Transcriptly

Fix your AI-limited mindset in 12 mins | Caltech, Anima Anandkumar

EO11:54Added Aug 31, 2026

Caltech's Professor Anima Anandkumar states, "A lot of these AI tools are getting better, but you still need to provide AI what to do." So how can we discove...

Watch on YouTube →
Contributed by 刘嘉琪

Transcript

Transcript format

  1. Intro

  2. 00:00A lot of these AI tools are getting
  3. 00:01better, but that only means they can do
  4. 00:04a certain set of instructions which are
  5. 00:06seen in data. You still need to provide
  6. 00:09AI what to do, right? You still need to
  7. 00:12be able to describe it. And that ability
  8. 00:15to describe what are the tasks AI should
  9. 00:17do, what are the programs to be written
  10. 00:19is still important. I think one job that
  11. 00:22will not be replaced by AI is the
  12. 00:25ability to be curious and go after hard
  13. 00:28problems. So for young people, my advice
  14. 00:30is not to be afraid of AI or worry what
  15. 00:34skills to learn that AI may replace them
  16. 00:37with, but really be in that path of
  17. 00:40curiosity. I'm an animant Kumar. I'm
  18. 00:43brand professor at Caltech. I've been a
  19. 00:45professor at Caltech for about 8 years
  20. 00:47now and during that time I also had
  21. 00:50stints in industry. I was principal
  22. 00:52scientist at Amazon Web Services. So at
  23. 00:55Caltech, I lead the AI and science lab,
  24. 00:58which means uh working on some of the
  25. 01:01hardest challenges we see in science and
  26. 01:03engineering and how we can not only use
  27. 01:07existing AI methods to solve them, but
  28. 01:09really develop new ones.
  29. 01:12[Music]
  30. 01:20I think one job that will not be
  31. 01:22replaced by AI is the ability to be

  32. Start Where You Are Curious

  33. 01:25curious and go after hard problems. For
  34. 01:28a lot of students, there is a strong
  35. 01:30motivation to just conform and go ahead.
  36. 01:33Right? The number one thing I would ask
  37. 01:36is to question everything. Think
  38. 01:38critically. I always begin the classes
  39. 01:40by asking questions, not writing down
  40. 01:42math equations, right? Not going into
  41. 01:44the details but just intuitively based
  42. 01:47on everything that you've seen you've
  43. 01:49done what do you think I asked them
  44. 01:52simple cases for instance if you were to
  45. 01:54design for a fire alarm when should it
  46. 01:57say that there is fire or not that's
  47. 01:59something you can change right you can
  48. 02:01put a threshold you know what level of
  49. 02:03smoke or when does it think it's smoky
  50. 02:05and that's a very practical question so
  51. 02:08if you just had fire alarms every day
  52. 02:10we'd be just out and it would be useless
  53. 02:13we would not have a functional office
  54. 02:14space. But on the other hand, if we
  55. 02:17never set fire, that would be bad, too.
  56. 02:19So, how to balance this and how to model
  57. 02:22how noisy this sensor could be. And
  58. 02:25sometimes I see with students who may
  59. 02:27not have the mathematical training, but
  60. 02:29they're very intuitive and practical.
  61. 02:31They may be like, oh, I would go and
  62. 02:32measure how noisy it is, or I would show
  63. 02:34different levels of smoke, have like
  64. 02:36candles of different sizes and go and
  65. 02:39test it. that and you know many times
  66. 02:41people already maybe somewhat aware of
  67. 02:43this so they have some intuitive ideas
  68. 02:45so that already is a good starting point
  69. 02:47and sometimes they are maybe really
  70. 02:49wrong they have an intuition but that's
  71. 02:51a wrong intuition which is still okay
  72. 02:53because intuitions are not always the
  73. 02:56only answer right so I think a lot of it
  74. 02:58comes by asking questions and now you
  75. 03:01can use these AI tools to get answers
  76. 03:04very quickly and also verify them a lot
  77. 03:06of it comes from being just like curious
  78. 03:09or interested and that could be one
  79. 03:11specific topic and if somebody's
  80. 03:13interested in music they can delve
  81. 03:15deeper into that if somebody is
  82. 03:16interested in art so it just has to that
  83. 03:19spark has to come from within and I
  84. 03:21think giving students more the freedom
  85. 03:24to pursue where they are passionate
  86. 03:26where they have a spark I think is going
  87. 03:29to be the future and that's the right
  88. 03:31thing rather than forcing everybody to
  89. 03:33learn
  90. 03:35everything I'm always motivated by the

