YouTube2Text

How I Use AI to Learn Things — Transcript

by Eero Alvar · 2,662 words · 425 segments · language en · Watch on YouTube

Full transcript

  1. 0:00How could we use AI to learn better? I
  2. 0:02mean, there's got to be some way to
  3. 0:04optimize learning with AI. I think it's
  4. 0:07the perfect tool for it. Uh it is not
  5. 0:10entirely clear yet how. So, yeah. Uh
  6. 0:13this video will be me sharing my own
  7. 0:15current approach to this. First, uh I'll
  8. 0:18go through the reasoning behind the
  9. 0:21approach and the design, and then we'll
  10. 0:24do a live demo. All right. So, this is
  11. 0:26the standard way of learning that we're
  12. 0:28used to. Each outlet uh so, a teacher, a
  13. 0:32book, a course, whatever, uh
  14. 0:35they all are designed to teach many
  15. 0:38students. And also, each students learns
  16. 0:42from many outlets. So, there's a
  17. 0:45many-to-many relation between the
  18. 0:47learners and the outlets, and both
  19. 0:50directions of it produce their own kind
  20. 0:53of inefficiency. First, one outlet
  21. 0:56teaches many. An outlet that's designed
  22. 0:59for many people cannot be optimal for
  23. 1:02any one of them. That's because optimal,
  24. 1:05quote-unquote optimal, teaching
  25. 1:07uh that would clearly depend entirely on
  26. 1:09the learner's current understanding. And
  27. 1:12this includes both the whole teaching
  28. 1:15arc from their current understanding to
  29. 1:18their goal understanding, as well as
  30. 1:20each explanation along the way. An
  31. 1:22optimal teaching path, then, uh would be
  32. 1:25one that minimizes teaching stuff that
  33. 1:28the learner already holds, and also
  34. 1:30stuff that they can't yet understand.
  35. 1:33So, it would work exactly at the edge of
  36. 1:36their understanding. So, ideally, a
  37. 1:39teacher would have exactly one student.
  38. 1:42All right. So, second direction, uh one
  39. 1:44student learns from many outlets. This
  40. 1:47means many teaching styles to get used
  41. 1:51to, many notations, many levels of
  42. 1:54reliability, many interfaces, and the
  43. 1:57switching between them uh costs mental
  44. 2:01effort that's not gone into learning the
  45. 2:04material. But, I think there's a much
  46. 2:06deeper cost in this, which is trust.
  47. 2:10With an unfamiliar outlet, the brain I I
  48. 2:14think the brain kind of hedges. So, it
  49. 2:16won't fully commit to accepting a fact
  50. 2:19until the outlet has proven itself and
  51. 2:22is familiar enough. I think a good way
  52. 2:24to illustrate this with is with the
  53. 2:27following example. Let's say we want to
  54. 2:29understand the hairy ball theorem, and
  55. 2:33we have two identical explanations. But,
  56. 2:35the other one is in a 3Blue1Brown video,
  57. 2:38and the other is on X posted by some
  58. 2:42random guy. Now, even though the
  59. 2:44explanations are identical, I think it
  60. 2:46is clear that we're going to learn a
  61. 2:48better from the familiar source. The
  62. 2:51brain is going to have a much easier
  63. 2:53time internalizing the information when
  64. 2:56it trusts the outlet, even though the
  65. 3:00explanation is identical.
  66. 3:03All right? So, ideally, a student would
  67. 3:06have exactly one teacher. Now, we've
  68. 3:08established that the ideal scenario is
  69. 3:11one-to-one in both directions, but
  70. 3:15there's an obvious objection to this,
  71. 3:18which is that having only one teacher
  72. 3:22means getting only one perspective. But,
  73. 3:25I think that the objection conflates a
  74. 3:27source with an interface. So, a teacher
  75. 3:31doesn't reduce the number of sources,
  76. 3:35perspectives.
  77. 3:36Uh instead, it aggregates all of them
  78. 3:39and delivers them through one interface.
