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

1. 0:00 – 1:46

   Intro

   1. SPSpeaker0:00

      One of the things that I don't think the AI native people fall trap to is that the AI native people understand what the capabilities are like today, and they treat the AI in that very powerful way. Versus someone who was using AI tools in 2022, they sort of see it still as like an assistant. Maybe they're misjudging the capabilities. And so what we all need to do is we need to sort of think like the AI native person, someone who just grew up using AI from day one. People often are saying, "Okay, what's the step-by-step playbook of how to do things?" That made sense last year, and I think as we're looking to where the technology is now, what we really need is to give more of this context to the AI and then let it do its thing. And in some ways, it, it is a social skill. It's not just a technical skill. Sure, early days of AI was how do you prompt it this and that way, but ultimately you have to treat this more as a colleague, as a collaborator. And so then it becomes more like a social skill. My name is Drew Bent. I lead education at Anthropic. My whole career has been focused on tutoring. So my parents are educators. I grew up just loving learning and loving, you know, peer tutoring people at my school. Later, I founded a tutoring nonprofit, became a high school math teacher, and now I'm working on AI tutors, trying to figure out what does great tutoring look like, and more importantly, how do we scale that up? A world-class education to everyone everywhere, it was worth fighting for, but to do it at scale was always gonna be the challenge. But now finally with AI, we're finally able to take something that was previously only available if you had that amount of money, and now bring it to the whole world. [keyboard clicking]

2. 1:46 – 4:58

   Don’t Underestimate AI’s Capabilities

   1. SPSpeaker1:49

      One of the things that I think holds us all back is we give AI tools pretty simple problems when we could be giving them much more complex problems. And so that's the thing that I've struggled with where you kind of know the AI models of, you know, last year, and so you sort of give it a problem that was like the hardest type of problem that an AI tool could have done last year. But then when I look at my colleagues, they're doing all sorts of powerful things that the AI tools where they're able to do all sorts of things that I didn't even know were possible. And so a lot of this is just kind of seeing what others are doing, being able to open up your eyes to what the current AI models are able to do today, and then lots of experimentation. Ultimately, you're gonna need to feel it yourself. You're gonna need to be the one to go in and try it out, but what that means is you're constantly gonna need to be raising your ambition of the type of problems. You were probably handholding these AI tools originally, but how can you start to give it more latitude to actually make more judgment calls? But then I've also spent time in places like Rwanda and India where people are AI native, right? Some of the first technology they've been using outside of just, you know, cell phone and WhatsApp technology are AI tools. And of course, there's a lot of similarities, but one of the things that I don't think the AI native people fall trap to is that the AI native people understand what the capabilities are like today, and they treat the AI in that very powerful way. Versus someone who was using AI tools in 2022, they sort of see it still as like an assistant. Maybe they're misjudging the capabilities. Someone who may have never played with this technology before in some ways has the upper hand because they can kind of see what the technology's capable of today and where it's headed. And so what we all need to do is we need to sort of think like the AI native person, someone who just grew up using AI from day one, and say, "What are the models capable of today, and what will they be capable of tomorrow?" Elevating our ambition with what we can do with these AI tools. As the AI tools get smarter, where could I push you further? Where could you have pushed me further? And you really need to elevate your ambition with what is possible with these tools, and always be pushing yourself to the limits of trying things that are not quite possible in today's AI models or today's harnesses. Because then when the next model comes out, you try that same thing, and then you're at the cutting edge of something that no one else has figured out they could do. We as humans are very used to linear or, like, static things. And so it's like you have a colleague, and every day you come into work, they're just getting exponentially better. We as humans don't have a, a way to think about that, and so you're always treating your colleague like they had the capabilities from last month. But actually, it could have doubled in the past month, and that's like a really, really hard thing to grok. But those I find who are really, really good at, at using the AI tools are treating it like something that is gone from an assistant to a collaborator. And eventually, if we're looking ahead, there will be, in some cases, I think this inversion of control, where actually the AI model is doing some of the highest level strategic thinking and then delegating to you, the human, for the areas that require human taste, human agency. [keyboard clicking]

3. 4:58 – 7:09

   How to Be a Top 1% Learner with AI

   1. SPSpeaker4:58

      There was a study that Anthropic did recently, one of our research fellows, where they looked at it particularly with coding education and CS education. And what they did is they created two groups. Both had to do some assignment, coding assignment. One of them could use AI tools. One group couldn't. And what they found, of course, as you may expect, is the, the group that could use the, the AI tools were able to, you know, finish the assignment much quicker. But what they then did is they had a separate assessment afterwards where they couldn't use AI at all in either group to assess their understanding of the concepts that they had just used in that assignment before. And what they found is that the group that didn't use AI tools actually performed 17% better. They understood the concepts better because they had sort of slogged through this work without AI, and they really had to internalize it. And that's, I think, the warning for all of AI in education, is that the more we use these tools, the more it raises questions of this, like, skill atrophy. Now, the good news with that particular study was that not all of them did worse. And actually, the ones that were using the AI tools not in such a transactional way, but more of in an inquiry way and they were probing and asking questions, they actually were able to perform well on the final assessment. How you use the AI tools really matters. It's not just that you're using them or you're not using them, it's the way in which you're engaging. What is your goal when you show up to this AI tool? Are you just trying to get the task done as quickly as possible? And we all fall trap to that. Or are you really trying to maybe do the task, but also get better and smarter as you're doing it? People often go to the AI tool just with a particular solution in mind, and you ask a very particular question 'cause you're trying to get that solution, where really the best practice is probably to come with the problem, not the solution, and say, "This is the thing I'm wrestling with." And, you know, maybe you know some potential solutions, but actually when you just come into an AI bot with the particular solution, it's gonna give a very narrow answer around that. Versus if you come with a much more open-ended problem, the AI models of today are actually pretty good at helping you wrestle with this problem. In 2026, we can come with the hairier problem and really try to get that input from the AI tool.