  91. How I Started Where I Was Curious

  92. 03:38hardest challenge challenges. You know,
  93. 03:39I want to know what is difficult, but
  94. 03:41also why it's difficult, right? And even
  95. 03:44though I may not be able to solve it
  96. 03:45today, how do we build up the
  97. 03:47foundations to get there? Growing up in
  98. 03:50my sore as a kid, I loved just solving
  99. 03:54math problems, you know, going to my
  100. 03:56parents' factory. I was reading up their
  101. 03:59program manuals. I was, you know,
  102. 04:01learning how they could be programmed.
  103. 04:03And unlike in other computer programs
  104. 04:05here if something was wrong that would
  105. 04:08lead to like physical failure parts
  106. 04:10being like not manufactured correctly
  107. 04:13and I'm like oh but how does it go into
  108. 04:14the computer and how does the computer
  109. 04:16tell the machine what to do so there
  110. 04:18were always gaps because as a kid you
  111. 04:21don't know everything but to me it was
  112. 04:23like observing and then understanding
  113. 04:26what the gap is and even if I didn't get
  114. 04:28an immediate answer I would remember
  115. 04:30that there is a gap and then later when
  116. 04:32I was introduce use those topics. I was
  117. 04:34in my mind I was like, "Oh, that's what
  118. 04:36it relates to." So somehow I had built
  119. 04:39up that mental map and I had put places
  120. 04:42of where things I knew and things I
  121. 04:44didn't know and I kept kind of growing
  122. 04:47that in my mind as I progressed through
  123. 04:50the years. So when I was growing up, AI

  124. Developing the AI That Changes the Real World

  125. 04:53was considered science fiction.
  126. 04:55Naturally, there were lots of science
  127. 04:57fiction movies where I saw and was
  128. 05:00fascinated. But that's not something
  129. 05:02people thought were practical. Uh since
  130. 05:04I was in middle and high school and now
  131. 05:07it's almost 30 years, right? It is a
  132. 05:10long time, but the amount of progress
  133. 05:12that has happened is also so astounding
  134. 05:15in so many ways. So since I joined
  135. 05:17Caltech in 2017, the timing just felt
  136. 05:20right to use AI as a tool and a
  137. 05:23framework to solve some of the hardest
  138. 05:26problems which until that point was not
  139. 05:28considered practical. So after I joined
  140. 05:31Caltech and wanted to explore problems
  141. 05:34at the intersection of AI and science
  142. 05:36and I was talking to everybody on
  143. 05:37campus, I was like okay do you need
  144. 05:40compute? What do you need it for? Let me
  145. 05:42understand the problems that you're
  146. 05:44tackling. And you know I can't possibly
  147. 05:46go and solve each one of them problems
  148. 05:48myself. Right? My question then was are
  149. 05:51there general tools we can develop that
  150. 05:53could impact so many different areas?
  151. 05:55And that again put me back to
  152. 05:57mathematical foundations. So because a
  153. 06:00lot of these different real world
  154. 06:01phenomena are modeled by partial
  155. 06:03differential equations. So now can we
  156. 06:05design AI that can solve this and do it
  157. 06:08much faster do it much better than what
  158. 06:11is currently being done with traditional
  159. 06:13simulations. And to do that we developed
  160. 06:16neural operators. We've invented an AI
  161. 06:19technology called neural operators that
  162. 06:21is trained to understand physical
  163. 06:24behaviors not just highle reasoning with
  164. 06:27text. So think of a hurricane. The
  165. 06:29hurricane if you will just eyeball it,
  166. 06:32can you tell where it's going to go? You
  167. 06:34know, most humans cannot, right? It's a
  168. 06:36superhuman skill to predict where
  169. 06:39hurricanes are going to go. And for to
  170. 06:42do that, we need finecale information
  171. 06:44and finecale modeling. So this cannot be
  172. 06:46just a core scale image like the image
  173. 06:49of a cat where even if it's gets blurry,
  174. 06:52you know it's a cat. The same techniques
  175. 06:54don't work for phenomena like
  176. 06:56hurricanes. Once we developed tools and
  177. 06:58then the next natural question is what
  178. 07:01were practical use cases that involved
  179. 07:03and the weather models was a natural one
  180. 07:06because it's widely used. It has huge
  181. 07:09implications on our lives especially if
  182. 07:11extreme weather events like hurricanes
  183. 07:13if we get them right that has the
  184. 07:15potential to save human lives and also
  185. 07:18bring down economic costs. So I was
  186. 07:20motivated by how it can be helpful to
  187. 07:22people, but I was also motivated by that
  188. 07:26being considered a very hard technical
  189. 07:28challenge. In fact, just a few months
  190. 07:30before we released our model, there were
  191. 07:32a group of very wellrespected weather
  192. 07:35scientists who published in the Royal
  193. 07:37Society Journal thinking they're they
  194. 07:39were under the impression that AI would
  195. 07:42take more than a decade or even longer
  196. 07:44to replace traditional ways to forecast
  197. 07:47weather. and they felt AI was just not
  198. 07:49ready. This problem is way too
  199. 07:51difficult. And we released this and it
  200. 07:53just took everybody by surprise. It was
  201. 07:55not only accurate, it was tens of
  202. 07:58thousands of times faster. So what would
  203. 08:00take a big supercomput for traditional
  204. 08:03weather models can now be run on a local
  205. 08:06gaming PC with just a consumer GPU. So
  206. 08:09that's the beauty of uh machine learning
  207. 08:12as a field. We are not always stopped by
  208. 08:16what others think as difficult. As long
  209. 08:18as we can get the data and we can design
  210. 08:21the methods, we can just go and try
  211. 08:23it. So my mission is to constantly be

