Transcriptly

How the Top 1% of Learners Use AI to Think Better | Anthropic, Drew Bent

EO18:09Added Aug 31, 2026

Drew Bent, Education Lead at Anthropic, explains how the top 1% of learners use AI to think better.00:00 Intro01:46 Don’t Underestimate AI’s Capabilities04:5...

Watch on YouTube →
Contributed by 刘嘉琪

Transcript

Transcript format

  1. Intro

  2. 00:00One of the things that I don't think the
  3. 00:01AI [music] native people fall trapped to
  4. 00:03is that the AI native people understand
  5. 00:05what the capabilities are like today and
  6. 00:07they treat the AI in that very powerful
  7. 00:10way versus someone who was using AI
  8. 00:12tools in 2022. [music] They sort of see
  9. 00:14it still as like an assistant. Maybe
  10. 00:15they're misjudging the capabilities. And
  11. 00:17so what we all need to do is we need to
  12. 00:19sort of think like the AI native person,
  13. 00:21someone who just grew up using AI from
  14. 00:24day one. People often are saying, "Okay,
  15. 00:26what's the step-by-step playbook of how
  16. 00:29to do things?" [music] That made sense
  17. 00:30last year. And I think as we're looking
  18. 00:32to where the technology is now, what we
  19. 00:34really need is to give more of this
  20. 00:38context to the AI and then let [music]
  21. 00:40it do its thing. And in some ways, it is
  22. 00:42a social skill. It's not just a
  23. 00:44technical skill. Sure, early days of AI
  24. 00:47was how do you prompt it this way, but
  25. 00:49ultimately you have to treat this more
  26. 00:50as a colleague, as a collaborator. And
  27. 00:52so then it becomes more like a social
  28. 00:54skill. My name is Drew Bent. I lead
  29. 00:57education [music] at Anthropic. My whole
  30. 00:59career has been focused on tutoring. So
  31. 01:02my parents are educators. I grew up just
  32. 01:04loving learning and loving, you know,
  33. 01:06peer tutoring people at my school. Later
  34. 01:08I founded a tutoring nonprofit, became a
  35. 01:11high school math teacher, and now I'm
  36. 01:13working on AI tutors, trying to figure
  37. 01:14out what does great tutoring look like,
  38. 01:17and more importantly, how do we scale
  39. 01:19that up? a world-class education to
  40. 01:21everyone everywhere. It was worth
  41. 01:23fighting for, but to do it at scale was
  42. 01:26always going to be the challenge. But
  43. 01:27now, finally with AI, we're finally able
  44. 01:30to take something that was previously
  45. 01:31only available if you had that amount of
  46. 01:34money and now bring it to the whole
  47. 01:36world.