  79. 3:42So,
  80. 3:43we don't lose many perspectives. Now,
  81. 3:47with AI being the teacher, trust isn't
  82. 3:49really built over time. Rather, it's
  83. 3:52engineered into the system. The reason
  84. 3:55we have to have reliable verification
  85. 3:58and fact-checking is one, correct
  86. 4:01information. Obviously, we absolutely
  87. 4:04don't want the AI to hallucinate false
  88. 4:08information. But also, two, cuz it makes
  89. 4:11learning easier when we know that the
  90. 4:14system is reliable. So, one interface
  91. 4:18fitted to one mind over all sources.
  92. 4:22Now, the two inefficiencies that we went
  93. 4:24over, they give us the two principles
  94. 4:27that the system runs on. First,
  95. 4:30optimized teaching, the answer to the
  96. 4:33first inefficiency. And two, optimized
  97. 4:37allocation of mental resources, answer
  98. 4:40to the second one. And this does not
  99. 4:43mean removing difficulty. Instead, it's
  100. 4:46about concentrating all cognitive work
  101. 4:49into the material itself. We want to
  102. 4:52maximize struggle. We want to learn
  103. 4:56difficult things, so struggling is very
  104. 4:58important, but it has to be in the
  105. 5:01material itself and not in logistics,
  106. 5:05planning, finding resources, verifying
  107. 5:07facts, figuring out what to learn and in
  108. 5:09what order. All of that is for the
  109. 5:11system to absorb. Next, the actual
  110. 5:14process,
  111. 5:16how this is implemented. First,
  112. 5:18optimized teaching, to teach optimally,
  113. 5:22whatever this might be very different
  114. 5:24from person to person, but in order to
  115. 5:28do that, the system obviously needs to
  116. 5:31know the person's exact current
  117. 5:34understanding. So, it has to measure,
  118. 5:36and it does this with a quiz tool, which
  119. 5:40lets it ask graded multiple-choice
  120. 5:42questions. It starts off with very broad
  121. 5:45of
  122. 5:46and then basically
  123. 5:48binary search is the edge on every
  124. 5:52possible strand that the lesson will
  125. 5:54depend on. So, it's going to get a very
  126. 5:57detailed map of the learner's
  127. 6:00understanding. So, that's phase one,
  128. 6:03probe. Second, phase two is plan.
  129. 6:07Basically just reasons everything out.
  130. 6:09How do I teach this mind this specific
  131. 6:12thing? And here's also where it fires
  132. 6:15off its first verification and
  133. 6:18fact-checking sub-agents. And then
  134. 6:21presents the plan as a mermaid graph and
  135. 6:26diagram. And okay, basically two reasons
  136. 6:29for why why it has to show a graph. One,
  137. 6:33it gives the learner a better sense of
  138. 6:36what's to come. And also two, the actual
  139. 6:39reason I implemented it is to actually
  140. 6:42force the AI to reason everything out.
  141. 6:46That's the real reason.
  142. 6:49So,
  143. 6:50it cannot just cheat and wing it. And
  144. 6:53then finally, the teaching itself. And
  145. 6:56this phase is is going to be very
  146. 6:58different cuz it's going to depend on it
  147. 7:00entirely how you want to learn. You want
  148. 7:02to install the system with your own
  149. 7:05learning philosophy and how you learn
  150. 7:07best. But one detail that I think is
  151. 7:11going to be important regardless of the
  152. 7:12exact way of learning is feedback. So,
  153. 7:17the AI is going to quiz you periodically
  154. 7:20to check whether you actually understood
  155. 7:22the thing. And this is important for
  156. 7:25three reasons. One, it's very easy to
  157. 7:28sort of gaslight yourself into thinking
  158. 7:31that you understood something,
  159. 7:33especially when learning with AI. So,
  160. 7:36actually testing your understanding is
  161. 7:39important feedback for yourself. Two,
  162. 7:41the system needs continuous feedback to
  163. 7:44stay calibrated. And three,
  164. 7:48applying the material, doing practice
  165. 7:50problems, using the stuff that you
  166. 7:53learned, also just helps you learn
  167. 7:55better. It's It's a part of how the new
  168. 7:57information and the new understanding
  169. 8:00locks in. So, yeah.
