Transcriptly

A CS Professor on Why Slow Learning Wins in the AI Era | CU Boulder, Tom Yeh

EO12:37Added Aug 31, 2026

Tom Yeh, Associate Professor of Computer Science at the University of Colorado Boulder & creator of AI by Hand, breaks down why he teaches AI by hand at huma...

Watch on YouTube →
Contributed by 刘嘉琪

Transcript

Transcript format

  1. Intro

  2. 00:00My name is Tom Yed. I'm a professor of
  3. 00:02computer science at the University of
  4. 00:05Colorado in Boulder. I'm also a founder
  5. 00:07of AI by hand is a global education
  6. 00:10initiative to make AI inside the black
  7. 00:12box accessible and approachable. Writing
  8. 00:15out all the maths with hand in doing so
  9. 00:17understanding that AI is not a big
  10. 00:20mystery [music] is something we all can
  11. 00:21understand. What is the purpose of
  12. 00:23learning? Having the answer doesn't mean
  13. 00:25you know it. People can buy degree by
  14. 00:27certificate. Do you have ownership of
  15. 00:29this particular idea core and
  16. 00:32foundational that doesn't change is
  17. 00:34evergreen? AI cannot change people but
  18. 00:37you can change AI.

  19. Why I Teach AI by Hand- The Power of Learning "Slow"

  20. 00:48A transformer is to me to process
  21. 00:52individual words in your sentence. And I
  22. 00:54can started with say well at the token
  23. 00:57number four in my sentence and then just
  24. 00:59then this is a boxes I draw to show that
  25. 01:01well each token actually has multiple
  26. 01:04number in this case simple case that
  27. 01:05says three numbers. Unfortunately while
  28. 01:08I was a student I missed deep learning
  29. 01:11entirely. I was a bit too old. So I
  30. 01:14studied support vector machines
  31. 01:15traditional [music] machine learning
  32. 01:17method. So when then I became a
  33. 01:18professor all of a sudden all these
  34. 01:20people are doing deep learning. I have
  35. 01:22to learn deep learning all over again.
  36. 01:24So what I'm doing today with AI by hand
  37. 01:27is to share my learning journey. How I
  38. 01:29as a old professor trying to learn deep
  39. 01:31learning from scratch.
  40. 01:34All the things I'm sharing has been
  41. 01:37about to show my own struggle with
  42. 01:39understanding AI model math and
  43. 01:41algorithms. Only way I can get it is I
  44. 01:43[music] get to draw or write on the
  45. 01:46paper. This is what oh I actually get
  46. 01:47it. So when I share my drawing, shared
  47. 01:50my way to map out the math and a lot of
  48. 01:53people resonated and people started to
  49. 01:55comment out, hey, I really like your
  50. 01:56approach to breaking down by hand. I
  51. 01:59started, hey, maybe I just call it AI by
  52. 02:01hand. There's a reason that people
  53. 02:03resonate with this. Connect with this
  54. 02:06thing that I'm doing by hand.
  55. 02:09Why do we like to calculate this by hand
  56. 02:13when AI can do this very well? when I
  57. 02:16was teaching introduction to programming
  58. 02:19and I have a feedback like I going too
  59. 02:21fast. I'm going to s slides and so on
  60. 02:24and just [music] too fast because I
  61. 02:26really like to share a lot of my
  62. 02:27teaching and knowledge with my students.
  63. 02:30So I decided to teach the entire
  64. 02:32semester of C++ programming on the
  65. 02:34blackboard instead of doing live coding.
  66. 02:36So I decided to do that. So I have a
  67. 02:38whole semester of doing writing the I
  68. 02:41this my notes about I wrote it down on
  69. 02:44the piece of paper as a semester
  70. 02:46progressing I see a few benefits. One is
  71. 02:48that I can only go at the humanly
  72. 02:51possible speed of my writing. I cannot
  73. 02:53go any faster they can write. Second
  74. 02:56student can only learn at a humanly
  75. 02:59possible speed. They can only follow how
  76. 03:01much I write. And number three is that
  77. 03:03if I get my student to use their hand to
  78. 03:06copy my notes on a notebook, their hands
  79. 03:08are not on their keyboard checking their
  80. 03:10Instagram messages. So help with focus
  81. 03:13as well. The by hand is really about a
  82. 03:15way to connect back to our humans. So
  83. 03:18using your hand, you can go at a human
  84. 03:20speed. You commute in a human way. And
  85. 03:23over time I learned this. I started to
  86. 03:25see this values of going back to the old
  87. 03:28school by hand.
  88. 03:31What is the purpose of learning? Is it
  89. 03:33about that physical or digital artifact
  90. 03:36that prove they have learning or
  91. 03:38something you feel like you actually
  92. 03:40internalize you actually own this kind
  93. 03:42of [music] stuff? Do you have ownership
  94. 03:43of this particular idea? Well, AI give
  95. 03:45me an answer right away. Having the
  96. 03:47answer doesn't mean you know it people
  97. 03:49can buy degree by certificate. Whether
  98. 03:52you or not you own something, you value
  99. 03:54something is actually proportional to
  100. 03:56how much [music] time you spend um
  101. 03:58acquiring that piece of knowledge. You
  102. 04:00have to first define what learning
  103. 04:02actually means to you.