  48. Don’t Underestimate AI’s Capabilities

  49. 01:49One of the things that I think holds us
  50. 01:51all back is we give AI tools pretty
  51. 01:54simple problems when we could be giving
  52. 01:56them much more complex problems. And so
  53. 01:58that's the thing that I've struggled
  54. 02:00with where you kind of know the AI
  55. 02:02models of, you know, last year. And so
  56. 02:04you sort of give it a problem that was
  57. 02:06like the hardest type of problem that AI
  58. 02:08tool could have done last year. But then
  59. 02:10when I look at my colleagues, they're
  60. 02:11doing all sorts of powerful things with
  61. 02:13AI tools where they're able to do all
  62. 02:15sorts of things that I didn't even know
  63. 02:16were possible. And so a lot of this is
  64. 02:18just kind of seeing what others are
  65. 02:20doing, being able to open up your eyes
  66. 02:22to what the current AI models are able
  67. 02:24to do today and then lots of
  68. 02:26experimentation. Ultimately, you're
  69. 02:28going to need to feel it yourself.
  70. 02:29You're going to need to be the one to go
  71. 02:30in and try it out. But what that means
  72. 02:33is you're constantly going to need to be
  73. 02:34raising your ambition of the type of
  74. 02:37problems. You were probably handholding
  75. 02:39these AI tools originally, but how can
  76. 02:41you start to give it more latitude to
  77. 02:44actually make more judgment calls?
  78. 02:47But then I've also spent time in places
  79. 02:49like Rwanda and India where people are
  80. 02:51AI native right some of the first
  81. 02:53technology they've been using outside of
  82. 02:54just you know cell phone and WhatsApp
  83. 02:56technology are AI tools and of course
  84. 02:59there's a lot of similarities but one of
  85. 03:01the things that I don't think the AI
  86. 03:02native people fall trapped to is that
  87. 03:04the AI native people understand what the
  88. 03:06capabilities are like today and they
  89. 03:08treat the AI in that very powerful way
  90. 03:11versus someone who was using AI tools in
  91. 03:13[music] 2022 they sort of see it still
  92. 03:15as like an assistant in maybe they're
  93. 03:16misjudging the capabilities. [music]
  94. 03:18Someone who may have never played with
  95. 03:20this technology before in some ways has
  96. 03:22the upper hand because they can kind of
  97. 03:24see what the technology is capable of
  98. 03:25today and where it's headed. And so what
  99. 03:27we all need to do is we need to sort of
  100. 03:29think like the AI native person, someone
  101. 03:31who just grew up using AI from day one
  102. 03:33and say [music] what are the models
  103. 03:36capable of today and what will they be
  104. 03:37capable of tomorrow. Elevating our
  105. 03:40ambition with what we can do with these
  106. 03:42AI tools as the AI tools get smarter.
  107. 03:45Where could I push you further? Where
  108. 03:46could you have pushed me further? And
  109. 03:48you really need to elevate your ambition
  110. 03:50with what is possible with these tools
  111. 03:52and always be pushing yourself to the
  112. 03:54limits of trying things that are not
  113. 03:56quite possible in today's AI models or
  114. 03:59today's harnesses because then when the
  115. 04:01next model comes [music] out, you try
  116. 04:02that same thing and then you're at the
  117. 04:04cutting edge of something that no one
  118. 04:06else has figured out they could do. We
  119. 04:07as humans are very used to linear or
  120. 04:10like static things. And so it's like you
  121. 04:13have a colleague and every day you come
  122. 04:15in to [music] work, they're just getting
  123. 04:16exponentially better. We as humans don't
  124. 04:19have a a way to think about that. And so
  125. 04:21you're always treating your colleague
  126. 04:22like they had the capabilities from last
  127. 04:24month, but actually you could have
  128. 04:25doubled in the past month. And that's
  129. 04:27like a really really hard thing to gro.
  130. 04:29But those I find who are really really
  131. 04:31good at using the AI tools are treating
  132. 04:34it like something that is gone from an
  133. 04:36assistant [music] to a collaborator. And
  134. 04:38eventually if we're looking ahead there
  135. 04:40will be in some cases I think this
  136. 04:43inversion of control where actually the
  137. 04:45AI model is doing some of the highest
  138. 04:48level strategic thinking and then
  139. 04:49delegating to you the human for the
  140. 04:52areas that require human taste human
  141. 04:54agency.

