now essentially the next 10 years of your life
736. 00:33:25 are all figured out.
737. 00:33:26 I don't know, it just really rubs me the wrong way
738. 00:33:29 and this brought up that, like you're saying
739. 00:33:32 and I'm like, I wonder how we're doing this now
740. 00:33:34 with predictive AI.
741. 00:33:35 Like how many times have I been placed into a bucket
742. 00:33:38 by some predictive algorithm and not even known it
743. 00:33:41 and had no clue that I'm capped at my ability
744. 00:33:45 to engage with your company or your organization
745. 00:33:47 or whatever it is because I was automatically labeled
746. 00:33:51 and put in some bucket.
747. 00:33:53 It's wild.
748. 00:33:55 - Yeah, history repeats itself.
749. 00:33:58 But yeah, that's the big thing I had from this chapter.
750. 00:34:01 I don't know that I really got a solid answer
751. 00:34:04 for why AI can't predict the future.
752. 00:34:09 - I would think the closest thing I got to it was,
753. 00:34:13 where was it?
754. 00:34:14 Basically it was about data and it was the fact that,
755. 00:34:19 oh man, I'm trying to find it.
756. 00:34:21 It was the fact that basically it's only good
757. 00:34:23 on what it's trained on and then--
758. 00:34:25 - Sure.
759. 00:34:26 Man, there's another one here.
760. 00:34:29 I've lost it though.
761. 00:34:31 - Well, there's another thing from here
762. 00:34:33 which I think is important based on one of the fundamental
763. 00:34:36 reasons why AI can't predict the future.
764. 00:34:38 It's not the reason but it's one of the things
765. 00:34:40 that goes into this and I've actually heard some people
766. 00:34:43 that I look up to mention this too,
767. 00:34:45 that luck plays a larger role in success
768. 00:34:47 than most people who are successful would care to admit.
769. 00:34:51 And the point that they make in this book
770. 00:34:53 is that luck plays a larger role in success
771. 00:34:55 than it does in failure.
772. 00:34:57 So you can do everything right
773. 00:35:01 and you can still fail.
774. 00:35:03 So if luck is a factor, you really can't build a model
775. 00:35:10 successfully off of that, in the aggregate,
776. 00:35:14 I guess you can because you can insert a certain number,
777. 00:35:17 a certain amount of randomness to the system.
778. 00:35:19 But when you're trying to determine
779. 00:35:21 if this person is gonna be able to do something,
780. 00:35:25 that's where all bets are off, I think.
781. 00:35:28 - Yeah, you made me think of it.
782. 00:35:30 - It was the idea of agency.
783. 00:35:32 Luck agency, this idea of we actually have agency as humans
784. 00:35:35 and we are not completely predictive or predictable,
785. 00:35:40 I guess if you will, I mean, isn't that the premise of,
786. 00:35:44 have you seen the last mission in possible movies?
787. 00:35:47 Like that's kind of the whole premise
788. 00:35:49 of the last mission in possible movies
789. 00:35:50 is we aren't completely predictable humans
790. 00:35:52 but I haven't seen the very last one.
791. 00:35:55 - I predict they will continue
792. 00:35:56 to make mission in possible movies.
793. 00:35:57 - As long as they keep making money,
794. 00:35:59 as long as they keep predicting they'll make money, right?
795. 00:36:02 - Yeah.
796. 00:36:03 - Are you ready for chapter four?
797. 00:36:05 - Yeah, this one I'm excited to talk about.
798. 00:36:06 - Me too, that's awesome that you and I
799. 00:36:09 are on the same page, maybe it's bad for the show
800. 00:36:11 but it's awesome that we're on the same page.
801. 00:36:13 - Chapter four is the long road to generative AI.
802. 00:36:16 So I'm a big fan of thinking about generative AI.
803. 00:36:19 I like generative AI.
804. 00:36:21 I wanna use it more and more every day.
805. 00:36:24 But they basically, for as hard as they go
806. 00:36:28 against predictive AI,
807. 00:36:32 they seem to be for generative AI
808. 00:36:34 and they see the value of it
809. 00:36:36 but they kind of temper with like,
810. 00:36:39 okay, well, how did it work and where is it coming?
811. 00:36:42 Now they still will throw out ideas of,
812. 00:36:44 we need to be careful of, we need to,
813. 00:36:46 cautious on how we're getting the data to train it.
814. 00:36:49 So they talk about the fact that there's the common crawl
815. 00:36:51 and scraping the entire internet and stealing.
816. 00:36:54 Basically everybody's thoughts and ideas and words
817. 00:36:57 and then regurgitating those.
818. 00:36:59 They talk about how there are like organizations
819. 00:37:05 or advocacy groups that are trying to think about these things.
820. 00:37:09 Is this where they talked about,
821. 00:37:12 man, there was the letter, Mike.
822. 00:37:14 The big letter they put out to all the AI people
823. 00:37:18 and they had them sign it that basically it was like,
824. 00:37:19 we need to slow down, we need to really pump the brakes
825. 00:37:22 and it was at the chapter.
826. 00:37:23 No, that was the general AI, the existential threat one.
827. 00:37:28 Okay.
828. 00:37:29 So that's next chapter.
829. 00:37:30 But this one's all about basically just using it
830. 00:37:32 for image generation, for text generation
831. 00:37:35 and boy, do I have thoughts on this.
832. 00:37:38 Let me begin with a story.
833. 00:37:41 So a long time ago we were working
834. 00:37:43 with a digital marketing agency.
835. 00:37:45 They were helping us with some blog posts.
836. 00:37:47 They put together these blog posts,
837. 00:37:49 published them on our site
838. 00:37:51 and they used stock photos when they put together the blog posts.
839. 00:37:57 We stopped working with the agency
840. 00:37:59 and I think like five, six years later,
841. 00:38:02 we're getting emails from Getty images saying,
842. 00:38:09 you need to take this down right now
843. 00:38:11 and we're suing you for $25,000 for each one of these posts.
844. 00:38:15 - Oh wow.
845. 00:38:16 - Unless you can provide the license
846. 00:38:18 for these images that you used.
847. 00:38:20 And we're like, well, the agency did it.
848. 00:38:22 Let me reach out to them.
849. 00:38:23 They had an intern or something working on it.
850. 00:38:26 So they didn't have the receipts for the files
851. 00:38:29 that they had.
852. 00:38:30 So we were on the hook for it
853. 00:38:32 because we couldn't produce the license for the stock photo
854. 00:38:34 that we paid somebody who did it legit,
855. 00:38:37 but didn't have the receipt trails.
856. 00:38:40 So not only is this extremely expensive
857. 00:38:42 from a financial perspective,
858. 00:38:44 but it's also a complete waste of time.
859. 00:38:46 - Yeah.
860. 00:38:48 There are literally companies, lawyers, firms, whatever.
861. 00:38:54 All they do is they scrape the web
862. 00:38:55 that find these copyrighted images
863. 00:38:57 and they go after people who maybe have used this incorrectly.
864. 00:39:01 And in our case, we did it correctly,
865. 00:39:03 but we didn't have the proof to show
866. 00:39:06 that we had done it correctly.
867. 00:39:08 So I guess, you know, don't work with digital marketing agencies.
868. 00:39:11 Is that the solution?
869. 00:39:12 Like no.
870. 00:39:13 - Yeah.
871. 00:39:14 - So I will be the first one to applaud
872. 00:39:17 the death of the stock photo industry.
873. 00:39:20 - Okay.
874. 00:39:21 - I apologize to every artist, every photographer,
875. 00:39:26 every videographer that that's how they make their living,
876. 00:39:29 but I do not care at all for any
877. 00:39:33 of these stock photo video sites.
878. 00:39:35 They are a bunch of parasites, in my opinion,
879. 00:39:38 and they should be gotten rid of.
880. 00:39:41 The quicker that AI can do this, the better.
881. 00:39:44 Now that's separate from the argument
882. 00:39:46 about they're trained on all this copy-related material.
883. 00:39:51 And so as a content creator myself,
884. 00:39:54 I understand that side of it too.
885. 00:39:57 I don't want people slurping up everything
886. 00:40:01 that I've ever written,
887. 00:40:01 everything that I've ever made,
888. 00:40:03 and just creating new stuff in my style.
889. 00:40:07 Looks like me, sounds like me, you know,
890. 00:40:10 and then using that to make money,
891. 00:40:13 which is essentially what these large language models
892. 00:40:16 are doing, that's trained on all this copy-related data.
893. 00:40:19 So here's where I land on this.
894. 00:40:22 I do not think this is a battle that should be resolved
895. 00:40:29 with the AI companies.
896. 00:40:31 I don't think it should be resolved
897. 00:40:32 with the independent creators.
898. 00:40:34 I think the place to get it resolved
899. 00:40:37 where it actually is gonna make a difference
900. 00:40:39 is in the courts.
901. 00:40:41 And the thing that sucks about this
902. 00:40:43 is how slow that system is
903. 00:40:46 to what is happening right now.
904. 00:40:49 Like New York Times, I know, was suing OpenAI,
905. 00:40:53 I think back in the day,
906. 00:40:54 because they had slurped up a bunch of articles
907. 00:40:56 that were behind the paywall.
908. 00:40:58 Like, if we could just resolve that issue
909. 00:41:01 in a matter of weeks instead of years,
910. 00:41:04 then that sets a precedent.
911. 00:41:05 And then all these other models
912. 00:41:09 aren't just training on everything that they can find.
913. 00:41:11 At this point, it's almost like too late.
914. 00:41:15 It's all in there anyways.
915. 00:41:16 You can't retroactively go back.
916. 00:41:19 So it sucks.
917. 00:41:21 But also that opens up another question is like,
918. 00:41:23 where do we go from here?
919. 00:41:25 You know, if you're an independent creator,
920. 00:41:27 I think you have to just accept
921. 00:41:29 that this is the world that we live in.
922. 00:41:31 The AI models are going to consume all of my stuff.
923. 00:41:37 And I can get mad at that and be like,
924. 00:41:39 no, I want this behind a paywall.
925. 00:41:41 Or I can actually think about,
926. 00:41:44 you know, how could this actually be good for me?
927. 00:41:46 And I think there are benefits to this.