  104. The Foundation That Doesn't Burn

  105. 04:08I remember when I was a undergrad, we're
  106. 04:10learning linear algebra as part of a
  107. 04:13requirement for getting a CS degree. We
  108. 04:15have learned linear algebra and we have
  109. 04:17learned max computation. I have no idea
  110. 04:19why it is even important. But then it
  111. 04:22turns out over time computer graphics
  112. 04:24became really popular because of
  113. 04:26Jurassic Park. They put this CGI up the
  114. 04:29front and people talk about hey
  115. 04:30everybody need to learn CGI and CGI uses
  116. 04:34a lot of matching application and then
  117. 04:36after a few years there was big data
  118. 04:38movement then it turns out you also need
  119. 04:40maximum application to do some sort of
  120. 04:43processing and then move to machine
  121. 04:44learning again forget about data science
  122. 04:47we should be machine learning specialist
  123. 04:48engineer again maximum application and
  124. 04:51today's AP is AI AI everybody need to be
  125. 04:54AI native we should raise our kids send
  126. 04:56into AI
  127. 04:57[music]
  128. 04:57M application in few years we all all we
  129. 05:00talk about is quantum computing is
  130. 05:02possible right and guess what m
  131. 05:04application again so you see this a
  132. 05:06trend that every time there's something
  133. 05:07that tool keep changing but it's always
  134. 05:10something that's core and foundational
  135. 05:12that doesn't change it's evergreen you
  136. 05:14could revisit a year from now two years
  137. 05:17from now it's still relevant people
  138. 05:18still care a lot about whereas the
  139. 05:20deepseek there was popular at the time
  140. 05:22but there's been a while now so deepse
  141. 05:24not as popular as before there's a new
  142. 05:26new thing for instance like this cloud
  143. 05:28super popular but let's see in two two
  144. 05:30months is it still popular we don't know
  145. 05:32but I'm pretty confident transformer
  146. 05:34topic is still going to be popular
  147. 05:37couple summers ago I had a opportunity
  148. 05:39to visit uh South Korea and I got to
  149. 05:42visit where everybody else will go the
  150. 05:45what's called Kukong there this palace
  151. 05:47that's that was just a view of the
  152. 05:49history very beautiful palace and then
  153. 05:51what just struck me when I learned about
  154. 05:53a bit more history the entire thing was
  155. 05:56burned down in like a 1500s except for
  156. 06:00the foundation that was made in Saudi
  157. 06:03rock. So in 1800s they rebuilt the
  158. 06:06entire palace based off the same
  159. 06:08foundation. Like to talk about this
  160. 06:10story because that reminds me of how
  161. 06:13this technology has been changing over
  162. 06:15and over again. But if you had a
  163. 06:18foundation on the maximum application,
  164. 06:20you could just apply it to AI. It
  165. 06:22doesn't really matter. We build your
  166. 06:23skill based on your solid foundation.
  167. 06:25[music] And so that's why I'm focusing
  168. 06:27on foundation because I believe there
  169. 06:29something you can rebuild. Doesn't
  170. 06:31really matter whether the new tools is
  171. 06:33obsolete. If you keep focus on the
  172. 06:35surface features, [music] the tools then
  173. 06:38forget about foundation. You just have
  174. 06:40to keep rebuilding your houses. You
  175. 06:42still never have your foundation up or
  176. 06:44can build a pump. So how can it apply to
  177. 06:47your own situation? Think about the way
