AI tutors and mastery learning: what educators and neuroscientists agree on

The Short Version

Across three very different conversations, an education entrepreneur, a computational neuroscientist and the founders of an AI code editor land on a shared idea: AI helps learning most when it gives each person feedback at their own level, fast, and keeps them working until they have really mastered the step in front of them. It helps least when it simply hands over answers. Joe Liemandt of Alpha School argues that AI tutors finally make Bloom's "two sigma result, one-on-one tutoring to mastery, affordable for every child, and says his students cover more than twice the usual material in about two hours a day [1]. Terry Sejnowski explains why that design fits the brain: learning runs on prediction and feedback, it needs both knowledge and practice, and it has to be active learning" rather than being told the answer [2]. The Cursor team never discusses onboarding or teaching directly, but their editor is built on the same principle for working programmers: instant, personal suggestions that the human judges and corrects, with the human kept in the driver's seat [3]. Put together, these sources move AI in education away from the cheating worry and toward a practical tool for faster, deeper learning, provided it is built to demand mastery rather than to do the work for you.

Table of contents

What They Said

Joe Liemandt: AI tutoring plus mastery, two hours a day

Joe Liemandt, a technology entrepreneur and principal of Alpha School, starts from history. About a century ago, societies decided to educate everyone and copied an industrial model from Prussia: a room of children with a teacher at the front [1]. That model is now nearly universal, from expensive private schools to poor districts in India, but outcomes vary enormously. His one-line diagnosis: "We have universal access, but not universal learning." [1] He argues that AI makes it possible, for the first time in 50 to 100 years, to redesign education from the ground up [1].

Bloom's two sigma, and the half everyone forgets. Liemandt's central evidence is a 1980s paper by Benjamin Bloom. In his words: "There's a famous paper from the 80s, Bloom's 2 Sigma program where he's, he's like, if you give an individual tutor to mastery level, kids can learn at 2 Sigma." [1] He insists the mastery part is what people drop: "Mastery is actually the part everybody forgets about on that. If you have a tutor that's not mastery based, you won't get to top 2%. One Sigma came from the tutor. One sigma comes from mastery." [1] In other words, a personal tutor alone gets you half the benefit; refusing to move on until the student has mastered the step gets you the other half. "You need individualized instruction at a mastery level to get that performance." [1]

One-on-one tutoring was always the best model, he says, but it wasn't affordable, so societies settled on 30-to-1 classrooms [1]. AI changes the cost: "And now with AI tutors, at cost effectively, you can give every kid perfectly tailored lessons at their level for them at a mastery level." [1]

Not chatbots. Liemandt is emphatic that this does not mean chatbots, which he says most students use to cheat [1]. What Alpha's software does instead is assess, personalize and loop: "I can give you an assessment. I can know exactly what you know, what you don't know. I can give you a lesson personalized to you that says you need this lesson, and then I give it to you. And then I'm going to hit you to mastery. I'm going to keep you in that loop to mastery." [1] Only after mastery does the student move to the next topic [1]. This distinction, between a tool that answers for you and a tool that diagnoses and drills you, is the hinge of this whole insight.

The results he reports. Alpha students averaged 121 minutes a day in the apps last year and, by Liemandt's account, learned more than twice what peers learned in a full school day plus homework, which he frames as roughly five to ten times normal speed [1]. The summary notes that these results are Alpha's own reports, with randomized controlled trials with MIT's Blueprint Labs still underway [1].

The sports ethos. His favorite analogy is coaching. "We don't do the basics. We need to nail the basics, just like in sports." [1] If a player loses the ball 20% of the time, a coach works on ball handling, not dunks: "20% isn't acceptable. 15% isn't acceptable. You can't be 85, 80% accurate." [1] School does the opposite: "You just finished fourth grade at 80%. Let's get you into that fifth grade stuff. And the answer is no, no, because then all those problems compound." [1] He sums up the double standard as "we all accept mastery in athletics, but we reject it, explicitly reject it in education." [1]