928. 00:41:48 I think it's kind of coming around
929. 00:41:51 where a lot of the AI models now will have the citations
930. 00:41:54 at the bottom of the thing that they spit back
931. 00:41:57 because people don't just trust it
932. 00:41:58 when it gives them a blanket answer.
933. 00:42:00 They want to see where AI got this.
934. 00:42:02 You know, I feel like that's kind of
935. 00:42:03 the natural evolution of this process.
936. 00:42:05 And so that's SEO 2.0 in a way.
937. 00:42:09 You know, people are going to go look at, look at my stuff
938. 00:42:11 that way and it's not the same.
939. 00:42:13 And I can get upset because it's not the same
940. 00:42:15 or I could just be like, oh, this is the way it is.
941. 00:42:16 You know, what are the advantages of doing things this way?
942. 00:42:19 'Cause I think there are some advantages to it as well.
943. 00:42:23 But, and rant.
944. 00:42:26 - I can't disagree with you on anything you said.
945. 00:42:29 The thing that I think is really interesting about this
946. 00:42:31 is a lot of these companies are like,
947. 00:42:33 we're just going to do it because nobody's governing it
948. 00:42:38 right now and nobody's regulating it right now.
949. 00:42:39 And we'll deal with the court stuff on the back end.
950. 00:42:42 Oh, but by the way, we made so much progress
951. 00:42:46 and we've become the, you know, the household name
952. 00:42:49 and we've become the brand that everybody goes to
953. 00:42:51 that's like, yeah, we'll resolve that all later
954. 00:42:53 down the road because what we're trying to do
955. 00:42:54 is we're trying to be first to market right now.
956. 00:42:56 We're trying to be the place that everybody goes.
957. 00:42:59 And I think that's just the nature of kind of the,
958. 00:43:02 you know, they say it's about the Wild West.
959. 00:43:04 That's the nature of the Wild West of AI right now
960. 00:43:06 is in less, yeah, unless, and they have a whole chapter
961. 00:43:10 on this which Mike and I are going to get to later.
962. 00:43:13 Unless the regulators come in and try to get faster than them,
963. 00:43:16 but I don't even know if they can get faster than them
964. 00:43:18 with the way the system works and the way the courts work
965. 00:43:21 and all of that, I don't think you can do that.
966. 00:43:25 I think about this too.
967. 00:43:26 You talked about it from your creator business.
968. 00:43:29 I think about it from an education business as well
969. 00:43:31 or for the education side as well.
970. 00:43:33 I'm up against the AI tool, the LLM tool
971. 00:43:38 that's in every kid's pocket when they walk into my classroom.
972. 00:43:42 So instead of them asking me, the first place they'll go
973. 00:43:45 is they'll go to their AI tool and ask the AI tool.
974. 00:43:48 And then after that, they'll ask me,
975. 00:43:51 but I can't tell you that I think
976. 00:43:52 that's a completely horrible, no good system.
977. 00:43:56 I think that, one, I think it over import,
978. 00:44:00 makes me overly important, right?
979. 00:44:01 Which I don't agree with that at all.
980. 00:44:03 It's more a matter of it just changes the way I teach.
981. 00:44:06 Like I have to teach in a different way
982. 00:44:07 and I have to present information in a different way
983. 00:44:10 and I have to kind of lean into that and say,
984. 00:44:12 there's gonna be some of this stuff
985. 00:44:13 that you're gonna go to the AI tool
986. 00:44:15 and learn it through there because it's more efficient
987. 00:44:17 and it's faster and that's good.
988. 00:44:18 And then there's gonna be some of the stuff
989. 00:44:19 you're gonna engage with me, engage with me with.
990. 00:44:21 Where I get, the one that actually makes me
991. 00:44:24 the most nervous, I think is the audio video.
992. 00:44:26 And they talk about this towards the end of this is
993. 00:44:28 how quickly and readily and well
994. 00:44:32 these tools can replicate your voice
995. 00:44:35 and then how quickly and well they can actually physically
996. 00:44:39 make it look like you are doing something that you didn't do.
997. 00:44:44 And I think that is a very scary world
998. 00:44:49 until we figure out detection
999. 00:44:52 and then we start playing the dance game of,
1000. 00:44:54 well, we figured out better ways to detect it.
1001. 00:44:56 Oh, they figured out better ways to trick that.
1002. 00:44:58 Oh, we figured out better ways to detect it.
1003. 00:44:59 They figured out better ways to trick it.
1004. 00:45:00 And once we get into that dance game,
1005. 00:45:02 I actually think things level out a little bit.
1006. 00:45:05 'Cause you're really only gonna have the fringe,
1007. 00:45:07 the front-end of that be an issue.
1008. 00:45:09 But right now, I think it can be an issue for anybody.
1009. 00:45:12 We're still in that world where,
1010. 00:45:14 somebody could make a video of you saying anything crazy
1011. 00:45:19 and there's really very little we can do to detect that
1012. 00:45:22 other than you coming out and being like, that wasn't me.
1013. 00:45:24 It's completely AI-generated.
1014. 00:45:26 And that's really the recourse you have at this point.
1015. 00:45:28 And that's a scary world to me right now.
1016. 00:45:31 - Dude, people are doing that intentionally.
1017. 00:45:33 I was at Crafty Commerce this year.
1018. 00:45:34 One of the sponsors was a company called Delphi,
1019. 00:45:36 where you literally make an AI version of yourself.
1020. 00:45:40 They demoed on stage somebody calling and getting
1021. 00:45:43 workout advice from an Arnold Schwarzenegger AI voice bot.
1022. 00:45:51 So it is scary and they talk a little bit
1023. 00:45:55 about the deep fake stuff and how that can be bad.
1024. 00:45:59 But I also think that there are people
1025. 00:46:00 who are just leaning into this technology
1026. 00:46:02 be like, yeah, I don't growl trick people.
1027. 00:46:04 They'll think it's me.
1028. 00:46:04 Not that they actually think it's you, but I don't know.
1029. 00:46:10 The thing to remember about all of this technology,
1030. 00:46:12 I think, and we'll get into this I think
1031. 00:46:14 as we go further down the book is that these are tools
1032. 00:46:19 and they have no intrinsic moral value.
1033. 00:46:22 They are not good.
1034. 00:46:23 They are not bad.
1035. 00:46:24 They can be used for both good and bad.
1036. 00:46:28 - Do you think they are, so what I would say,
1037. 00:46:32 there's a book that we read in with my seniors every year,
1038. 00:46:36 but there's an inherent bias
1039. 00:46:38 and/or an inherent value built into things.
1040. 00:46:40 So we would say they are value-laden designs.
1041. 00:46:43 And value-laden designs, they're neither good nor bad.
1042. 00:46:46 You can't say they're all one thing or all the other,
1043. 00:46:48 but you can say that the way they were designed,
1044. 00:46:51 they lean more one way or they lean more another way.
1045. 00:46:54 So my question to you, Mike, put you in the spot.
1046. 00:46:56 Would you say it's value-laden in a way
1047. 00:46:58 that it leans a different way,
1048. 00:47:02 one of the ways or another?
1049. 00:47:04 - No, because fundamentally what it does
1050. 00:47:06 is it just tries to predict the next thing in the chain.
1051. 00:47:09 So it needs data to be trained on
1052. 00:47:12 and the data you can argue about
1053. 00:47:14 whether that was sourced ethically,
1054. 00:47:15 I would argue that it's not.
1055. 00:47:17 Maybe it wasn't illegal, but it certainly wasn't ethical.
1056. 00:47:20 - Agreed.
1057. 00:47:21 - But that data now that is in there,
1058. 00:47:23 the LLMs, the way that they function,
1059. 00:47:25 they're just going to try and fill in the next blank.
1060. 00:47:28 And it's going to use tokens to do that.
1061. 00:47:31 They mentioned a single token,
1062. 00:47:32 requires one quadrillion calculations,
1063. 00:47:36 which is a million trillion, I think, or a million billion?
1064. 00:47:41 I forget.
1065. 00:47:42 - I have to think hard about it.
1066. 00:47:44 - One, zero, zero, zero, zero, zero, zero, zero, zero, zero,
1067. 00:47:48 zero, zero, zero, zero calculations.
1068. 00:47:50 And that's why they say the long road to Generative AI
1069. 00:47:53 because those calculations,
1070. 00:47:55 that computers have been a long time coming
1071. 00:47:58 and they keep getting more powerful
1072. 00:47:59 and they can do more of these calculations.
1073. 00:48:01 And so this is kind of just like the next step
1074. 00:48:03 in the chain for that.
1075. 00:48:05 And the thing to recognize with these models
1076. 00:48:08 is that they are trained to produce
1077. 00:48:11 plausible text, not true statements.
1078. 00:48:13 So you can't get mad at an LLM
1079. 00:48:15 because it hallucinates, in my opinion.
1080. 00:48:17 That's part of the deal.
1081. 00:48:19 But hallucinations aren't all bad either.
1082. 00:48:21 It's kind of interesting because I literally just did
1083. 00:48:23 a talk at max-stock on AI.
1084. 00:48:26 (laughs)
1085. 00:48:27 And I talked about using it as a creative collaborator.
1086. 00:48:30 You know, when you're in the creative process,
1087. 00:48:32 you have a divergence of ideas
1088. 00:48:35 where you consider all the options
1089. 00:48:36 and then a convergence when you bring it back together.
1090. 00:48:39 During the divergence piece,
1091. 00:48:40 I mean, that's where hallucinations can be good.
1092. 00:48:42 You want to consider every off the wall possibility
1093. 00:48:45 that you possibly can.
1094. 00:48:47 But then at some point you have to be like,
1095. 00:48:48 okay, you know, I have what I need.
1096. 00:48:50 Now I'm gonna actually do something with it.
1097. 00:48:53 So yeah, I honestly believe that this doesn't lean
1098. 00:48:57 one way or the other.
1099. 00:48:58 It's just a technological tool.
1100. 00:48:59 Just like a computer is a technological tool.
1101. 00:49:01 And how you use it really is where the morality
1102. 00:49:06 of the technology comes from.
1103. 00:49:08 - I think it's well said.
1104. 00:49:10 Let's go on to five if you're ready.
1105. 00:49:14 - Yes, natural extension of what we're just talking about
1106. 00:49:17 in my opinion.
1107. 00:49:18 Is advanced AI and existential threat?