  212. Can AI Replace Scientists?

  213. 08:26curious and learning and not assume that
  214. 08:31any problem is easy. I can't imagine a
  215. 08:33world where scientists will be out of
  216. 08:35jobs because the definition of a
  217. 08:37scientist is somebody who tackles open
  218. 08:40problems, right? So there are harder and
  219. 08:43harder problems to tackle. You know, if
  220. 08:45you want to look at the deep secrets of
  221. 08:48our universe, go down to the smallest
  222. 08:50scale and understand at the atomic and
  223. 08:53subatomic level how matter is
  224. 08:56constructed to of course level of galaxy
  225. 08:59and beyond and understand how the
  226. 09:00universe is put together. There are
  227. 09:03still lots of open challenges. Many
  228. 09:06other teams such as Google Deepine focus
  229. 09:08on what is known as an AI scientist.
  230. 09:11Meaning AI that comes up with new ideas.
  231. 09:14But so much of scientific progress is
  232. 09:16not limited by the lack of new ideas,
  233. 09:19right? Lots of people have lots of
  234. 09:21ideas. But the bottleneck is going to
  235. 09:24the lab or going to the real world and
  236. 09:26testing them. That is slow. That is
  237. 09:28expensive. So my focus is how we can
  238. 09:31replace those lab experiments. Can we
  239. 09:34come up with AI that inherently
  240. 09:35understands the physics better? So we
  241. 09:38can completely avoid the lab experiments
  242. 09:40or maybe only do it to do the final
  243. 09:43testing. And so with that focus and with
  244. 09:45that physical knowledge, we can come up
  245. 09:48with AI designed answers that we can go
  246. 09:52directly to the real world and minimize
  247. 09:55this need for testing. To me, human
  248. 09:58agency is driving AI to do something

  249. Does AI Kill Curiosity?

  250. 10:01that you want it to be done. You have
  251. 10:04the agency as a human to decide what
  252. 10:07tasks AI does and then you're evaluating
  253. 10:11and you're in charge, right? So, you go
  254. 10:13and verify whether what AI is saying is
  255. 10:16true or not and then over time provide
  256. 10:18that feedback to AI and make it
  257. 10:21better. AI is a tool. It can both help
  258. 10:25curiosity but also kill it depending on
  259. 10:27how it's used. Right? So for young
  260. 10:30people, my advice is not to be afraid of
  261. 10:32AI or worry what skills to learn that AI
  262. 10:36may replace them with, but really be in
  263. 10:39that path of curiosity, right? Use AI as
  264. 10:42a tool to drive that curiosity, learn
  265. 10:45new skills, new knowledge, and you can
  266. 10:48do that in a much more interactive way.
  267. 10:50Even when it comes to writing computer
  268. 10:52programs, you know, a lot of these AI
  269. 10:54tools are getting better, but that only
  270. 10:56means they can do a certain set of
  271. 10:58instructions which are seen in data. You
  272. 11:01still need to provide AI what to do,
  273. 11:05right? You still need to be able to
  274. 11:06describe it. And that ability to
  275. 11:09describe what are the tasks AI should
  276. 11:11do. What are kind of highlevel
  277. 11:13understanding of what AI is doing when
  278. 11:16it writes these computer programs is
  279. 11:18still important because a bad programmer
  280. 11:21who is not better than AI will be
  281. 11:23replaced. But a great programmer who can
  282. 11:26assess what AI is doing, make fixes,
  283. 11:29ensure those programs are written well
  284. 11:32will be in more demand than ever.
Fix your AI-limited mindset in 12 mins | Caltech, Anima Anandkumar — Transcriptly