Working memory and fluency. Liemandt's most practical argument is about working memory, which has a limited number of slots [1]. Anything memorized to fluency does not use a slot, so a student who has to calculate seven times eight is spending capacity she needs for a multi-step algebra problem [1]. His example: a student stuck in the low 700s on SAT math, making "careless errors, went back and memorized her multiplication tables and then scored 790 [1]. He adds that gifted children who do everything in their heads eventually hit a wall, because they never built good habits for multi-step work, and that schools bore gifted children and teach them bad habits for 12 years [1]. He describes knowledge as layered: chemistry depends on algebra, algebra on fractions, fractions on math facts [1]. When he talks about what predicts performance, he says IQ, you get more working memory slots, with Big Five conscientiousness as the other factor, and then adds that our educational system isn't a mastery-based system." [1]

Gaps are closable, and fast. Students transferring into Alpha from expensive private schools averaged "2.2 grade levels behind" on Alpha's assessment [1]. Because a grade level in one subject takes 20 to 30 hours for K–8 students, he argues a gap of years can be closed in months [1]. He cites NWEA MAP data showing the average student barely improves between grades 8 and 12, roughly 233 to 234 on a 300-point scale [1].

The data loop as a scientific instrument. Liemandt compares AI to the microscope or telescope: because each student works one-on-one with software, Alpha can test learning-science ideas cleanly [1]. His example is a live experiment on initial mastery. Published papers suggested "you don't need initial mastery. You can do it with space repetition," and in grades 8–12 "we crushed it with just worked examples and spaced repetition. Because once you get into algebra, the multi-line problems, worked examples really matter." [1] But in grades 4–7, where the problems are easier, scores dipped, so "we then just added initial mastery back to it. So you get the worked example, you get the initial mastery and then spaced repetition." [1] Pressed by Huberman, his conclusion is blunt: "Initial mastery always matters. Just it always does." [1]

Teachers who were never taught the science. He says schools of education rarely teach cognitive load theory, direct instruction or Project Follow Through [1]. At a staff day with hundreds of teachers, he asked who had learned direct instruction: "You know, 5% raise their hand." [1] Project Follow Through, he says, was "the largest educational study ever about effective ways to teach kids," and none of his teachers had read it [1].

Motivation, shame and growth. Liemandt calls motivation "90% of learning [1]. Huberman points out that individualized learning removes shame, since nobody knows an eighth grader is working on third-grade material [1]. Liemandt wants schools judged on growth rate, the slope of the line, rather than raw achievement, which mostly reflects selection [1]. The adults' job is to find and remove each child's blocks, such as I'm not a math person [1]. That is where his controversial practice of paying kids comes in: his daughter, who didn't see herself as a math-and-science kid, reached the top 1% after he bet her $1,000, and later said it changed how she saw herself [1]. His 100 for 100" program paid seventh graders $100 for each perfect score on state tests, starting at third grade, which pushed them to fill old gaps [1]. He believes paying for a breakthrough can create intrinsic motivation afterward, though some Alpha parents object [1].

High standards, high support. Drawing on Dr. Yeager's book 10 to 25, Alpha trains its guides in a "high standards, high support" framework [1]. Liemandt's summary of child development: "The key to developing a child is that they struggle and fail on their road to success supported by a caring adult." [1] He reports that 90-some percent of Alpha students say they love school and 40 to 60 percent would choose school over a vacation, and he argues that happiness comes from the high standards, not despite them [1]. He also says ADHD and gender gaps largely disappear in the model, with 25-minute Pomodoro blocks for students with ADHD [1].

AI frees the humans. Because the software handles academics in about two hours, afternoons go to building: a student-run music festival that raised $25,000 in sponsorships, vending machines and food trucks, a Broadway-style musical, and a tracked 1,000-hour passion project [1]. AI, he says, frees human guides to do what teachers originally signed up for: developing kids [1]. He plans to scale through vouchers, public-school pilots at roughly $20,000 per student, and a free learning game, noting that "20 grand a year, you can make a 10-year-old, like, love school." [1] He also cautions that changing the whole system will take about 20 years [1].

Terry Sejnowski: the neuroscience of feedback, practice and active learning

Terry Sejnowski, a professor at the Salk Institute who directs its computational neurobiology laboratory and helped pioneer neural-network AI, supplies the brain-level explanation [2].