1108. 00:49:20 They set up the argument here that a lot of people,
1109. 00:49:23 so we get into the new term
1110. 00:49:25 of artificial general intelligence now.
1111. 00:49:28 So AGI, so they add more to it.
1112. 00:49:32 But that they say a lot of experts in the field
1113. 00:49:34 say that AGI is an imminent existential threat.
1114. 00:49:38 They then immediately, or not immediately,
1115. 00:49:40 but they unpack their argument that say,
1116. 00:49:42 we actually should be more worried
1117. 00:49:43 about what people will do with AI
1118. 00:49:46 than AI doing things on its own.
1119. 00:49:48 Like they're not as worried about AI
1120. 00:49:50 kind of running a muck and doing things on its own.
1121. 00:49:54 And they say that there are two options for safety.
1122. 00:49:58 There's a line AI with humanity's interests
1123. 00:50:00 and then stop powerful AI from being built at all.
1124. 00:50:03 And they basically say, and neither of these
1125. 00:50:05 are gonna work because we've already either
1126. 00:50:08 let the cat out of the bag or we already
1127. 00:50:10 haven't aligned the AI or whatever it is.
1128. 00:50:13 So it's just really funny to me like how,
1129. 00:50:17 here's what we should do.
1130. 00:50:18 And here's the way we should address this
1131. 00:50:19 and none of that's gonna work, so moving on.
1132. 00:50:22 And it's just like a weird feel for the book.
1133. 00:50:24 It's like very information-based,
1134. 00:50:28 not very story or not very my opinion kind of a thing.
1135. 00:50:34 I mean, I guess maybe they do give their opinion,
1136. 00:50:36 but they give it in a very factual way,
1137. 00:50:38 like in a very information-based way.
1138. 00:50:40 And it just reads very, very differently.
1139. 00:50:44 So they come in kind of at the very end of this
1140. 00:50:47 and they say, so how should we go about
1141. 00:50:49 defending against bad actors who wanna use it?
1142. 00:50:51 Well, we can defend against their specific threats
1143. 00:50:53 or we can strengthen democracy.
1144. 00:50:55 And like, I like all of this, these are all good ideas.
1145. 00:50:59 And maybe this is where we need to be
1146. 00:51:01 at this point in the discussion.
1147. 00:51:02 In September of 2024, when they're writing this,
1148. 00:51:05 but I kind of feel like we're just painting
1149. 00:51:07 like gigantic broad strokes
1150. 00:51:09 and we're not getting to like things
1151. 00:51:11 that can actually help people.
1152. 00:51:13 So what should we do?
1153. 00:51:14 We should defend against specific threats.
1154. 00:51:16 Okay, great, great.
1155. 00:51:18 Like cyber security, great.
1156. 00:51:19 Okay, fantastic.
1157. 00:51:21 And then what should we do?
1158. 00:51:21 We should strengthen democracy.
1159. 00:51:23 Oh yeah, 'cause that's easy, right?
1160. 00:51:24 Like let's just strengthen democracy.
1161. 00:51:26 Here we go.
1162. 00:51:27 So I mean, I know I'm being like kind of tongue in cheek
1163. 00:51:29 as I talk about it, but I just get this feel for it
1164. 00:51:32 that it's the point of this book isn't to,
1165. 00:51:37 or it's not as much to solve problems.
1166. 00:51:39 It's more a matter of to call out something
1167. 00:51:42 that I think these authors don't think
1168. 00:51:44 is being discussed enough in the industry
1169. 00:51:47 and in the policy section surrounding the industry.
1170. 00:51:50 So it's just a very different feel
1171. 00:51:51 and a very different read for the book,
1172. 00:51:53 at least for me, anyhow.
1173. 00:51:56 - Yeah, so I had a maybe different perspective
1174. 00:52:00 on this chapter, but I love this chapter, to be honest.
1175. 00:52:03 So artificial general intelligence,
1176. 00:52:06 they define that as AI that can perform most
1177. 00:52:09 or all economically relevant tasks
1178. 00:52:10 as effectively as any human.
1179. 00:52:12 And at that point, that's where they lay out the argument
1180. 00:52:17 that most people have this Terminator style picture of AI
1181. 00:52:21 once it gets beyond the capabilities of humans,
1182. 00:52:24 but that's not necessarily the case.
1183. 00:52:26 I kind of feel like that's irrelevant
1184. 00:52:29 to the discussion, to be honest.
1185. 00:52:31 And they kind of mentioned how that is probably
1186. 00:52:35 a very, very tiny chance of actually happening.
1187. 00:52:39 But you mentioned this when you were,
1188. 00:52:42 at the beginning of your introduction to this chapter,
1189. 00:52:44 we should be far more worried about what humans will do
1190. 00:52:46 with AI than what AI will do on its own.
1191. 00:52:48 Going back to the previous chapter's discussion
1192. 00:52:50 on like the morality of the tool.
1193. 00:52:53 Okay, and then they say that we should assume
1194. 00:52:55 that bad actors will have access to state of the art AI.
1195. 00:52:59 I read a book back in the day called Tools and Weapons
1196. 00:53:03 by Brad Smith, who was the president of Microsoft
1197. 00:53:06 forward by Bill Gates.
1198. 00:53:09 And the subtitle of this book is The Promise
1199. 00:53:11 and the Peril of the Digital Age.
1200. 00:53:12 It was written in 2019.
1201. 00:53:14 It came on my radar because Brad Smith
1202. 00:53:17 was actually in Appleton, the town that I live.
1203. 00:53:20 So he went to Appleton, West, and Microsoft
1204. 00:53:22 is actually invested in a local up in Green Bay,
1205. 00:53:27 what do they call it?
1206. 00:53:28 It's like an accelerator.
1207. 00:53:32 They call it Title Town Tech, basically,
1208. 00:53:34 where they are investing in helping get technology companies
1209. 00:53:40 funded.
1210. 00:53:41 - Yeah, like an incubator accelerator kind of thing.
1211. 00:53:43 - Yes, yes, exactly, in Wisconsin, right?
1212. 00:53:47 So this book, 2019.
1213. 00:53:50 So this was before a lot of this AI stuff.
1214. 00:53:53 But I remember he was in Appleton.
1215. 00:53:55 He was giving a talk at the local college university.
1216. 00:54:00 So I went and attended it.
1217. 00:54:02 And there were a lot of questions in that session
1218. 00:54:08 about things that bad actors were trying to do
1219. 00:54:13 with technology because he's at Microsoft.
1220. 00:54:16 They're trying to build all these crazy technological tools.
1221. 00:54:19 So he's sharing stories about how,
1222. 00:54:22 what's the term that he used dictator governments
1223. 00:54:26 or something like that?
1224. 00:54:28 Examples of surveillance states and stuff like that.
1225. 00:54:33 And he's like, they reach out to us all the time.
1226. 00:54:35 They want us to unlock these things for him.
1227. 00:54:37 And we're just like, no, we're not gonna do that.
1228. 00:54:38 He's like, this is a very fine line that you are trying
1229. 00:54:44 to walk here because on the one hand,
1230. 00:54:46 like the technology can save lives,
1231. 00:54:47 it can improve people's existence.
1232. 00:54:50 It can do a lot of good for humanity,
1233. 00:54:53 but then you have to recognize that there are always people
1234. 00:54:54 who are going to use it for various purposes.
1235. 00:54:59 And I feel like AI is simply that, you know?
1236. 00:55:02 And yeah, maybe at some point it gained sentience
1237. 00:55:06 and it didn't slaves humanity.
1238. 00:55:09 I mean, that sounds ridiculous, but is it a possibility?
1239. 00:55:14 Like no one really knows whether that could be or not,
1240. 00:55:19 but it's also not worth worrying about,
1241. 00:55:22 what you should be worrying about is how people are using
1242. 00:55:25 the tools that are available now to steal your credit card
1243. 00:55:29 numbers and ruin your life right now
1244. 00:55:31 because they're actively trying to do that
1245. 00:55:34 and figuring out what you can do to protect yourself
1246. 00:55:37 in the digital culture that we live in.
1247. 00:55:41 So yeah, if I were to just answer this,
1248. 00:55:44 is advanced AI an existential threat?
1249. 00:55:46 Question mark, I don't really care.
1250. 00:55:49 Not relevant, at least not right now.
1251. 00:55:51 So I listened to a couple of different things,
1252. 00:55:54 I was probably a year and a half ago,
1253. 00:55:55 maybe two years ago around this.
1254. 00:55:58 And there were people that were very much the way
1255. 00:56:00 they described them that, you know,
1256. 00:56:02 saying AI is an existential threat.
1257. 00:56:06 And in the talk, I mean, it was a podcast
1258. 00:56:09 and there were three, I don't know,
1259. 00:56:11 three hours into the podcast or whatever,
1260. 00:56:12 two and a half hours, three hours in the podcast.
1261. 00:56:14 And they were like, okay, there were like four or five things
1262. 00:56:17 we can't do with AI.
1263. 00:56:18 One was give it access to internet.
1264. 00:56:20 Two was teach it had a program and they listened off
1265. 00:56:24 and they were like, and we've done them all.
1266. 00:56:26 So yeah, and like that was kind of the end
1267. 00:56:30 of the conversation, it was like,
1268. 00:56:31 so we've already let the cat out of the bag.
1269. 00:56:34 I mean, what are we gonna do now?
1270. 00:56:35 But I agree with you in the sense of,
1271. 00:56:39 I think we need to be thinking,
1272. 00:56:41 well, so I agree with you in the times of,
1273. 00:56:43 we need to focus on what's happening right now.
1274. 00:56:45 But I also do think we need to be thinking about like,
1275. 00:56:47 what is the general trajectory of where we're going?
1276. 00:56:52 And do we like that general trajectory?
1277. 00:56:54 I don't necessarily think we can fully define
1278. 00:56:57 where it's going, you know, those would be futurists,
1279. 00:56:59 people who think about this stuff professionally.
1280. 00:57:01 But like, I think we just need to make sure
1281. 00:57:04 we know where it's pointing or where we think it's pointing
1282. 00:57:06 and be like, are we okay with that?
1283. 00:57:07 Are we okay with it pointing that way?
1284. 00:57:08 Because we have no idea, Mike, if in like a year,
1285. 00:57:12 some other companies are gonna come out
1286. 00:57:14 and they're gonna drop another bombshell.