  170. 8:02Uh that's the That's the general idea.
  171. 8:04Now, let's Let's see it in action. All
  172. 8:08right, we're in the Pi Agent harness now
  173. 8:11in my learning directory. And yeah, this
  174. 8:14is where I've got everything set up. So,
  175. 8:16in the dot Pi folder, I've got the teach
  176. 8:18skill, some visualization stuff,
  177. 8:21uh the quiz extension, the MD log
  178. 8:24extension, which I'll show you in a bit,
  179. 8:27and two subagents to make visuals.
  180. 8:30Hopefully, we'll see those also. So,
  181. 8:33yeah. I think the only way to demo a
  182. 8:36system like this is actually to try to
  183. 8:38learn something. And to really go full
  184. 8:41circle, uh in my previous
  185. 8:43learning-related video, I mentioned the
  186. 8:46Maxwell's equations and how they can be
  187. 8:49expressed in just two equations using
  188. 8:53differential forms. And as you can see,
  189. 8:56uh I don't really know anything about
  190. 8:59differential forms. So, I think this is
  191. 9:01what we're going to learn today. Now,
  192. 9:03obviously, I don't We're not going to
  193. 9:05get to this level, but uh I think this
  194. 9:07is a very good example to showcase the
  195. 9:09system with a an actual learning
  196. 9:12process. So, yeah. Uh let's begin. Let
  197. 9:15me pull up Obsidian here cuz this is
  198. 9:17what I use.
  199. 9:19It's It's sort of like the UI for
  200. 9:22everything. Learning differential forms.
  201. 9:25This is what the MD log extension is
  202. 9:28for.
  203. 9:28It lets me link a Markdown file to the
  204. 9:31session, and then everything is going to
  205. 9:33get printed right here. Just a nicer way
  206. 9:37to see things and also get the LaTeX
  207. 9:39rendering. So, yeah, that's my solution.
  208. 9:42And yeah, it's also nice to have like
  209. 9:44persistence artifacts from each learning
  210. 9:47session. So, yeah,
  211. 9:49let's begin teach. I want to get like a
  212. 9:52solid introduction to differential
  213. 9:54forms.
  214. 9:55And we're using Kimik 3 on Macs cuz I
  215. 10:00really found that the intelligence of
  216. 10:02the model really matters
  217. 10:04in teaching. The teaching instructions
  218. 10:07are quite
  219. 10:09They're quite specific.
  220. 10:10Anyway, [snorts]
  221. 10:12let's see how this goes. We're going to
  222. 10:13get questions now.
  223. 10:15A force field F acts on a particle as it
  224. 10:18moves along a curve C. What does the
  225. 10:21line integral
  226. 10:23compute?
  227. 10:24Let me also I've got this note here
  228. 10:26where I can just yap into. Basically,
  229. 10:29just to give it more context and sort of
  230. 10:32talk through my reasoning.
  231. 10:34It's not going to be number one.
  232. 10:36I mean, number one does hold for
  233. 10:38conservative fields, but for any force
  234. 10:41field F, it's going to be number three.
  235. 10:44The network done by the field on the
  236. 10:47particle.
  237. 10:50All right, next one. The divergence of a
  238. 10:53vector field at a point measures the net
  239. 10:56outward flux per unit volume at that
  240. 10:59point.
  241. 11:02All right, now the Stokes theorem. Yes.
  242. 11:07And now we get Faraday's law of
  243. 11:09induction. I mean, you can always just
  244. 11:10give it more context at the beginning
  245. 11:13about what stuff you already understand
  246. 11:17very thoroughly.
  247. 11:18In this case, I gave you very little
  248. 11:19context, so we're going to get a long
  249. 11:22probing phase.
  250. 11:26Oh we're going to have to go into
  251. 11:27relativity.
  252. 11:29In special relativity, when you change
  253. 11:31to a moving reference frame, what
  254. 11:33happens to the electric and magnetic
  255. 11:35fields? Okay, let's let's read through
  256. 11:38these. Both are invariants. All
  257. 11:40observers measure the same E and B at
  258. 11:43each event. They mix. A purely electric
  259. 11:46field in one frame has both electric and
  260. 11:48magnetic parts in another. Hold on. Only
  261. 11:51E changes. B stays fixed.