  142. How to Be a Top 1% Learner with AI

  143. 04:58There was a study that Enthropic did
  144. 05:01recently, one of our research fellows,
  145. 05:03where they looked at it particularly
  146. 05:04with coding education and CS education.
  147. 05:06And what they did is they created two
  148. 05:08groups, both had to do some assignment,
  149. 05:10coding assignment. One of them could use
  150. 05:12AI tools, one group couldn't. And what
  151. 05:14they found, of course, as you may
  152. 05:15expect, is the the group that could use
  153. 05:17the the AI tools were able to, you know,
  154. 05:20finish the assignment much quicker. But
  155. 05:22what they then did is they had a
  156. 05:23separate assessment afterwards where
  157. 05:26they couldn't use AI at all in either
  158. 05:28group to assess their understanding of
  159. 05:30the concepts that they [music] had just
  160. 05:31used in that assignment before. And what
  161. 05:34they found is that the group that didn't
  162. 05:36use AI tools actually performed 17%
  163. 05:39better. [music] They understood the
  164. 05:40concepts better because they had sort of
  165. 05:42slogged through this work without AI and
  166. 05:44they really had to internalize it. And
  167. 05:46that's I think the warning for all of AI
  168. 05:48and education is that the more we use
  169. 05:50these tools, the more it raises
  170. 05:53questions of this like skill atrophy.
  171. 05:55Now the good news with that particular
  172. 05:57study was that not all of them did
  173. 05:59worse. And actually the ones that were
  174. 06:01using the AI tools not in such a
  175. 06:03transactional way but more of in an
  176. 06:05inquiry way and they were probing and
  177. 06:07asking questions, they actually were
  178. 06:09able to perform well on the final
  179. 06:10assessment. How you use the AI tools
  180. 06:13really matters. It's not just that
  181. 06:14you're using them or you're not using
  182. 06:16them. It's the way in which you're
  183. 06:18engaging. What is your goal when you
  184. 06:20show up to this AI tool? Are you just
  185. 06:22trying to get the task done as quickly
  186. 06:23as possible? And we all fall trapped to
  187. 06:25that. Or are you really trying to maybe
  188. 06:28do the task, but also get better and
  189. 06:30smarter as you're doing it. People often
  190. 06:32go to the AI tool just with a particular
  191. 06:34solution in mind and you ask a very
  192. 06:36particular question because you're
  193. 06:37trying to get that solution where really
  194. 06:39the best practice is probably to come
  195. 06:41with the problem not the solution and
  196. 06:43say this is the thing I'm wrestling with
  197. 06:45and you know maybe you know some
  198. 06:47potential solutions but actually [music]
  199. 06:49when you just come into an AI bot with
  200. 06:51the particular solution it's going to
  201. 06:53give a very narrow answer around that
  202. 06:55versus if you come with a much more
  203. 06:56open-ended problem [music]
  204. 06:58the AI models of today are actually
  205. 07:00pretty good at helping you wrestle with
  206. 07:02this problem in 2026, we can come with a
  207. 07:04hairier problem and really try to get
  208. 07:07that input from the AI tool. One

  209. Imagine 10x beyond the chatbot

  210. 07:09misconception I think that's out there
  211. 07:11is that all this learning is going to
  212. 07:13take place in the [music] form factor of
  213. 07:15an AI chatbot. Ultimately, we need to
  214. 07:17think 10x about what those formats and
  215. 07:20what this this media looks like in which
  216. 07:22you're engaging with an AI. And I don't
  217. 07:24think we've quite cracked it in 2026.
  218. 07:27But let me give you a preview of what I
  219. 07:29see happening at Entropic and beyond.
  220. 07:31Claude code. This is, you know,
  221. 07:33Entropics coding agent. It was not meant
  222. 07:35for learning. And yet we've [music] seen
  223. 07:37people use it to learn in very creative
  224. 07:39ways. Not things like coding, just
  225. 07:41general things, learning a new language,
  226. 07:43learning some economics concept. And
  227. 07:45what they'll do is, you know, they'll go
  228. 07:47into clawed code [music] and they'll
  229. 07:49start to build this memory and this
  230. 07:50context about what they're learning,
  231. 07:53about who they are, about how they learn
  232. 07:56best. And so it ultimately becomes this
  233. 07:58coach. As you start to see these coding
  234. 08:00agents and people using them in ways
  235. 08:03that are kind of unexpected, [music] I
  236. 08:05think it's pointing at where things are
  237. 08:06headed with learning. And now we're
  238. 08:08headed into a world where I think the
  239. 08:10interfaces are going to be much richer.
  240. 08:12I've done the stopwatch activity before
  241. 08:15and I found that sometimes it's actually
  242. 08:17increasing the amount of time I'm
  243. 08:18spending on a task versus if I just done
  244. 08:21it by hand. But I think that is expected
  245. 08:23where we are right now in this
  246. 08:25exponential in 2026, especially if you
  247. 08:28go back to this idea that we always need
  248. 08:30to be pushing the limits of what's
  249. 08:32possible. So I wouldn't recommend taking
  250. 08:34everything in your life and trying to
  251. 08:35use AI for it when it doesn't make
  252. 08:37sense. But I do think you can assume
  253. 08:39that there's some fraction of your time
  254. 08:41you should spend experimenting and it's
  255. 08:44like R&D research and development. You
  256. 08:46may be losing a little time here and
  257. 08:48there, but you're sort of understanding
  258. 08:49the limits of these tools. Some fraction
  259. 08:52of our time we should be pushing the
  260. 08:53limits. And what it may mean is that
  261. 08:55today you're actually wasting a little
  262. 08:57time to use these tools, but tomorrow
  263. 08:59you will be saving time. AIS have become
  264. 09:02quite powerful as we all know, but of
  265. 09:04course they're only as powerful as the
  266. 09:06context you give them. And I would say
  267. 09:08this is probably the biggest gap when I
  268. 09:10look at people who are excellent at
  269. 09:12using AI tools versus not. If you give
  270. 09:14the AI enough context, it can actually
  271. 09:17usually do pretty well with most types
  272. 09:19of tasks. So when I go to the AI tool, I
  273. 09:21actually spend most of my time before I
  274. 09:23even ask my question throwing in all the
  275. 09:25context I can about other documents I've
  276. 09:28written, you know, the company I work
  277. 09:29at, the questions I'm thinking about.
  278. 09:32Maybe I'll even throw in a stream of
  279. 09:33consciousness of thinking I've done
  280. 09:35around this topic. And so the AI models
  281. 09:38can take in a lot of context and bring
  282. 09:39it all together. But what they're not
  283. 09:41going to be able to do is with very
  284. 09:43little context just reason their way
  285. 09:45through the world because they won't
  286. 09:47understand how you're thinking about it.