1287. 00:57:16 And like, I mean, it's gonna take a hard left
1288. 00:57:18 and we're gonna be like, oh, no,
1289. 00:57:20 we weren't ready for that at all.
1290. 00:57:22 And that's-
1291. 00:57:23 - Correct, correct.
1292. 00:57:25 So they do mention at one point,
1293. 00:57:27 I think it was in this chapter about how like,
1294. 00:57:29 there was push for pausing the advancement
1295. 00:57:31 of these large language models.
1296. 00:57:34 I, okay, so is the cat out of the bag?
1297. 00:57:38 Maybe, maybe not.
1298. 00:57:40 I think we are still very, very early stages
1299. 00:57:44 when it comes to what actually is AI
1300. 00:57:47 and how does it get implemented into our lives.
1301. 00:57:50 With AI, just like any technology,
1302. 00:57:54 there's always a large rate of growth
1303. 00:57:59 at the beginning.
1304. 00:58:00 And I wish I had the link to the podcast
1305. 00:58:03 that I had listened to that shared this,
1306. 00:58:05 but basically that six month pause has come and gone.
1307. 00:58:11 And there's no GPT-5.
1308. 00:58:14 Do you know why there's no GPT-5?
1309. 00:58:15 It's because they're not getting the same rate of advancement.
1310. 00:58:18 You can't just throw more compute power at this
1311. 00:58:22 and get the same returns.
1312. 00:58:24 So if they were to release GPT-5,
1313. 00:58:26 then it's, people are gonna be like, meh.
1314. 00:58:30 It's not that much better than the previous one.
1315. 00:58:33 And so I think with large language models
1316. 00:58:36 is probably the tail end of the curve
1317. 00:58:38 and things are flattening out a little bit.
1318. 00:58:41 And there will be something else that will come after that.
1319. 00:58:44 But that's just, that's the way technology advances.
1320. 00:58:49 It's not surprising to me.
1321. 00:58:52 And I kind of think you kind of don't know
1322. 00:58:54 where this is going to end up.
1323. 00:58:56 I think it's foolish to say,
1324. 00:59:02 how do I say this?
1325. 00:59:02 I don't wanna say something that I'll regret.
1326. 00:59:04 I get it.
1327. 00:59:06 I personally though think it's a little bit foolish
1328. 00:59:09 to say, no, this is inherently bad
1329. 00:59:12 and we just shouldn't touch this with a 39 and a half foot pole.
1330. 00:59:16 And instead we should be thinking about
1331. 00:59:17 how do we actually do good with this?
1332. 00:59:20 There's probably specific things where it's like,
1333. 00:59:22 eh, no, that's where I would draw the line.
1334. 00:59:26 But I don't see that with AI yet.
1335. 00:59:29 I'm personally not comfortable with OpenAI
1336. 00:59:33 and the other company's goal of creating
1337. 00:59:36 this artificial general intelligence.
1338. 00:59:38 But I also think the fact that they wanna do that
1339. 00:59:40 is not make them inherently evil
1340. 00:59:44 and this is guaranteed doomsday.
1341. 00:59:49 I don't know, I feel like the larger question to be asked
1342. 00:59:51 and really something that society needs to ask.
1343. 00:59:54 So kind of like when I say, don't worry about it,
1344. 00:59:56 it's because me as an individual,
1345. 00:59:58 what am I really able to,
1346. 01:00:01 I cannot, Mike Schmitt's, stop the AI avalanche.
1347. 01:00:09 So I'm not gonna worry about that,
1348. 01:00:11 but I am going to ask how can I use the tools
1349. 01:00:15 that are available to me to do more
1350. 01:00:17 of what actually matters, to make a bigger impact,
1351. 01:00:20 positive impact in the lives of the people around me.
1352. 01:00:24 And I do think AI is useful for that.
1353. 01:00:28 - All right, you ready to go to social media, chapter six?
1354. 01:00:31 - Great segue into social media.
1355. 01:00:33 - Yeah.
1356. 01:00:34 - So chapter six is why can't AI fix social media?
1357. 01:00:38 I do not like the title of this chapter.
1358. 01:00:40 I understand why they did it,
1359. 01:00:42 but I don't like the title of the chapter because really,
1360. 01:00:44 I think this chapter is less about social media
1361. 01:00:47 and it's more about content moderation.
1362. 01:00:48 I think they should have called this
1363. 01:00:50 the content moderation chapter
1364. 01:00:52 because that's more accurate to what they're talking about.
1365. 01:00:54 Now, granted, where does most content moderation happen?
1366. 01:00:57 Probably on the internet right now,
1367. 01:00:58 it happens in social media
1368. 01:01:00 because there are currently humans
1369. 01:01:04 that are paid a very, very low wage
1370. 01:01:06 to sit there and look at content and flag it
1371. 01:01:08 for whatever it needs to be flagged as.
1372. 01:01:11 I personally think this is the worst job
1373. 01:01:16 or it was probably the worst job on the planet,
1374. 01:01:18 like one of the worst jobs on the planet.
1375. 01:01:19 I mean, the filth that is thrown out there
1376. 01:01:23 and just scours the internet and your job all day
1377. 01:01:25 is to identify and block and identify and shut down
1378. 01:01:28 and identify and shut down.
1379. 01:01:30 It's like you, it's akin to being the remote drone pilot
1380. 01:01:36 in the military and like dealing
1381. 01:01:38 with the psychological effects of that
1382. 01:01:39 just in a different scale, right?
1383. 01:01:41 Like, so I feel 100% for these people,
1384. 01:01:45 it hurts me to think about that.
1385. 01:01:48 Let's get to the idea of AI doing content moderation.
1386. 01:01:51 Basically, their whole premise of this chapter is
1387. 01:01:53 it can't do it.
1388. 01:01:55 One, does it even translate it correctly?
1389. 01:01:57 They're like, can it do that?
1390. 01:01:59 Translation's gotten better,
1391. 01:02:00 but can it translate it correctly?
1392. 01:02:01 Two, humans are really good
1393. 01:02:03 at getting around content moderation.
1394. 01:02:05 There's nuance. They throw in different like phrases
1395. 01:02:08 that say a certain thing without using the words
1396. 01:02:11 that get flagged automatically.
1397. 01:02:14 Like, so their whole point I think are around here
1398. 01:02:17 are how people evade this content.
1399. 01:02:18 They throw out ideas about how people evade it.
1400. 01:02:21 They give you examples of that.
1401. 01:02:23 They talk about like algo speak
1402. 01:02:25 where you figure out how to speak in these algorithms
1403. 01:02:27 that the algorithm then won't pick up
1404. 01:02:29 or these like pseudo algorithms, whatever it might be.
1405. 01:02:32 I, did you know of the Overton window
1406. 01:02:34 before you read this chapter?
1407. 01:02:36 Like, that was interesting to me.
1408. 01:02:37 I'd never heard of the Overton window.
1409. 01:02:38 So basically the Overton window is a horizontal scale
1410. 01:02:42 from right to left, basically arrows on either side
1411. 01:02:46 and it goes from unthinkable, radical, acceptable, popular.
1412. 01:02:49 And then on the other side of that scale
1413. 01:02:51 is the acceptable, radical, and unthinkable.
1414. 01:02:54 And I just never heard of that, right?
1415. 01:02:56 So that's a thing in the content moderation world.
1416. 01:02:59 They give a table of the seven reasons
1417. 01:03:01 why content moderation is hard.
1418. 01:03:03 I will spare you reading all of those.
1419. 01:03:05 I think it's really interesting though.
1420. 01:03:07 And then they tell you what, you know,
1421. 01:03:09 there's current limitations, there's inherent limitations,
1422. 01:03:11 whether it's cost prohibitive
1423. 01:03:13 or whether it's inherent to social media.
1424. 01:03:16 I think there's a lot of really interesting things
1425. 01:03:19 in this chapter, but like my overall idea is
1426. 01:03:23 I think it kind of comes out of nowhere.
1427. 01:03:25 Really, I just don't know like in the flow of everything
1428. 01:03:29 we just came from existential threat.
1429. 01:03:30 And now we go into content moderation.
1430. 01:03:32 And it's like, okay, cool, like this is good
1431. 01:03:37 and I'm happy that we're talking about this,
1432. 01:03:39 but where, where did it come from?
1433. 01:03:40 So that's just kind of my general take on it.
1434. 01:03:43 - Yeah, so I think I see how we get there.
1435. 01:03:48 I mean, one of the things that they mentioned earlier
1436. 01:03:51 about bolstering democracy, so one of the arguments
1437. 01:03:53 that you hear people make is all the election was rigged
1438. 01:03:56 because of all the Russian interference
1439. 01:03:57 and they actually share statistics that really was not a thing.
1440. 01:04:02 But people think it was.
1441. 01:04:03 So hard to argue with what people think in their heads
1442. 01:04:07 because the internet is a grocery store for facts
1443. 01:04:10 and you can find whatever facts you want
1444. 01:04:11 to support whatever crackpot beliefs you want to hold.
1445. 01:04:13 - Yes, you can.
1446. 01:04:14 Yes, you can.
1447. 01:04:15 - And I'm not saying that that one specifically,
1448. 01:04:16 although maybe, you know, it's been proven otherwise.
1449. 01:04:22 Anyways, so like I think it's interesting to go from,
1450. 01:04:27 is AI and existential threat to social media
1451. 01:04:31 because that's the argument, I think,
1452. 01:04:34 is like what appears on social media influences society.
1453. 01:04:38 And I think it influences individuals much more
1454. 01:04:42 than it does societies as a whole.
1455. 01:04:47 I guess you could say that that stuff in aggregate,
1456. 01:04:50 you know, has an effect, but I would be more concerned
1457. 01:04:54 about what the content that the algorithm is surfacing
1458. 01:04:59 to any individual is causing to that person
1459. 01:05:03 than I would be, you know, the election is actually being rigged
1460. 01:05:08 because of Russian troll farms,
1461. 01:05:11 the term that they used in this particular chapter.
1462. 01:05:14 So I think this is an interesting problem
1463. 01:05:18 without a good solution.
1464. 01:05:20 You know, you mentioned that it's been done by humans
1465. 01:05:22 and that job kind of sucks, and I think I would agree with that.