4. 7:09 – 9:51

   Imagine 10x beyond the chatbot

   1. SPSpeaker7:09

      One misconception I think that's out there is that all this learning is gonna take place in the form factor of an AI chatbot. Ultimately, we need to think- 10x about what those formats and what this, this media looks like in which you're engaging with an AI. And I don't think we've quite cracked it in 2026, but let me give you a preview of what I see happening at Anthropic and beyond. Claude Code, this is, you know, Anthropic's coding agent, it was not meant for learning. And yet we've seen people use it to learn in very creative ways. Not things like coding, just general things. Learning a new language, learning some economics concept. And what they'll do is, you know, they'll go into Claude Code, and they'll start to build this memory and this context about what they're learning, about who they are, about how they learn best. And so it ultimately becomes this coach. As you start to see these coding agents and people using them in ways that are kind of unexpected, I think it's pointing at where things are headed with learning. And now we're headed into a world where I think the interfaces are gonna be much richer. I've done the stopwatch activity before, and I've found that sometimes it's actually increasing the amount of time I'm spending on a task versus if I'd just done it by hand. But I think that is expected where we are right now in this exponential in 2026, especially if you go back to this idea that we always need to be pushing the limits of what's possible. So I wouldn't recommend taking everything in your life and trying to use AI for it when it doesn't make sense, but I do think you can assume that there's some fraction of your time you should spend experimenting. And it's like R&D, research and development. You may be losing a little time here and there, but you're sort of understanding the limits of these tools. Some fraction of our time we should be pushing the limits, and what it may mean is that today you're actually wasting a little time to use these tools, but tomorrow you will be saving time. AIs have become quite powerful, as we all know, but of course they're only as powerful as the context you give them. And I would say this is probably the biggest gap when I look at people who are excellent at using AI tools versus not. If you give the AI enough context, it can actually usually do pretty well with most types of tasks. So when I go to the AI tool, I actually spend most of my time before I even ask my question throwing in all the context I can about other documents I've written, you know, the company I work at, the questions I'm thinking about. Maybe I'll even throw in a stream of consciousness of thinking I've done around this topic. And so the AI models can take in a lot of context and bring it all together, but what they're not gonna be able to do is, with very little context, just reason their way through the world. Because they won't understand how you're thinking about it. [keyboard clacking]

5. 9:51 – 10:27

   The 2030 Classroom Is Nothing Like You’d Expect

   1. SPSpeaker9:52

      But it was around a year ago that we looked into the data around how people were using the tool, and we knew that people were using Anthropic for coding and software development and in work. But what I think was surprising to us was that one of the top use cases was also educators using it, students using it. But when we looked further into the data, we were a little dismayed around how people were using it. And often it was being used as kind of a crutch, you know? Very transactionally, students were just going in and asking for, you know, answers to their homework and, and those are the things that, I mean, honestly keep me up at night. When it comes to education,

6. 10:27 – 11:15

   how do we scale up truly personalized learning?

   1. SPSpeaker10:27

      there's been a lot of research showing that one-on-one tutoring can be one of the most effective ways to learn, and so that's why I think the sort of like elusive dream in AI and, and technology and education has been, how do we scale up this truly personalized learning? But I will say there's a little more than just that, because we want learning that's personalized, but you also hear it in the word personalized. We want learning that is personal. And ultimately, that's human to human interaction. And you go to school not just to learn a particular algebra concept or some historical figure, you go to learn to also how to interact with your, you know, fellow citizens and your, you know, colleagues and your friends. And so when we think about the role of AI in education, how can we have AI further the human to human connections in the classroom, putting humans at the center?