  262. 11:54No, obviously not. Three, each
  263. 11:57transforms independently.
  264. 12:00And I'll I'll go ahead and answer I
  265. 12:01don't know. And it was two.
  266. 12:04They mix.
  267. 12:06All right, it's stopped with questions.
  268. 12:09Um so yeah, uh finally done with the
  269. 12:12probing phase.
  270. 12:14Uh it did ask a lot of questions, but
  271. 12:17now it's got a very detailed image of my
  272. 12:19understanding. And it's also just a nice
  273. 12:22warm-up as well. And I don't really got
  274. 12:24to worry about anything. Just answer the
  275. 12:26questions and it handles all the
  276. 12:29logistics. Now we've got a researcher
  277. 12:31still going on fact-checking stuff.
  278. 12:34Although usually math isn't really
  279. 12:36something it would need to fact-check.
  280. 12:39But well, it's it's good practice
  281. 12:41anyway. Then after this, it's going to
  282. 12:43fully plan everything out. All right,
  283. 12:45finally, we're getting the plan.
  284. 12:50I did start rendering
  285. 12:52the mermaid in the 2A, but
  286. 12:55don't know what happened. It was
  287. 12:56probably
  288. 12:57in the most recent release, I believe,
  289. 12:59added
  290. 13:01mermaids. But anyway, we get it in the
  291. 13:03obsidian. This is the plan. Let's go. A
  292. 13:07lot of things I'm going to be reading
  293. 13:09reading the plan. All right, so that's
  294. 13:11cool. We're going to get to generalized
  295. 13:13Stokes well before the actual goal.
  296. 13:16Okay, perfect. It's it's using the
  297. 13:18visualization skill. So we're going to
  298. 13:20get some visuals soon. All right, it's
  299. 13:23making an SVG. And the reason that these
  300. 13:26are done in sub agents is to obviously
  301. 13:29preserve context but also
  302. 13:32the sub agents will look at the image to
  303. 13:35verify whether it actually looks
  304. 13:38correct. It wrote an SVG and now
  305. 13:41it's going to view it. It's going to
  306. 13:43edit it a bit, look at it again.
  307. 13:47Yes. So, we're going to get that in the
  308. 13:49next message. But now the first
  309. 13:52node. So, it's going to It's going to
  310. 13:55slowly
  311. 13:56walking down the path one reasoning step
  312. 14:00at a time. Because what usually happens
  313. 14:03if you're like talking to chat GPT or
  314. 14:05something trying to get it to explain or
  315. 14:07teach you anything it usually just it's
  316. 14:10way too excited and it rushes through
  317. 14:13the whole thing. And so, I like to keep
  318. 14:14this very slow, one reasoning step at a
  319. 14:18time. So, now we're introducing
  320. 14:20co-vectors. Let me
  321. 14:22read through this.
  322. 14:24All right. So, now it's introduced
  323. 14:26co-vectors
  324. 14:27and also given a new perspective on the
  325. 14:31X.
  326. 14:32This is really cool. Yeah, it's going to
  327. 14:34quiz me on the this reasoning step to
  328. 14:38confirm that everything is actually
  329. 14:39understood, to confirm my understanding.
  330. 14:41Now, yeah, it's going to incorporate the
  331. 14:44visual
  332. 14:45which will be all rendered into Obsidian
  333. 14:48and automatically here as an embed
  334. 14:52embedded file. Let's alpha be 3D X minus
  335. 14:572D Y and V this. What is alpha of V?
  336. 15:01Which is going to be -4.
  337. 15:04Okay, this makes perfect sense.
  338. 15:06The picture helped. Oh, yeah, the
  339. 15:09Pi also got like built-in LaTeX
  340. 15:12rendering. Okay, which is kind of cool.
  341. 15:13Yeah, it does the job but
  342. 15:16it's definitely not the same as this.