1466. 01:05:26 But also content moderation in general is sort of something
1467. 01:05:30 I feel like social networks do because they have to,
1468. 01:05:34 not because they actually care,
1469. 01:05:36 not because they want to.
1470. 01:05:38 It's in direct contrast to their business interests
1471. 01:05:43 because they share in here that content that is closer
1472. 01:05:46 to violating content moderation policies
1473. 01:05:49 is actually more likely to be engaged with.
1474. 01:05:51 So that creates this incentive to get it as close
1475. 01:05:55 to the edge as you can without it going over.
1476. 01:05:58 So if they were really concerned about this
1477. 01:06:01 the closer that it got to that point,
1478. 01:06:03 the more they would moderate it,
1479. 01:06:05 the more they would clamp down on it.
1480. 01:06:06 And the demand curve would not, you know,
1481. 01:06:09 trickle up at that point.
1482. 01:06:10 It would be flattened at that point.
1483. 01:06:12 It's like, ah, you're getting to a point
1484. 01:06:14 where this is dangerous.
1485. 01:06:15 Now, people in the US don't like that
1486. 01:06:19 because of the First Amendment.
1487. 01:06:21 And that's kind of the argument that people always have
1488. 01:06:24 is like, well, I could say whatever the heck I want.
1489. 01:06:27 I think with social media,
1490. 01:06:28 you should not be able to say whatever the heck you want.
1491. 01:06:30 - I agree.
1492. 01:06:31 - I think you want to go out into the streets
1493. 01:06:32 and yell in a megaphone,
1494. 01:06:34 go ahead and say whatever the heck you want.
1495. 01:06:37 But when you are using these tools
1496. 01:06:41 that give people platforms,
1497. 01:06:43 I think there does need to be some sort of,
1498. 01:06:46 some sort of moderation on this.
1499. 01:06:53 So yeah, I get why AI isn't able to fix this
1500. 01:06:58 because even if it was doing
1501. 01:07:00 what the jobs of the humans were doing,
1502. 01:07:03 it's not able to do it perfectly
1503. 01:07:06 and it's gonna have all of the biases
1504. 01:07:08 that are inherent with the things
1505. 01:07:09 that they're trying to moderate.
1506. 01:07:11 So the real problem with social media
1507. 01:07:14 is the algorithms, in my opinion.
1508. 01:07:16 - And the humans.
1509. 01:07:17 - I'm not, yeah, yeah.
1510. 01:07:19 I'm not a fan of algorithm-based social media.
1511. 01:07:23 Now, I need to temper that
1512. 01:07:25 because I have also said that YouTube
1513. 01:07:27 is the one quote unquote social media platform
1514. 01:07:29 that I like to engage on.
1515. 01:07:33 And that is because as a creator
1516. 01:07:35 who's trying to build a business,
1517. 01:07:37 I feel like YouTube is the one where
1518. 01:07:39 I can reach the people that I want to reach
1519. 01:07:40 and I can have a positive impact
1520. 01:07:42 if that's what I want to do with my videos,
1521. 01:07:44 which is in fact what I want to do with my videos.
1522. 01:07:46 And the algorithm is going to help the people
1523. 01:07:48 who want to see what I make.
1524. 01:07:50 It's gonna help me get in front of them.
1525. 01:07:52 It's gonna help us get connected.
1526. 01:07:53 And so it's gonna help my channel grow.
1527. 01:07:56 But in the back of my mind is this existential threat
1528. 01:08:01 for my YouTube channel.
1529. 01:08:03 They mentioned that you get three copyright strikes,
1530. 01:08:05 which is automated based on the AI,
1531. 01:08:09 then your YouTube channel gets banned.
1532. 01:08:11 And I know a lot of people who have succumbed to that,
1533. 01:08:13 not they fell into it, like it was their fault or anything,
1534. 01:08:16 but just for whatever reason, YouTube decided,
1535. 01:08:19 you use copyrighted material and they don't have...
1536. 01:08:23 There's no one at YouTube that they can reach out to
1537. 01:08:26 to argue with.
1538. 01:08:28 And so that's kind of scary.
1539. 01:08:32 But overall, recognize what these algorithms
1540. 01:08:36 are doing to you.
1541. 01:08:38 They're trying to feed you more and more extreme content
1542. 01:08:44 to get you more and more emotionally invested,
1543. 01:08:47 usually negatively, get you more upset
1544. 01:08:50 so that you continue to scroll,
1545. 01:08:52 so that you continue to engage with the platform.
1546. 01:08:54 You don't put the phone down and take the tech-free week
1547. 01:08:57 that Corey and I talked about in the pro show.
1548. 01:09:02 And there's a great documentary, The Social Dilemma,
1549. 01:09:05 which is older now, but I think speaks to this very clearly.
1550. 01:09:09 And I'm not one who would say,
1551. 01:09:12 just don't use social media,
1552. 01:09:14 but do use it intentionally and go into it with your eyes open.
1553. 01:09:18 These aren't benevolent dictators who are looking out for you
1554. 01:09:23 because they have a digital well-being feature
1555. 01:09:25 or they're investing all this money in content moderation.
1556. 01:09:28 No, they're doing that because it is literally
1557. 01:09:30 the bare minimum that they can do
1558. 01:09:34 to not be held responsible for what happens on their platform.
1559. 01:09:38 - Exactly, I'm gonna make a prediction here, Mike.
1560. 01:09:40 And I'm intrigued to see if this comes true.
1561. 01:09:44 Hold me accountable to it if your memory serves you so well.
1562. 01:09:48 I think AI and automated content generation
1563. 01:09:54 is going to destroy social media.
1564. 01:09:56 And I think it's going to push us into,
1565. 01:09:59 you're already starting to see it.
1566. 01:10:01 It's gonna push us more into communities
1567. 01:10:03 where there is more moderation,
1568. 01:10:05 but it's appropriate moderation
1569. 01:10:07 and we choose to be involved in that.
1570. 01:10:08 So for instance, I wanna be in the productivity space,
1571. 01:10:13 like I wanna learn about that,
1572. 01:10:14 where I used to go to all of the social media platforms
1573. 01:10:18 and I would follow the creators that I wanted to follow.
1574. 01:10:20 Well, now I joined the four or five creators
1575. 01:10:23 that I really wanna be part of their communities
1576. 01:10:26 and that's where I get those things.
1577. 01:10:28 And I don't worry about the broader completely,
1578. 01:10:31 well, not completely, but 90% automated social media stuff anymore.
1579. 01:10:36 It's just not, as you would say,
1580. 01:10:39 'cause you like to say this,
1581. 01:10:40 there's too much noise and not enough signal.
1582. 01:10:43 So I'm going to put myself in the place
1583. 01:10:45 where there's more signal and less noise.
1584. 01:10:47 I think we're gonna see that continue to increase
1585. 01:10:49 over the next five, 10 years.
1586. 01:10:52 I mean, I hope so, because I'm building a community.
1587. 01:10:56 But I think you've already seen it.
1588. 01:10:57 Like I think you've already seen people
1589. 01:10:58 who are willing to do that.
1590. 01:10:59 They're willing to join, invest,
1591. 01:11:03 whether it's invest in time, attention, money,
1592. 01:11:06 whatever it is.
1593. 01:11:07 They're willing to do that to get out of that environment
1594. 01:11:10 that's just so noisy and you know
1595. 01:11:12 that every second, third, fifth, 10th thing
1596. 01:11:16 that you're seeing is just completely artificial and made up.
1597. 01:11:19 So hopefully that's true.
1598. 01:11:20 - Yeah, I agree.
1599. 01:11:21 - I think it would make a better internet.
1600. 01:11:23 - Agreed.
1601. 01:11:25 - All right, chapter seven,
1602. 01:11:26 why do myths about AI persist?
1603. 01:11:29 So in chapter seven, we get into this idea of hype.
1604. 01:11:34 There are major sources of hype.
1605. 01:11:35 There are companies, researchers, journalists,
1606. 01:11:38 public figures.
1607. 01:11:40 They basically are very aggressive.
1608. 01:11:43 The authors are very aggressive in saying
1609. 01:11:45 how they exploit cognitive biases to misinform the public.
1610. 01:11:50 And I was like, well, that's a stance
1611. 01:11:53 that they showed this Gartner hype cycle,
1612. 01:11:56 which I had heard of this before,
1613. 01:11:57 but I'd never seen it graphed out before.
1614. 01:11:59 So if you're listening to us,
1615. 01:12:02 there's visibility on the y-axis.
1616. 01:12:03 There's time on the x-axis.
1617. 01:12:06 The new technology happens
1618. 01:12:08 and you get this peak upward.
1619. 01:12:11 It's not perfectly straight, but it's a peak upward.
1620. 01:12:14 And that's when there's all these inflated expectations.
1621. 01:12:17 AI's gonna save the world.
1622. 01:12:18 AI's gonna do this thing.
1623. 01:12:19 And it's gonna make you not have to do your job.
1624. 01:12:21 And you're only gonna have to work 10 hours a week
1625. 01:12:22 and all this stuff.
1626. 01:12:24 And then you get into the traffic disillusionment
1627. 01:12:26 so you go down hard.
1628. 01:12:28 You get to the slope of enlightenment,
1629. 01:12:29 which is basically where everybody's like, calm down.
1630. 01:12:33 Let's figure out what this thing will actually do.
1631. 01:12:35 And then you get to the plateau of productivity.
1632. 01:12:37 And they're saying that technologies
1633. 01:12:39 go through this Gartner hype cycle.
1634. 01:12:42 So they're making a claim that AI is in this hype cycle.
1635. 01:12:47 Although they say it's not a good way
1636. 01:12:50 to track the adoption and usefulness of AI
1637. 01:12:52 because it actually doesn't work that well.
1638. 01:12:55 So I think it's really interesting.
1639. 01:12:56 They present this Gartner hype cycle
1640. 01:12:58 to try to explain it to us and they're basically like,
1641. 01:13:00 "But wait, it actually doesn't work that well."
1642. 01:13:05 So I was kind of taken back as I'm reading this,
1643. 01:13:08 I was like, well, you just presented this hype cycle
1644. 01:13:09 and then you just told me that it doesn't really work.
1645. 01:13:12 Anyhow, so I don't really know what to do with that.