7. 11:15 – 18:09

   Technology will be invisible in the future classroom

   1. SPSpeaker11:15

      What I would love to see if you go into a classroom in 2030 is you walk in and you don't even see the technology, and it's operating behind the scenes. You know, maybe teachers are using it to save time and build personalized lesson plans and group students, and it's just a rich learning environment and something that you may see right now if you go into, you know, a private school. You have the resources to build this type of environment, but you bring that type of education to every school. I'm in a WhatsApp group with teachers across the world, actually, in a bunch of different countries, and every morning I wake up, I'll see new WhatsApp messages from these teachers where they're sharing about how they built, you know, these custom tools with AIs like Claude. And it's not just, like, custom lesson plans. I think that's, like, the first level. But these are teachers who are building, you know, flashcard apps for their students and, you know, formative assessments, and they're able to sort of just envision something that previously, you know, you could have thought about, but it may take a year to build a new curriculum, a new lesson plan, a new way to assess your students. And they can just every morning wake up, come up with this idea, build the tool with, you know, Claude or Claude Code or Claude Artifact, and then just deploy it in their class with their students. And it just builds for a much, much richer learning environment. What I'm excited about in 2030, I'll sort of paint a picture of what I think this could look like. So first of all, these AI tools are gonna need to know a lot about your context. Let's say you're a student in school. They're not just gonna be using some generic curriculum. They're gonna know the context of your school's curriculum, your state curriculum, and they're gonna be able to tie everything back to that. So one, they know more about your context. Two, they're gonna know more about you. You know, I was a teacher before, and it would've been very disappointing to all my students if every time I walked into the classroom, I said, "Oh, yeah. Remind me, what are you working on? And what do you know?" And, and that's the type of interaction we're having right now in 2026, and I think where we have to head is a world where this is a real learning companion, and it understands where you're coming from, and you get to choose to give it this context so that you're growing with it, and the AI tool's also growing with you. So I've known Sal Khan since I was a high schooler when I first interned at Khan Academy. And Sal Khan, before he created Khan Academy, he was doing one-on-one tutoring with his cousins over Skype. And so actually, his initial vision with Khan Academy was, "How can I scale this type of tutoring to more people?" But of course, Khan Academy was sort of based on YouTube videos, and so in 2020 we started to wonder, how else could we recreate this one-on-one tutoring experience with Sal Khan but make that available to everyone? And so that's why we built this peer-to-peer tutoring platform called Schoolhouse, where anyone in the world log on and receive free peer-to-peer tutoring in subjects like math, science, test prep. And then also, if you wanna give back, you can go volunteer and be a tutor. What we notice with Schoolhouse is that if we c- create a global learning community, there's actually more shared elements than you may initially think. I've tutored, you know, tutoring sessions on Schoolhouse where I have someone from Russia, someone from Colombia, someone from the US, and someone from China all in the same Zoom call together learning the same topics, and then the students will start to share, "Oh, in my school we learn it this way. In my school we learn it that way." And there's been something kinda magical where you can actually bring people together. The world feels a lot smaller, and you have these high schoolers learning together and also forming connections that would previously not have been possible. And so to actually bring them into the same room, to have them learn together, I think is just incredible. For me, an incredibly compelling way to scale up tutoring. Sure, AI tutoring is a part of it and we're gonna have amazing AI tutors. At the same time, there's a, a personal dynamic to it, and having someone who cares about your progress, who holds you accountable, is equally important, if not more important. We essentially have birthed this new s- species of artificial intelligence into the world. And then when I think about my own education experience, I went through 15-plus years of school learning how to interact with other humans. But now we're entering on a w- a world where we have another type of being, which is artificial intelligence, which is gonna be in our workplaces as a colleague. You're gonna just see it show up in Slack one day as, as a coworker. It's gonna be showing up in schools. This is a whole type of different skill, which is, how do you not just interact with other humans productively, but how do you collaborate with an AI? And in some ways it, it is a social skill. It's not just a technical skill. Sure, early days of AI was how do you prompt it this and that way, but I think that era's over. We're, we're done with the days of just, like, technical, like, how do you prompt? But ultimately you have to treat this more as a colleague, as a collaborator, and so then it becomes more like a social skill. Like, what is the way in which I can interact with an AI so that it understands more about what I'm looking for, I understand more about how the AI works? And that's like a type of dialogue that is similar to how we're building these social skills with other humans. Now we're gonna be in a very multifaceted environment where you have multiple humans working with shared AIs, and that's why I keep going back to this idea of practice and reps. You just spend time with an AI, start treating it more like a colleague, interacting with it, understanding its limitations today, its capabilities. But that's gonna come through practice and repetitions.

   2. DBDrew Bent16:44

      We should all be having dozens or hundreds of agents working for us. So over time I built up this team of 40 AI marketing agents to work with me. I'm the only marketing person at RelatedApp. I have an AI coach that once a week tells me, "Here are the themes that seem to be really working well, or the styles that are working really well for you. Do more of this, do less of that." That's why I've had a few very successful posts on LinkedIn. So for example, the post was, "Here are the 40 marketing agents that I actually use," and that post so far has 1.5 million impressions and the most comments ever on a LinkedIn post. If your job is as a junior content marketer to take a YouTube video and write a blog post based on that YouTube video, that job won't exist anymore. A high-quality marketing contractor, let's say I had four people at $12,500 per month each, that would be $50,000 a month, and my AI bill is $500 a month. Building AI agents is the fundamental skill that will define every professional's career for the next 30 years. It's a requirement. In the same way that knowing how to use spreadsheets like Microsoft Excel was a requirement for the last 40 years, knowing how to build AI agents is a requirement for the next 40 years. We are all using it for probably 1% of what we should be using it for. Here's what I'll say. AI is way under-hyped. [upbeat music]

Episode duration: 18:09