  343. 15:20But yeah,
  344. 15:21cool features though. All right, so now
  345. 15:23we're extending
  346. 15:25the covectors to a covector field. All
  347. 15:28right, so now we get a new perspective
  348. 15:30on what the line integral actually is.
  349. 15:33Yeah.
  350. 15:34One thing I don't like about this is
  351. 15:37it's I haven't really configured its
  352. 15:41style of speaking. So we get a lot of it
  353. 15:44sounds very AI, which I don't really
  354. 15:47mind cuz I don't want to cram it with
  355. 15:49too many instructions to worry about. So
  356. 15:53I'm fine with these LLM-isms.
  357. 15:56It's not X, it's Y.
  358. 15:59Wow, insane. The quality of the teaching
  359. 16:02is what matters. But yeah, it's
  360. 16:04basically walking down the dependency
  361. 16:07tree. Seems like we're still here.
  362. 16:10So when we're going to get I think this
  363. 16:11is the wedge product. But yeah, I like
  364. 16:14that it moves one reasoning step at a
  365. 16:17time. So if at any point I have
  366. 16:18questions, I can always ask.
  367. 16:21And it doesn't rush forwards. The each
  368. 16:25step is very easy to digest. And it's
  369. 16:28going to give me everything that I can
  370. 16:30accept at face value like this. All
  371. 16:32integration over a one-dimensional thing
  372. 16:35is the integration of a one-form. Now, I
  373. 16:38assume that this continues, yeah.
  374. 16:41A K-form will be the kind of thing a
  375. 16:44K-dimensional surface can eat.
  376. 16:47Sure.
  377. 16:49All right, we're getting a new new SVG.
  378. 16:54Seems like our
  379. 16:56uh SVG maker
  380. 16:58died. Agent error. Overloaded. Bro, not
  381. 17:03now. Trying to learn differential forms.
  382. 17:05Oh no, we're getting it. Now we're
  383. 17:06getting the wedge products. That's cool.
  384. 17:10Machines that are bilinear and
  385. 17:11antisymmetric. Yeah, where do we get
  386. 17:13them? I'm sure this is not
  387. 17:15antisymmetric. The minimal fix is the
  388. 17:18oldest trick in the book. Aha. Yeah,
  389. 17:21that gives something anti-symmetric.
  390. 17:24So, is that just what it is? Sure, okay,
  391. 17:27fine. But yeah, anyway, this is we're
  392. 17:29obviously not going to go
  393. 17:31through the whole dependency graph here.
  394. 17:35Um this is going to be this one, I
  395. 17:37believe. Yeah.
  396. 17:40But yeah, anyway, where did we get to in
  397. 17:42the DAG? Here.
  398. 17:44Yeah, I think we got here. Yeah, we're
  399. 17:46approaching generalized Stokes now.
  400. 17:49Yeah, that's cool. Um but yeah, I think
  401. 17:53this is enough for a demo. So, this has
  402. 17:55been sort of a second part or
  403. 17:57continuation of where we left off last
  404. 17:59time in my previous learning related
  405. 18:01video. So, what this video has been
  406. 18:04really about is how to take a learning
  407. 18:07philosophy or way of teaching and
  408. 18:10implement that as an AI system. I think
  409. 18:13the main reason for why this has worked
  410. 18:15so well for me is none of the stuff I
  411. 18:19talked here about, rather just the style
  412. 18:21of teaching I covered in the previous
  413. 18:23video, which affects exactly two things:
  414. 18:26the learning arc, the path that we're
  415. 18:30going to take from our current
  416. 18:31understanding to our goal understanding,
  417. 18:33and two, the individual steps and
  418. 18:36explanations along the way.
  419. 18:39Yes. Anyway, just some ideas for you for
  420. 18:43you to think about. I'd like to hear
  421. 18:45your your thoughts on this. And how how
  422. 18:46how could the system be improved? How do
  423. 18:49you use AI in learning? I'd like to
  424. 18:51know. I'm very invested into refining
  425. 18:55this further. So, yeah, uh that's it.

About this transcript

This page contains the full transcript of How I Use AI to Learn Things by Eero Alvar, generated from the public captions YouTube serves with the video. The transcript has 2,662 words across 425 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.