CHAPTERS
- 0:00 – 0:30
Thinking like an AI-native: from “assistant” to powerful collaborator
Drew contrasts early AI usage—treating tools as simple assistants—with an AI-native mindset that assumes much stronger capabilities. He argues the biggest unlock is raising ambition and engaging AI as a collaborator, not a step-by-step vending machine.
- •AI-native users start with today’s capabilities, not outdated mental models from 2022
- •Shift from “playbooks” and rigid prompting to giving context and letting AI reason
- •Using AI well resembles a social skill: collaborating, not commanding
- •Adopt the mindset of someone who grew up with AI from day one
- 0:30 – 1:30
Drew Bent’s tutoring mission: scaling world-class education with AI
Drew introduces his background in tutoring, teaching, and building education programs, culminating in his role leading education at Anthropic. His north star is scaling the benefits of high-quality tutoring to everyone, which AI may finally enable.
- •Career arc: peer tutoring → nonprofit → high school teacher → AI tutors at Anthropic
- •One-on-one tutoring is effective but historically hard to scale
- •AI makes elite learning support more accessible globally
- •Central question: what does great tutoring look like at scale?
- 1:30 – 3:31
Stop giving AI “small problems”: raise your ambition through experimentation
He explains how many users under-challenge modern models because they’re anchored to what last year’s tools could do. Progress comes from experimentation, learning from others, and gradually giving AI more latitude to make judgment calls.
- •People often cap tasks at what prior models could handle
- •Watch what power users do; then try it yourself
- •Continuously increase task complexity and delegate more judgment
- •Hands-on experimentation is the only way to internalize capabilities
- 3:31 – 4:32
Exponential improvement and the risk of outdated expectations
Drew notes humans struggle to intuit exponential change, so we keep treating AI like it hasn’t improved. Top users behave as if the tool is rapidly evolving, and they test ideas that may be barely possible today so they’re ready for the next model.
- •Humans default to linear thinking; AI capabilities jump quickly
- •Treating AI like last month’s version limits outcomes
- •Try “near-impossible” workflows to stay at the cutting edge
- •Future possibility: inversion of control—AI does strategy, humans provide taste/agency
- 4:32 – 6:34
The learning tradeoff: speed vs understanding (and skill atrophy)
He cites an Anthropic study in coding education showing AI can speed completion but may reduce conceptual understanding when used transactionally. However, learners who use AI inquisitively—probing and questioning—can retain or improve understanding.
- •AI-enabled group finished faster, but non-AI group scored 17% higher on later assessment
- •Key risk: skill atrophy when AI is used as a shortcut
- •Inquiry-based use (asking why/how) supports learning better than answer-seeking
- •The goal should include getting smarter, not just finishing faster
- 6:34 – 7:04
Ask with problems, not solutions: open-ended framing for better thinking
Drew recommends approaching AI with the real problem you’re wrestling with rather than steering it toward a predetermined answer. Open-ended problem framing helps the model explore broader solution spaces and become a genuine thought partner.
- •Narrow questions produce narrow answers
- •Bring the messy problem and constraints, not only your preferred solution
- •Modern models are better at “wrestling” with complex, ambiguous issues
- •Use AI to expand options, not just confirm a direction
- 7:04 – 8:05
Beyond the chatbot: richer learning interfaces and unexpected tools like coding agents
He challenges the assumption that AI learning will primarily happen via chatbots. Drew describes how tools like Claude Code—built for coding—are already being repurposed as learning coaches by accumulating memory and personalized context.
- •Chatbot form factor is not the end state for AI learning
- •Claude Code is used creatively to learn non-coding topics too
- •Persistent context and “memory” can turn tools into coaches
- •Next wave: richer, more integrated interfaces for learning
- 8:05 – 9:05
Experimentation as R&D: when AI may slow you down today to speed you up tomorrow
Drew frames AI adoption as personal R&D: some workflows may be slower initially, but they build intuition about limits and future leverage. He advises selectively experimenting rather than forcing AI into every task.