1646. 01:13:15 I don't know if you wanna go anywhere with that.
1647. 01:13:17 Do you have thoughts on it or?
1648. 01:13:20 - Well, I think the hype cycle is important to understand.
1649. 01:13:27 I don't know that I have a whole lot to discuss about this.
1650. 01:13:30 The one thing that stood out to me from this chapter
1651. 01:13:32 was that the difference between the crypto hype
1652. 01:13:36 and the AI hype is socially beneficial uses,
1653. 01:13:38 which I thought was interesting.
1654. 01:13:40 (laughs)
1655. 01:13:41 Basically said that there are no socially beneficial uses
1656. 01:13:43 for the crypto hype.
1657. 01:13:46 Yeah, but AI hype, I agree that it can be socially beneficial,
1658. 01:13:50 but also, yeah, recognize that there's a bunch
1659. 01:13:52 of really big companies that are funding this hype.
1660. 01:13:58 I also thought it was interesting,
1661. 01:14:00 the curve that you mentioned,
1662. 01:14:02 how it goes back and forth between winters and springs.
1663. 01:14:05 There's not just one winter, there's not just one spring,
1664. 01:14:09 but it's constantly fluctuating.
1665. 01:14:10 I think that's kind of interesting.
1666. 01:14:14 Yeah, I don't know that I have really any big takeaway
1667. 01:14:20 from this, other than don't believe the hype
1668. 01:14:25 and don't believe the news sources
1669. 01:14:26 when they talk about AI, (laughs)
1670. 01:14:29 recognize where the hype is coming from.
1671. 01:14:32 It mentioned that AI research relies on corporate funding,
1672. 01:14:35 so literally the people who are paying
1673. 01:14:37 for all the research being done
1674. 01:14:39 on the effectiveness of all this stuff
1675. 01:14:41 are the people who are trying to sell it.
1676. 01:14:44 Regardless of the name, open AI,
1677. 01:14:45 it's not a non-profit organization.
1678. 01:14:48 It's a for-profit company.
1679. 01:14:49 It reminds me a lot of the nutrition industry
1680. 01:14:53 to where a lot of the nutrition research that gets done
1681. 01:14:56 are by the companies that want to sell you
1682. 01:14:58 the latest protein or the latest weight loss
1683. 01:15:02 or whatever it might be.
1684. 01:15:04 And it's really interesting to me
1685. 01:15:05 that that popped into my head about these two connections.
1686. 01:15:10 It's like, oh, we found this new thing
1687. 01:15:11 and it's gonna revolutionize whatever it might be.
1688. 01:15:14 And is it really?
1689. 01:15:15 And do we take a temper and step back
1690. 01:15:18 and do we let that play itself out?
1691. 01:15:21 I think it's really interesting that time,
1692. 01:15:26 it seems to be a factor here
1693. 01:15:27 that they didn't really bring up,
1694. 01:15:29 but if we just wait a little bit and don't go crazy,
1695. 01:15:34 it'll pan out.
1696. 01:15:35 We'll know how things are gonna work out
1697. 01:15:37 or what's gonna stay around and what's not.
1698. 01:15:39 Now, the people who invested in Bitcoin and a dollar
1699. 01:15:43 and went up to 50, whatever, however many thousands of dollars,
1700. 01:15:47 they're like, oh yeah, but if you'd have waited,
1701. 01:15:50 it's like, yeah, but you're like the 1% of people
1702. 01:15:51 who made money on it.
1703. 01:15:52 Like the other 99% of people lost their shirt on that.
1704. 01:15:58 So I think it's really,
1705. 01:16:00 I think it's an interesting chapter to think about hype
1706. 01:16:03 and think about the way the media
1707. 01:16:06 and the popular figures and the researchers
1708. 01:16:09 and these companies all play into this.
1709. 01:16:12 They state their take-home message here,
1710. 01:16:15 results from AI-based science should be treated
1711. 01:16:17 with extreme caution.
1712. 01:16:19 So be aware, be careful of that.
1713. 01:16:22 They say the most important questions about AI
1714. 01:16:24 aren't about the internals.
1715. 01:16:26 They're not about the way it works
1716. 01:16:27 and how efficiently it works and any of that stuff.
1717. 01:16:29 It's really more about the ethics and the use cases
1718. 01:16:33 and those type of things, policies around them.
1719. 01:16:36 I mean, I just, I think it's an interesting chapter.
1720. 01:16:40 They hit biases again here,
1721. 01:16:41 which I feel like this is like the 10th time
1722. 01:16:43 they've hit these cognitive biases
1723. 01:16:45 that's been all through the book.
1724. 01:16:46 So many biases.
1725. 01:16:48 - Yep.
1726. 01:16:49 So I don't really know if I have much else on this chapter,
1727. 01:16:52 so I'm ready to move if you are.
1728. 01:16:54 - Yeah, let's do it.
1729. 01:16:55 - Okay.
1730. 01:16:55 All right, so the eighth and final chapter,
1731. 01:16:58 where do we go from here?
1732. 01:16:59 So this is their kind of launch off,
1733. 01:17:01 trying to wrap up the book.
1734. 01:17:03 So they started with the introduction
1735. 01:17:04 to lay the foundation of essentially predictive AI bad
1736. 01:17:08 and then they've unpacked some things
1737. 01:17:10 and now we're, where do we go?
1738. 01:17:13 They're gonna say that generative AI
1739. 01:17:15 is going to continue to improve, you know, chat GPT educators.
1740. 01:17:20 They introduced this idea of partial lotteries,
1741. 01:17:23 which Mike, I'd never thought about this,
1742. 01:17:25 but this is a really interesting concept.
1743. 01:17:27 So essentially, if you meet a bar,
1744. 01:17:31 you're then entered into a lottery
1745. 01:17:33 and the rest of the decision making
1746. 01:17:36 isn't trying to be predictive.
1747. 01:17:37 It's just putting you in the lottery
1748. 01:17:38 and said you met the bar, here you go.
1749. 01:17:40 And then from there, it'll let down people's cognitive,
1750. 01:17:47 I want to say cognitive intensity, but that's not right.
1751. 01:17:49 Like my attachment to like, oh, there's something wrong
1752. 01:17:52 with me because I didn't make it into the club
1753. 01:17:55 or the university or whatever it might be.
1754. 01:17:57 It's like, no, I hit the bar and the lottery weeded me out.
1755. 01:18:01 And okay, so they make this idea
1756. 01:18:03 that these partial lotteries could help.
1757. 01:18:08 What else?
1758. 01:18:09 Coming out of here, regulatory capture,
1759. 01:18:11 this idea that, you know, the companies
1760. 01:18:14 are going to continue to try to influence lobbying
1761. 01:18:16 and influence regulations.
1762. 01:18:18 So they're going to keep playing into that
1763. 01:18:20 and we need to think about how that happens.
1764. 01:18:24 Regulation was a big part of this chapter
1765. 01:18:26 'cause they are pro-regulation.
1766. 01:18:29 They think we need to be regulating this,
1767. 01:18:32 that more regulation is better, not less regulation.
1768. 01:18:38 And what else did you get out of this one, Mike?
1769. 01:18:40 - Which I think is true, by the way,
1770. 01:18:43 but that's a tricky thing to unpack.
1771. 01:18:50 Yeah, so this last chapter,
1772. 01:18:52 they have the two worlds that they describe,
1773. 01:18:55 between the two people at the end,
1774. 01:18:57 which is kind of interesting.
1775. 01:18:58 Yeah, and that, so that's interesting to me
1776. 01:19:02 because it's kind of, you know,
1777. 01:19:03 if we're going to really clamp down
1778. 01:19:04 and we're going to be scared of AI versus we're going to
1779. 01:19:07 just embrace it and figure out how to use it.
1780. 01:19:10 Rachel is struggling with this right now
1781. 01:19:12 with her classical conversations campus.
1782. 01:19:15 She's a director this year and someone floated the idea
1783. 01:19:19 of having in the policy handbook statement about AI
1784. 01:19:23 and then she was gone for something for a couple of days
1785. 01:19:26 who didn't respond right away
1786. 01:19:27 and someone tried to jump in and create an AI statement
1787. 01:19:30 to add to the handbook.
1788. 01:19:31 And I was like, I disagree with every word of this.
1789. 01:19:33 - Okay, okay.
1790. 01:19:34 (laughing)
1791. 01:19:36 They're basically just saying like completely, you know,
1792. 01:19:38 cut it out, you can't use it at all.
1793. 01:19:40 And that's, you know, one of the worlds and the other one,
1794. 01:19:43 which is painted a lot more positively is the one
1795. 01:19:45 that's, you know, let's figure out what to do with this.
1796. 01:19:47 I don't know, that's a effective way to end it.
1797. 01:19:49 I think it illustrates their point,
1798. 01:19:51 which is really, you don't need to be scared of this stuff.
1799. 01:19:55 One of the interesting facts that they shared here,
1800. 01:19:57 which, you know, I guess I'd have to look at the one
1801. 01:20:00 of many citations they have behind the things
1802. 01:20:03 that they added in this book.
1803. 01:20:04 But they mentioned that the only job category
1804. 01:20:07 to be completely replaced by technology
1805. 01:20:09 was the elevator operator.
1806. 01:20:10 - Yep, yep.
1807. 01:20:12 - Which kind of tempers the fears about AI
1808. 01:20:19 causing massive joblessness.
1809. 01:20:21 You know, they mentioned that the,
1810. 01:20:23 while AI will impact many jobs,
1811. 01:20:25 the fears of massive joblessness are overblown.
1812. 01:20:27 And I think they're probably right about that.
1813. 01:20:30 I think it's real hard from a bookworm perspective
1814. 01:20:35 to grab onto something specific
1815. 01:20:38 that you're going to do as a result of this last chapter
1816. 01:20:42 or even this book.
1817. 01:20:43 - I agree to you.
1818. 01:20:46 - But it's good information, I guess.
1819. 01:20:49 - Yeah, yeah, I couldn't agree with you more, right?
1820. 01:20:52 I was thinking about where we're gonna go
1821. 01:20:54 and next part of the conversation.
1822. 01:20:55 I was like, action items.
1823. 01:20:56 I was like, I don't have any action items,
1824. 01:20:57 but we'll get there, we'll get there later.