  178. 06:49you grew up. Maybe your parents sent you
  179. 06:51to a soccer game or maybe the piano and
  180. 06:54you have something some skill you have
  181. 06:57become good at. Is it piano? Is it
  182. 07:00chess? And think about the way you
  183. 07:01acquire the skill and that is actually
  184. 07:05something that doesn't change. So as AI
  185. 07:08tool come every day the fact that you
  186. 07:10could acquire very difficult skill that
  187. 07:12is something that part of your identity
  188. 07:14that doesn't change. If you continue to
  189. 07:16focus [music] on that and you have
  190. 07:17ability to realize, hey, I could acquire
  191. 07:20this, I can learn this skill, I can
  192. 07:21become really good at it. That's when
  193. 07:23you could continue to apply that skill
  194. 07:25to a new AI tool. So, I'm a good
  195. 07:28example. I was falling behind on deep
  196. 07:30learning for quite a while. But I have
  197. 07:32learned skills really trying to break
  198. 07:35down difficult [music] topics by
  199. 07:37patiently writing everything down on
  200. 07:39paper. So that skill I able to
  201. 07:41eventually catch up. I caught up on deep
  202. 07:44learning by from a position that's way
  203. 07:46behind from people who actually been
  204. 07:49working on deep learning for a long
  205. 07:50time. So for you your piano skill, your
  206. 07:53soccer skill is not useless. That will
  207. 07:55be a skill to help you eventually once
  208. 07:57we figure out all this like crazy stuff
  209. 07:59and this one tool. You just skip this
  210. 08:02tool. I think it's absolutely fine. Who
  211. 08:05knows what are you going to next? But if
  212. 08:07you skip your next piano practice, you
  213. 08:08skip your next soccer practice [music]
  214. 08:10and you give up on that, that's not fine
  215. 08:13because that is going to be eventually a
  216. 08:15long run define who you are. But not
  217. 08:17this tool, not just one tool.
  218. 08:20At this moment in my career as a
  219. 08:22educator, I started to care more about
  220. 08:24not that they learned this math. A lot
  221. 08:26of times when I teach, oh, you showed
  222. 08:28up, you listen to me, you're trying to
  223. 08:29go through it, but I bet maybe a year
  224. 08:31from now, you don't remember anything.
  225. 08:33But what you can remember is more about
  226. 08:36you are able to understand this at the
  227. 08:38moment. You are willing to come here to
  228. 08:40understand foundation. You are willing
  229. 08:41to open up the black box. That
  230. 08:43willingness is what set you apart from
  231. 08:45others. [music] Others would never try,
  232. 08:47never attempted, never took out a
  233. 08:49challenge. That's why differentiate.
  234. 08:51It's not really about how much you
  235. 08:52remember the equation about [music]
  236. 08:54transformer about attention mechanism.
  237. 08:56It's really about there was once upon a
  238. 08:59time I try hard to memorize this. I stay
  239. 09:02in the library for hours of study. I was
  240. 09:05successful. So next time when there's a
  241. 09:07learning challenge I can learn this. So
  242. 09:09that is more important about what
  243. 09:11differentiate that people who know
  244. 09:13foundation they implies the process time
  245. 09:16the person invested in learning it. So
  246. 09:18that is what I value versus person who
  247. 09:20never really learn foundation implies
  248. 09:22the lack of effort the lack of
  249. 09:24willingness to invest in time and effort
  250. 09:26to take on a challenging learning task.