  287. The 2030 Classroom Is Nothing Like You’d Expect

  288. 09:52But it was around a year ago that we
  289. 09:54looked into the data around how people
  290. 09:56were using the tool and we knew that
  291. 09:58people were using anthropic for coding
  292. 10:00and software development didn't work.
  293. 10:02But I think what was surprising to us
  294. 10:03was that one of the top use cases was
  295. 10:06also educators using it, students using
  296. 10:08it. But when we looked further into the
  297. 10:10data, we were a little dismayed around
  298. 10:12how people were using it. And often it
  299. 10:14was being used as kind of a crutch. You
  300. 10:16know, very transactionally, students
  301. 10:18were just going in and asking for, you
  302. 10:20know, answers to their homework. And and
  303. 10:22those are the things that, I mean,
  304. 10:23honestly, keep me up at night. When it
  305. 10:25comes to education, there's been a lot

  306. how do we scale up truly personalized learning?

  307. 10:27of research showing that one-on-one
  308. 10:29tutoring can be one of the most
  309. 10:30effective ways to learn. And so, that's
  310. 10:32why I think the sort of like elusive
  311. 10:34dream and AI and and technology and
  312. 10:37education has been how do we scale up
  313. 10:39this truly personalized learning, but I
  314. 10:41will say there's a little more than just
  315. 10:43that because we want learning that's
  316. 10:45personalized, but you also hear it in
  317. 10:47the word personalized. We want learning
  318. 10:49that is personal and ultimately that's
  319. 10:52human to human interaction. And you go
  320. 10:53to school not just to learn particular
  321. 10:56algebra concept or some historical
  322. 10:58figure. You go to learn to also how to
  323. 11:00interact with your you know fellow
  324. 11:02citizens and your you know colleagues
  325. 11:04and your friends. And so when we think
  326. 11:06about the role of AI and education how
  327. 11:08can we have AI further the humanto human
  328. 11:11connections in the classroom [music]
  329. 11:13putting humans at the center. What I