- •Early AI usage can increase time-on-task in certain cases
- •Allocate a fraction of time to experimentation to learn the tool’s boundaries
- •Short-term inefficiency can yield long-term productivity gains
- •Don’t use AI everywhere—use it where it meaningfully helps or teaches you
- 9:05 – 9:35
The real differentiator: stuffing the model with context
He argues AI is only as powerful as the context you provide—and that top users invest heavily in pre-loading documents, goals, and even stream-of-consciousness thinking. With sparse context, models can’t reliably align to your intent or constraints.
- •High performers spend more time providing context than writing the question
- •Useful inputs: prior docs, organizational constraints, goals, partial thinking
- •Large context windows enable synthesis across many materials
- •Low-context prompts force the model to guess your perspective and priorities
- 9:35 – 10:05
Education reality check: widespread transactional use and “crutch” behavior
Anthropic observed educators and students as a top use case, but also saw discouraging patterns—students using AI primarily to get homework answers. This motivates the push toward designs that foster reasoning rather than replacement.
- •Education is a major real-world AI use case
- •Common misuse: transactional “give me the answer” behavior
- •This pattern threatens learning integrity and deep understanding
- •Better tools and norms should encourage inquiry and growth
- 10:05 – 11:06
Scaling personalized learning while keeping it personal and human
Drew distinguishes between personalized learning (tailored instruction) and personal learning (human connection). He argues the best classroom future uses AI to strengthen human-to-human interactions, not substitute for them.
- •One-on-one tutoring is powerful; scaling it is the “elusive dream”
- •Personalized ≠ personal: relationships and social learning still matter
- •AI’s role should reinforce human connections in classrooms
- •Design principle: humans at the center, AI in service of community
- 11:06 – 12:36
The 2030 classroom: invisible tech and teacher-built micro-tools
He paints a near-future classroom where AI runs behind the scenes to save teachers time, create lesson plans, and group students—without dominating the room. Teachers already share examples of rapidly building flashcards, assessments, and apps with tools like Claude.
- •Goal: walk into a classroom and not “see” the technology
- •AI supports planning, grouping, and personalized materials behind the scenes
- •Teachers are already building custom tools (flashcards, formative assessments) quickly
- •AI compresses curriculum/tool creation from months to days—or hours
- 12:36 – 13:08
Context-rich learning companions: curriculum alignment + persistent student understanding
Drew describes AI systems that know the local curriculum and also know the student—so interactions don’t reset each session. With user-approved context sharing, the AI becomes a companion that grows with the learner over time.
- •AI should align to school/state curricula, not generic content
- •Current tools feel like amnesia (“remind me what you’re working on”)
- •Future: persistent, opt-in context about the learner’s progress and needs
- •Companion model: student and AI improve together over time
- 13:08 – 15:09
Peer tutoring at global scale: Schoolhouse and the power of learning communities
He recounts building Schoolhouse (inspired by Sal Khan’s tutoring origins) to scale free peer tutoring globally. Mixed-country sessions create both academic support and meaningful cross-cultural connection—something AI alone can’t replicate.
- •Khan Academy began from one-on-one Skype tutoring; scaling tutoring remained the goal
- •Schoolhouse enables free peer tutoring and volunteering worldwide
- •Global groups reveal shared concepts and different instructional approaches
- •Community and accountability are core benefits alongside academic help
- 15:09 – 18:09
Collaboration as a new social skill—and the rise of many specialized agents
Drew argues society must learn to collaborate with AI as a new “being” in workplaces and schools. He and the host discuss practice, reps, and a future where professionals manage dozens of specialized agents—reshaping roles, costs, and career fundamentals.
- •Collaborating with AI is becoming a social skill, not just prompting technique
- •Future work: humans + shared AIs in multifaceted teams
- •Example: managing a fleet of specialized marketing agents and an AI coach
- •Jobs shift as agent workflows replace routine junior tasks; agent-building becomes a core career skill