  251. The 'AI-Native' Trap - AI can't change people, but people can change AI

  252. 09:32When I was teaching the intro to
  253. 09:34programming, you would really spend a
  254. 09:36lot of effort making new assignments a
  255. 09:38whole every semester because of check
  256. 09:41places that people share solutions and
  257. 09:43we have all these technical solutions.
  258. 09:46We're trying to check the IP addresses
  259. 09:48see whether people are accessing this
  260. 09:50and we even put some kind of a trap on
  261. 09:53the site. if somebody access that. We
  262. 09:56know they access this but at the time I
  263. 09:58was like okay well I hope Czech could
  264. 10:01get out of business maybe shut down by
  265. 10:04the government that they will help us
  266. 10:06educators and my dream came true because
  267. 10:09AI become a new treating tool and the
  268. 10:12check out the business [music] and then
  269. 10:13when I realized hey check was all the
  270. 10:15business my dream came true but problem
  271. 10:17is still there what's going on so it
  272. 10:19reminds me again we should keep thinking
  273. 10:22going back to the source there was the
  274. 10:24reason why people have to cheat at the
  275. 10:27first place. That is the main cause.
  276. 10:30It's not and check and AI just a
  277. 10:32symptoms. [music] So check is gone,
  278. 10:34people still cheat. I bet when AI is
  279. 10:36gone, people can still find way to
  280. 10:38cheat. Is this AI cheating that distract
  281. 10:40us from the bigger fundamental profit of
  282. 10:45the society's incentive system? Why
  283. 10:47[music] are student compelled to cheat?
  284. 10:50Why is it that this the system doesn't
  285. 10:52encourage real learning that has
  286. 10:54actually have spent time So when you
  287. 10:57hire somebody, what do I care about? Is
  288. 10:59does student have a good work ethics?
  289. 11:01What I care about? Is a student a good
  290. 11:03problem solver? Is this the person a
  291. 11:05team player that can actually
  292. 11:06communicate willing to work with others?
  293. 11:08[music] So when you hire somebody, we go
  294. 11:12back to those basic. That's what you
  295. 11:13actually care about. You want to keep
  296. 11:14those people as employees. And this AI
  297. 11:17thing is just going to be a byproduct of
  298. 11:20this. Think about it. When we hire
  299. 11:22someone because there's a problem solver
  300. 11:24in [music] order to solve problem that
  301. 11:26person is automatically just going to
  302. 11:28learn AI you don't have to tell them the
  303. 11:30reason why you had to force your AI
  304. 11:32native you had to go back maybe you
  305. 11:33didn't hire the right person you forgot
  306. 11:36to emphasize on person being a problem
  307. 11:39solver again similarly if you are hiring
  308. 11:42this person because this person team
  309. 11:43player because [music] the person is
  310. 11:45team player the person will learn how to
  311. 11:47use AI to facilitate team collaboration
  312. 11:49so you don't even have to tell the the
  313. 11:51person will automatically do that as
  314. 11:53well. Trust your instinct. [music]
  315. 11:54Continue to hire people like that
  316. 11:56because those people would automatically
  317. 11:58adopt AI. You're not a team player. AI
  318. 12:00not going to make you a team [music]
  319. 12:01player. You always look up your own
  320. 12:02interest. You do not respect others. AI
  321. 12:05not going to fix that. How can AI fix
  322. 12:07that? AI cannot change
  323. 12:10people only you. You [music] but you can
  324. 12:13change AI.