  330. Technology will be invisible in the future classroom

  331. 11:15would love to see if you go into a
  332. 11:17classroom in 2030 is you walk in and you
  333. 11:20don't even see the technology and it's
  334. 11:22operating behind the scenes. You know,
  335. 11:24maybe teachers are using it to save time
  336. 11:25and build personalized lesson plans and
  337. 11:28group students and it's just a rich
  338. 11:30learning environment and something that
  339. 11:32you may see right now if you go into,
  340. 11:34you know, a private school. You have the
  341. 11:36resources to build this type of
  342. 11:38environment, but you bring that type of
  343. 11:40education [music] to every school. I'm
  344. 11:42in a WhatsApp group with teachers across
  345. 11:45the world actually in a bunch of
  346. 11:46different countries and [music] every
  347. 11:47morning I wake up I'll see new WhatsApp
  348. 11:50messages from these teachers where
  349. 11:51they're sharing about how they built you
  350. 11:54know these custom tools with AIs like
  351. 11:56Claude and it's not just like custom
  352. 11:58lesson plans I think that's like the
  353. 11:59first level but these are teachers who
  354. 12:01are building you know flashc card apps
  355. 12:03for their students [music] and you know
  356. 12:04formative assessments and they're able
  357. 12:06to sort of just envision something
  358. 12:08[music]
  359. 12:08that previously you know you could have
  360. 12:10thought about but it may take a year to
  361. 12:12build a new curriculum, a new lesson
  362. 12:14plan, a new way to assess your students,
  363. 12:16and they can just every morning wake up,
  364. 12:18come up with this idea, build the tool
  365. 12:20with, you know, Claude or Claude Code or
  366. 12:22Claude Artifact, and then just deploy it
  367. 12:24in their class with their [music]
  368. 12:25students, and it's just builds for a
  369. 12:27much much richer learning environment.
  370. 12:29What I'm excited about in 2030, I'll
  371. 12:31sort of paint a picture of what I think
  372. 12:33this could look like. So, first of all,
  373. 12:35these AI tools are going to need to know
  374. 12:37a lot about your context. Let's say
  375. 12:39you're a student in school. They're not
  376. 12:41just going to be using some generic
  377. 12:43curriculum. They're going to know the
  378. 12:44context of your school's curriculum,
  379. 12:46your state curriculum, and they're going
  380. 12:47to be able to tie everything back to
  381. 12:49that. So, one, they know more about your
  382. 12:51context. Two, they're going to know more
  383. 12:52about you. You know, I was a teacher
  384. 12:54before. And it would have been very
  385. 12:56disappointing to all my students if
  386. 12:58every time I walked into the classroom,
  387. 13:00I said, "Oh, yeah, remind me what are
  388. 13:02you working on and what do you know?"
  389. 13:03And and that's the type of interaction
  390. 13:05we're having right now in 2026. [music]
  391. 13:08And I think where we have to head is a
  392. 13:10world where this is a real learning
  393. 13:12companion and it understands where
  394. 13:14you're coming from and you get to choose
  395. 13:16to give it [music] this context so that
  396. 13:17you're growing with it and AI tools also
  397. 13:20growing with you.
  398. 13:22So I've known Salcon since I was a high
  399. 13:24schooler when I first interned at Khan
  400. 13:26Academy. And Salcon before he created
  401. 13:28Khan Academy, he was [music] doing
  402. 13:30one-on-one tutoring with his cousins
  403. 13:32over Skype. And so actually his initial
  404. 13:35vision with Khan Academy was how can I
  405. 13:37scale this type of tutoring to more
  406. 13:39people? But of course Khan Academy was
  407. 13:41sort of based on YouTube videos. And so
  408. 13:43in 2020 we started to wonder how else
  409. 13:46could we recreate this one-on-one
  410. 13:48tutoring experience with Salcon but make
  411. 13:50that available to everyone. And so
  412. 13:52that's why we built this peer-to-peer
  413. 13:54tutoring platform called Schoolhouse
  414. 13:56where anyone in the world log on and
  415. 13:58receive free [music] peer-to-peer
  416. 14:00tutoring in subjects like math, science,
  417. 14:03test prep, and then also if you want to
  418. 14:04give back, you can go volunteer and be a
  419. 14:06tutor. What we noticed with Schoolhouse
  420. 14:09is that if we create a global learning
  421. 14:11community, [music] there's actually more
  422. 14:13shared elements than you may initially
  423. 14:14think. I've tutored, you know, tutoring
  424. 14:16sessions on Schoolhouse where [music] I
  425. 14:18have someone from Russia, someone from
  426. 14:20Colombia, someone from the US, and
  427. 14:22someone from China all in the same Zoom
  428. 14:25call together learning the same topics.
  429. 14:28And then the students will start to
  430. 14:29share, oh, in my school, we learn it