1825. 01:21:01 - Yeah, oh, one other thing.
1826. 01:21:02 They did mention that generative AI will fade
1827. 01:21:04 into the background meaning we'll use it all the time,
1828. 01:21:06 not just for specific tasks.
1829. 01:21:08 I think that's true.
1830. 01:21:10 And so that's the thing specifically
1831. 01:21:12 with the statement in the student handbook about,
1832. 01:21:15 you can't use AI because you're probably cheating.
1833. 01:21:18 I don't, yes, you can use it for creating a report,
1834. 01:21:25 you know, and you'll get a bad grade on it, whatever,
1835. 01:21:27 'cause it's not very good.
1836. 01:21:28 And AI tools are gonna continue to get better,
1837. 01:21:31 but I think this is gonna become part of everyday life.
1838. 01:21:34 And so we really need to figure out
1839. 01:21:36 where to use this appropriately and productively.
1840. 01:21:41 And I don't have a simple answer for this.
1841. 01:21:44 I think if you were to try to lay something out,
1842. 01:21:47 honestly, this is where you really have no idea.
1843. 01:21:49 You can't predict the future
1844. 01:21:50 and how this is gonna change and improve
1845. 01:21:53 and what it's gonna, next version of this
1846. 01:21:55 is gonna look like, but it is something
1847. 01:21:58 that we need to pay attention to.
1848. 01:22:00 - Yeah, I think, I don't wanna go too hard
1849. 01:22:03 on my educator wing of my life.
1850. 01:22:07 They make a statement here that tools
1851. 01:22:08 for detecting AI generated texts don't work.
1852. 01:22:11 And basically it leads to false accusations.
1853. 01:22:13 And you talked about that, you know,
1854. 01:22:15 saying people were cheating when they really weren't.
1855. 01:22:18 It gets really, I don't know,
1856. 01:22:22 I think we lose the force through the trees
1857. 01:22:24 when we try to get too intense about this, right?
1858. 01:22:30 I mean, students are gonna use it.
1859. 01:22:31 Like they're gonna want to use it
1860. 01:22:33 'cause it's a very effective tool to increase efficiency.
1861. 01:22:37 But at the same time, we have to still teach them
1862. 01:22:40 the core things that we have to teach them.
1863. 01:22:42 So it's a really big balance.
1864. 01:22:44 And I don't know if we know the answer to it.
1865. 01:22:46 I don't know if we're gonna know the answer to it.
1866. 01:22:47 I think we're gonna have to feel this out over time.
1867. 01:22:52 I don't know where to go, like you had mentioned it, Mike.
1868. 01:22:55 It's like, I don't know what to do.
1869. 01:22:57 I don't know where to go other than it's hard
1870. 01:22:59 and we need to keep thinking about it
1871. 01:23:00 and regulation needs to happen.
1872. 01:23:02 But companies are gonna keep trying to skirt the regulation.
1873. 01:23:05 And you know, it's like you get into this mind spiral of,
1874. 01:23:08 okay, I just ran on a treadmill, but I got nowhere.
1875. 01:23:11 So what do I do, what do I do with this?
1876. 01:23:13 So that's my end of the book.
1877. 01:23:15 That's my kind of wrap up.
1878. 01:23:17 I don't know if you have anything else you wanna say
1879. 01:23:18 before we move to action items.
1880. 01:23:21 - Nah, let's do action items.
1881. 01:23:24 - Okay.
1882. 01:23:25 Oh, there is one more, there is one more thing.
1883. 01:23:28 You had said about the jobs disappearing.
1884. 01:23:30 I had always been heard and it had always been told,
1885. 01:23:33 oh, radiologists, radiologists are gone, right?
1886. 01:23:36 Like radiologists are gone.
1887. 01:23:37 And they were like super clear on the fact
1888. 01:23:39 that radiologists aren't going away.
1889. 01:23:40 Everybody thinks radiologists are going away,
1890. 01:23:41 radiologists aren't going away.
1891. 01:23:42 And I was like, well, I was wrong
1892. 01:23:45 and everybody that I've ever listened to say that was wrong.
1893. 01:23:47 So I knew there was one other thing I wanted to mention.
1894. 01:23:49 - Interesting.
1895. 01:23:50 - Now we'll move into action items.
1896. 01:23:51 I will start off Mike.
1897. 01:23:53 So hopefully if you were in the same boat as I am,
1898. 01:23:55 you don't feel bad.
1899. 01:23:56 I have zero action items and I don't even feel bad about it.
1900. 01:24:00 Like I thought hard, I considered,
1901. 01:24:04 there are no action items that come out of this book for me.
1902. 01:24:07 I'm glad I read it, but there are no action items.
1903. 01:24:10 Done.
1904. 01:24:11 - Likewise, yeah.
1905. 01:24:15 I don't even think this book was written in a way
1906. 01:24:19 that they wanted you to actually do something with it.
1907. 01:24:24 It was just, here's the information.
1908. 01:24:26 Here's all the citations.
1909. 01:24:28 Hopefully we get tenure.
1910. 01:24:29 - Yeah, the closest action item I would have
1911. 01:24:31 is that reference is kind of interesting.
1912. 01:24:34 Maybe I should go find it, but it was too hard to then,
1913. 01:24:37 like you had to go search for it.
1914. 01:24:38 I was like, nah, I'm done, I'm done.
1915. 01:24:41 - There were a couple of those I tried to look up,
1916. 01:24:43 but you know, the content that they're,
1917. 01:24:45 like one of them specifically was a New York Times article
1918. 01:24:48 and I tried to access it and it's behind a pay wall.
1919. 01:24:50 And like, okay, fine, I won't look at this article.
1920. 01:24:56 - Exactly.
1921. 01:24:57 - I hate it when people do that.
1922. 01:24:58 - All right, so let's move on to style
1923. 01:24:59 I'm rating with zero action items.
1924. 01:25:01 Mike, is that a bookworm first?
1925. 01:25:03 Have you ever had a book where there are zero action items
1926. 01:25:05 coming out of the book?
1927. 01:25:07 - Good question.
1928. 01:25:09 I know that there have been books
1929. 01:25:10 where I've had no action items,
1930. 01:25:11 but I don't know that there's been a double no action item book.
1931. 01:25:15 I think that's a sign of what we're gonna talk about now
1932. 01:25:18 in style and rating.
1933. 01:25:19 So, style and rating, I will start, it is my book.
1934. 01:25:24 Mike, the rating on this one's hard.
1935. 01:25:26 I mean, it provided information
1936. 01:25:29 and what I think is good information.
1937. 01:25:32 It told a story that you don't often hear.
1938. 01:25:35 It told a data-driven story that you don't often hear.
1939. 01:25:38 They have analyzed it well.
1940. 01:25:40 I trust their thought process through things
1941. 01:25:45 and their experience through things
1942. 01:25:46 and the fact that they've researched
1943. 01:25:48 and they've thought critically about these things.
1944. 01:25:50 I liked all of that.
1945. 01:25:51 I feel as though I am more well-rounded
1946. 01:25:55 in my understanding of AI.
1947. 01:25:56 I feel as though I understand
1948. 01:25:58 some of the technical pieces better.
1949. 01:26:00 I definitely have a better perspective on the biases
1950. 01:26:03 and the potential implications of things.
1951. 01:26:06 I didn't care for the book very much at all.
1952. 01:26:11 If the information wouldn't have been really good,
1953. 01:26:14 I mean, this would maybe be in the one territory for me.
1954. 01:26:18 So, the fact that the information is pretty good
1955. 01:26:22 and I feel, I don't feel like I wasted my time.
1956. 01:26:24 So, I wanna make sure that that's clear.
1957. 01:26:25 Like, I don't feel like I wasted my time reading the book.
1958. 01:26:28 I would, I do feel a little bad
1959. 01:26:31 for the bookworm audience if they read along with us
1960. 01:26:33 because I was like, man, if they weren't into,
1961. 01:26:35 if they weren't into the AI scene,
1962. 01:26:37 I mean, this was a really tough read
1963. 01:26:39 and they probably stopped at some point through this.
1964. 01:26:42 So, I'm trying to get to my rating
1965. 01:26:45 and I'm trying to debate whether I think it's a two or a three.
1966. 01:26:48 I don't wanna do a half.
1967. 01:26:50 I'm going to give it the benefit of the doubt
1968. 01:26:53 and I'm going to rank it a three
1969. 01:26:55 even though my gut kind of feels like it's a two
1970. 01:26:57 but I'm gonna give it the benefit of the doubt
1971. 01:26:58 because it did have good information
1972. 01:27:01 that I think rounded out my picture of AI
1973. 01:27:04 and predictive AI in the industry.
1974. 01:27:07 So, I'm going to rank this a three.
1975. 01:27:09 Mike, what's yours?
1976. 01:27:10 All right, well, okay.
1977. 01:27:13 So, I actually, I think I liked the book
1978. 01:27:17 maybe better than you did.
1979. 01:27:19 I don't have a ton of notes from this book
1980. 01:27:22 because I quickly gave up trying to keep up
1981. 01:27:26 with every single argument that they were making
1982. 01:27:29 and just relied on the things that actually resonated
1983. 01:27:32 and there were definitely some things
1984. 01:27:35 from things that they shared that triggered some other things
1985. 01:27:38 which led to great conversation.
1986. 01:27:40 I don't think this book in isolation is great
1987. 01:27:45 but I do think it is good
1988. 01:27:47 and I think the arguments that they make are effective.
1989. 01:27:50 I think I kind of wish they would have skipped
1990. 01:27:56 chapters two and three altogether
1991. 01:27:59 and just said trust us predictive AI isn't a thing right now.
1992. 01:28:02 We'll revisit this in the future.
1993. 01:28:05 I don't, so I really don't care about the vendetta
1994. 01:28:08 that they have against predictive AI to be honest.
1995. 01:28:11 So, the AI snake oil, I don't like that framing.
1996. 01:28:13 I don't like that perspective necessarily
1997. 01:28:18 but I did enjoy several of the chapters specifically
1998. 01:28:21 and I think they have some interesting things
1999. 01:28:23 to add to the conversation
2000. 01:28:25 and I definitely walk out of this
2001. 01:28:27 with a better understanding,
2002. 01:28:28 not just because of the arguments in this book
2003. 01:28:30 but they fit in alongside the other things
2004. 01:28:32 that I'm aware of that are happening in the AI space
2005. 01:28:35 and the other books that we've read on AI in general.
2006. 01:28:38 So, like I'm trying to separate this
2007. 01:28:40 between like in an isolation read
2008. 01:28:43 and the Mortimer Adler, how does this contribute?
2009. 01:28:47 And I think that way it's actually more valuable.
2010. 01:28:51 So definitely not a five star book
2011. 01:28:53 but maybe a four star book.
2012. 01:28:55 The thing that's gonna keep me from giving it four stars though
2013. 01:28:57 is I don't think this age is very well.
2014. 01:29:01 - Yeah, okay.
2015. 01:29:01 This book was good right now as we read it
2016. 01:29:05 in July, August of 2025.
2017. 01:29:09 By the time January 1st, 2026 comes,
2018. 01:29:13 I'd really think there are gonna be parts of this
2019. 01:29:16 that are completely out of date
2020. 01:29:19 where not just, oh, that's old information
2021. 01:29:22 but no, it's changed since then.
2022. 01:29:25 And that's not necessarily their fault
2023. 01:29:29 but that's just my own feel for where things are at
2024. 01:29:32 currently with AI.
2025. 01:29:34 I was listening to, there's a podcast called AI
2026. 01:29:37 and I by Dan Shipper.
2027. 01:29:39 He's one of the people behind every.2.
2028. 01:29:43 And they're the ones that I think are one of the better voices
2029. 01:29:48 to listen to right now in terms of like where our AI is
2030. 01:29:52 and where it's going.
2031. 01:29:53 And there was an episode that I listened to.
2032. 01:29:54 I can't remember who it was with.
2033. 01:29:55 He was interviewing somebody
2034. 01:29:56 and they've worked with AI and they were at Stripe
2035. 01:30:01 and Google and all these different places.
2036. 01:30:03 And at the end of that episode,
2037. 01:30:04 they mentioned that people, they made a comment
2038. 01:30:07 about how people think that where we're at right now
2039. 01:30:11 with the LLMs is the end game.
2040. 01:30:15 He's like, this isn't even the end of the first inning.
2041. 01:30:18 He's like, this is going to change so incredibly much.
2042. 01:30:22 People have no idea.
2043. 01:30:25 And I feel like just the whole snake oil perspective
2044. 01:30:29 is sort of putting a stamp on, yeah, I can't ever do that.
2045. 01:30:34 You know, it can't do that right now.
2046. 01:30:36 And I think they would probably admit that.
2047. 01:30:38 But I don't believe that we're at a point
2048. 01:30:41 where the predictive AI stuff is just never, ever, ever,
2049. 01:30:44 ever going to work regardless of how much they leaned into this
2050. 01:30:47 about why I can't AI predict the future.
2051. 01:30:49 You give something enough data
2052. 01:30:51 that's going to be useful enough in some situations.
2053. 01:30:55 And we're not there yet, but I think we will be.
2054. 01:30:57 So I'm going to give this three stars
2055. 01:30:59 just because I don't want to come back to this rating
2056. 01:31:02 a year from now and be like, yeah, no.
2057. 01:31:05 - Yeah, you make a really, really good point
2058. 01:31:08 in terms of what it made me think of as this book say,
2059. 01:31:13 it's a dangerous book to write.
2060. 01:31:17 Right now you plant your flag really well
2061. 01:31:19 and you can get kind of popular with that.
2062. 01:31:22 But in 2026, everybody looks back and they're like,
2063. 01:31:25 hey, do you remember when those two guys were at that book
2064. 01:31:28 and how wrong they were about all this?
2065. 01:31:30 And it's like, that's a dangerous thing.
2066. 01:31:32 You don't know if that's going to happen or not.
2067. 01:31:33 And that's what, you know, when you were describing
2068. 01:31:35 your perspective on it or your view on it, it's like,
2069. 01:31:39 wow, I didn't think about, it's good for now.
2070. 01:31:41 It seems to have gotten popularity right now.
2071. 01:31:44 Hopefully that stays true.
2072. 01:31:46 Now the other option is you look back on it
2073. 01:31:47 and you're like, look how right they were.
2074. 01:31:49 They look how right they were the entire time.
2075. 01:31:50 And it's just a, it's a risk.
2076. 01:31:51 I mean, you're putting your flag in the ground
2077. 01:31:54 and you see where it goes, so.
2078. 01:31:56 - Which is not to say that they should not have written
2079. 01:31:59 this book or they should not have taken that approach.
2080. 01:32:01 It's just not the one that I would have chosen to read
2081. 01:32:05 from a book or in perspective.
2082. 01:32:07 I think as academics, they have alternate goals
2083. 01:32:11 with publishing their thoughts, beliefs.
2084. 01:32:14 - Without a doubt.
2085. 01:32:15 - Research around these things, which isn't, you know,
2086. 01:32:18 how do we make the best, most informative book
2087. 01:32:22 for a bookworm listener who's coming to this
2088. 01:32:26 and they want to know how they can actually use AI
2089. 01:32:28 in a positive way.
2090. 01:32:30 I think it was a great conversation.
2091. 01:32:32 I think it, you know, one of my,
2092. 01:32:34 probably one of my more, it's gotta be up there
2093. 01:32:37 in terms of like my favorite conversations that we've had
2094. 01:32:40 just 'cause there were so many topics in here
2095. 01:32:41 was like I have very strong opinions and feelings about this.
2096. 01:32:45 So hopefully it was entertaining for the listeners,
2097. 01:32:48 but yeah, that's the thing that causes me to say that
2098. 01:32:53 and you're right, you know, maybe come back to this
2099. 01:32:55 and, you know, all those arguments hold up,
2100. 01:32:57 but I kind of have this feeling that it's not gonna age
2101. 01:33:02 well and that's not their fault,
2102. 01:33:03 but it's gonna weigh into my decision.
2103. 01:33:05 - Got it.
2104. 01:33:06 All right, so let's put AI snake oil on the shelf.
2105. 01:33:09 What's next, Mike?
2106. 01:33:12 - Next is From Strength to Strength
2107. 01:33:14 by Arthur Brooks, which I have a couple of copies of.
2108. 01:33:19 So if you really need one, I could set you one.
2109. 01:33:22 I bought this and then my brother gave it to me
2110. 01:33:24 for Christmas, I think,
2111. 01:33:25 and it's been on my list for a while.
2112. 01:33:28 Looks really interesting and I'm looking forward
2113. 01:33:30 to going through this one with you.
2114. 01:33:32 - Nice.
2115. 01:33:33 After that, we're gonna do the Science of Scaling.
2116. 01:33:36 It's a business book by Benjamin Hardy and Blake Erickson.
2117. 01:33:41 So I'm looking forward to that one.
2118. 01:33:42 It's a new one that comes out
2119. 01:33:44 according to Amazon tomorrow.
2120. 01:33:46 So it's a brand new one on scaling your business.
2121. 01:33:50 - Well, tomorrow is we record this.
2122. 01:33:51 - Yeah, tomorrow is we record it.
2123. 01:33:52 - Most people who are listening to this.
2124. 01:33:54 - It will be out.
2125. 01:33:54 - It's available, so.
2126. 01:33:55 - Yes, tomorrow is we record, yep.
2127. 01:33:58 Mike, do you have any gap books right now?
2128. 01:34:02 - Good question.
2129. 01:34:03 I don't think, yeah, I didn't put anything in here,
2130. 01:34:06 but I actually do have one that I want to read.
2131. 01:34:09 I picked this book up in person at Craft & Commerce last year.
2132. 01:34:14 I have been kicking around the idea
2133. 01:34:16 of doing a coaching program
2134. 01:34:19 and John Meese is basically the guy for setting this up
2135. 01:34:24 and he wrote a book called Serve to Sell
2136. 01:34:27 and he gave it away at Craft & Commerce last year.
2137. 01:34:30 So I have a copy of it, it's fairly short
2138. 01:34:33 and I'm gonna try and get that one done
2139. 01:34:36 before from strength to strength.
2140. 01:34:38 - Although that one has a little bit shorter window
2141. 01:34:41 because we will actually be recording that one in person.
2142. 01:34:44 So yes.
2143. 01:34:45 - We will be recording in person,
2144. 01:34:46 which is gonna be an exciting adventure for us,
2145. 01:34:48 although Mike understands how to record in person,
2146. 01:34:50 so I feel good about that.
2147. 01:34:53 - Yeah.
2148. 01:34:54 - Mine is the same as it was before.
2149. 01:34:57 I kind of got into this AI book
2150. 01:35:00 and then didn't have tech for the week.
2151. 01:35:04 So I didn't have my head, my nose and anything else,
2152. 01:35:08 but I'm gonna finish scrolling ourselves to death,
2153. 01:35:10 which is a book about technology
2154. 01:35:12 and Christianity and the impact there,
2155. 01:35:15 the intersection there.
2156. 01:35:15 So that's the one I'm gonna finish as my gap book
2157. 01:35:19 before strength to strength.
2158. 01:35:21 - Nice.
2159. 01:35:22 - All righty, well, thank you all very much for listening.
2160. 01:35:26 If you are interested in more content that we make,
2161. 01:35:28 we do a pro show, it's a longer ad-free show.
2162. 01:35:33 You also will get the bootleg version of the show,
2163. 01:35:36 which releases fairly quickly after recording,
2164. 01:35:38 usually get it about a week or two early.
2165. 01:35:40 There's a live stream, so if you become a pro member,
2166. 01:35:43 you can get access to the live stream through Patreon.
2167. 01:35:48 We have a wallpaper there, it's $7 a month,
2168. 01:35:51 and it just helps us with the admin cost to run the show.
2169. 01:35:55 So for all the Patreon members,
2170. 01:35:57 we thank you so much for being Patreon members,
2171. 01:35:59 for anybody who's interested in kind of getting the perks
2172. 01:36:02 and just supporting the show,
2173. 01:36:04 we really would appreciate if you do that,
2174. 01:36:05 if you go to patreon.com/bookwormfm.
2175. 01:36:09 - All right, if you are reading along with us,
2176. 01:36:11 pick up from strength to strength by Arthur Brooks,
2177. 01:36:14 and we'll talk to you in a couple of weeks.