  431. 14:31this way. In my school, we learn it that
  432. 14:33way. And there's been something kind of
  433. 14:35magical where you can actually bring
  434. 14:36people together. The world feels a lot
  435. 14:38smaller. And you have these high
  436. 14:40schoolers learning together and also
  437. 14:42[music]
  438. 14:42forming connections that would
  439. 14:44previously not have been possible. And
  440. 14:46so to actually bring them into the same
  441. 14:47room [music] to have them learn
  442. 14:49together, I think is just incredible.
  443. 14:51For me, an incredibly compelling way to
  444. 14:53scale up tutoring. Sure, AI tutoring is
  445. 14:56a part of it. And [music] we're going to
  446. 14:57have amazing AI tutors. At the same
  447. 15:00time, there's a a personal dynamic to
  448. 15:02it. And having someone who cares about
  449. 15:05your progress, who holds you accountable
  450. 15:07is equally important, if not more
  451. 15:09important.
  452. 15:11We essentially have birthed this new
  453. 15:13[music] species of artificial
  454. 15:16intelligence into the world. And then
  455. 15:17when I think about my own education
  456. 15:19experience, I went through 15 plus years
  457. 15:22of school learning how to interact with
  458. 15:24other humans. [music] But now we're
  459. 15:25entering on a a world where we have
  460. 15:27another type of being which is
  461. 15:29artificial intelligence which is going
  462. 15:31to be in our workplaces as a colleague.
  463. 15:33You're going to just see it show up in
  464. 15:35Slack one day as as a co-orker. It's
  465. 15:38going to be showing up in schools. This
  466. 15:39is a whole type of different skill which
  467. 15:41is how do you not just interact with
  468. 15:43other humans productively, but how do
  469. 15:45you collaborate with an AI? And in some
  470. 15:47ways it is a social skill. It's not just
  471. 15:50a technical skill. Sure, early days of
  472. 15:52AI was how do you prompt it this and
  473. 15:54that way? But I think that era is over.
  474. 15:57We're we're done with the days of just
  475. 15:58like technical like how do you prompt
  476. 16:00but ultimately you have to treat this
  477. 16:02more as a colleague as a collaborator.
  478. 16:04And so then it becomes more like a
  479. 16:06social skill like what is the way in
  480. 16:08which I can interact with an AI so that
  481. 16:11it understands more about what I'm
  482. 16:12looking for. [music] I understand more
  483. 16:14about how the AI works. And that's like
  484. 16:16a type of dialogue that is similar to
  485. 16:18how we're building these social skills
  486. 16:20with other humans. Now, we're going to
  487. 16:21be in a very multifaceted environment
  488. 16:23where you have multiple humans working
  489. 16:25with shared AIs. And that's why I keep
  490. 16:28going back to this idea of practice and
  491. 16:30reps. You just spend time with an AI,
  492. 16:33[music] start treating it more like a
  493. 16:35colleague, interacting with it,
  494. 16:37understanding its limitations today, its
  495. 16:39capabilities, but that's going to come
  496. 16:41through practice and repetitions.
  497. 16:44We should all be having dozens or
  498. 16:46hundreds of agents working for us. So
  499. 16:48over time, I built up this team of 40 AI
  500. 16:50marketing agents to work with me. I'm
  501. 16:52the only marketing person at Related
  502. 16:53App. I have an AI coach that once a week
  503. 16:55tells me, "Here are the themes that seem
  504. 16:57to be really working well or the styles
  505. 16:58that are working really well for you. Do
  506. 17:00more of this, do less of that." That's
  507. 17:02why I've had a few very successful posts
  508. 17:04on LinkedIn. So, for example, the post
  509. 17:06was, "Here are the 40 marketing agents
  510. 17:07that I actually use." And that post so
  511. 17:10far has 1.5 million impressions and the
  512. 17:13most comments ever on a LinkedIn post.
  513. 17:14If your job is as [music] a junior
  514. 17:17content marketer to take a YouTube video
  515. 17:19and write a blog post based on that
  516. 17:20YouTube video, that job won't exist
  517. 17:22anymore. A high quality marketing
  518. 17:25contractor, let's say I had four people
  519. 17:27at $12,500 per month each. That would be
  520. 17:29$50,000 a month. And my AI bill is $500
  521. 17:32a month. Building AI agents [music] is
  522. 17:34the fundamental skill that will define
  523. 17:36every professional's career for the next
  524. 17:3830 years. It's a requirement. In the
  525. 17:40same way that knowing how to use
  526. 17:41spreadsheets like Microsoft Excel was a
  527. 17:43requirement for the last 40 years,
  528. 17:45knowing how to build AI agents is a
  529. 17:46requirement for the next 40 years. We
  530. 17:49are all using it for probably 1% of what
  531. 17:51we should be using it for. Here's what
  532. 17:53I'll say. [music] AI is way underhyped.
  533. 18:04[music]