Learning is prediction plus feedback. His flagship example is work from the 1990s with Peter Dayan and Read Montague on the basal ganglia, which learn sequences of actions toward a goal and also interact with the prefrontal cortex to shape thinking [2]. The algorithm, he says, is remarkably simple: "It's simply to predict the next reward you're going to get." [2] Each time you act, you compare what you got with what you expected and update your synapses, gradually building a value function, your accumulated sense of what is good or bad for you [2]. This temporal difference learning is the same algorithm DeepMind used in AlphaGo [2]. The value function, he adds, is "a very permanent part of your experience and who you are." [2]

He describes two routes through trial and error: "Small rewards are good because you're constantly coming closer and closer to getting what you're seeking, better tennis player or being able to make a friend. But the negative punishment is much more effective, one trial learning." [2] A shocked rat never forgets; one terrible relationship teaches lessons for life; PTSD shows how one experience can shape a lifetime [2].

Two learning systems that need each other. Sejnowski distinguishes cognitive learning, which is cortical, step by step and slower, from procedural learning, which is subcortical, runs through the basal ganglia and becomes automatic [2]. "We have two major learning systems. We have a cognitive learning system, which is cortical. We have a procedural learning system, which is sub cortical, basal ganglion. And the two go hand in hand. If you want to become good at anything, the two are going to help each other." [2] As a physics PhD, he found that working through problem sets, not memorizing equations, made him a physicist [2]; in the conversation, "doing your theoretical physics problem sets reinforces the examples that you learned in lecture and in your textbooks." [2] Huberman's version: imagine "a textbook on swimming. And then you're going to go out to the ocean someday and you will have never actually swam." [2]

Against removing practice. Sejnowski is sharply critical of a push in California schools to cut practice so students don't feel stress: "they're trying to get rid of the procedural. That's ridiculous. They don't want students to practice because it's going to be, you know, you're stressing them." [2]

Active learning, not being told. The key, he says, is that "it has to be active learning": wrestling with a problem yourself engages trial and error, whereas simply being told the answer stores a fact that won't come up when a similar but different problem appears [2]. That is also why large language models work: they are not parrots that memorize, they generalize to new cases, and "The brain is really, really good at generalizing", often from a single example [2]. This is the neuroscience behind Liemandt's warning about chatbots used to cheat: an answer handed over skips exactly the trial-and-error process that builds usable knowledge [1][2].

Learning how to learn, and self-testing. With engineer-turned-educator Barbara Oakley, whom he calls "the mastermind," Sejnowski created Learning How to Learn, a free online course taken by about 4 million people in 200 countries, ages 10 to 90, with roughly 98 percent approval [2]. It is 50 or 60 bite-sized videos of about 10 minutes each over a month, with quizzes, tests, forums and teaching assistants [2]. It deliberately avoids subject matter: "We're not trying to give you knowledge. We're trying to tell you how to acquire knowledge." [2] It covers exam anxiety and procrastination [2]. Surprisingly, "the peak of the demographic is 25 to 35": working adults with mortgages or children who need new skills but can't go back to school [2]. Huberman highlights self-testing, saying "we need to test ourselves on the material," not only to be evaluated but to find and correct errors [2].

Sleep, energy and reserve. Sejnowski's group found circular traveling waves across the cortex during non-REM sleep, the sleep spindles that help consolidate the day's learning without overwriting what you already know [2]. Sara Mednick's experiment showed the sleep drug zolpidem (Ambien) doubles spindles and roughly doubles next-morning recall, but it also wipes out memories of what happens after taking it; both men conclude you never get anything for free [2]. On energy, he says "That exercise is the best drug you could ever take," and describes his own short all-out running intervals [2]. He also recounts a Chinese study in which Alzheimer's began earliest in people with no education and latest in those with the most, suggesting that "the more you exercise your brain with education, the more reserve you have later in life." [2]

AI as a partner in thinking. Sejnowski is enthusiastic, not naive, about AI. He relays a story of a writer who was exhausted by ChatGPT until she treated it politely: "You treat it like a machine. It will treat you like a machine." [2] Colleague Rusty Gage feeds his lab's experiments and the literature into a language model and says "we use it as an idea pump" [2]. In a dermatology study, doctors and AI each classified skin lesions with about 90 percent accuracy, but doctors using the AI reached 98 percent, because the AI knew rare lesions and doctors knew common ones in depth; "It's not going to be adversarial. It's going to be a partnership." [2] Most relevant to tutoring, he once asked a model to explain a technical abstract to a ten-year-old, and it did better than he could [2]. That is the core of personalization: pitching an explanation at the learner's level. He also notes models improve over a single conversation through in-context learning, which remains a mystery [2].

The Cursor team: instant, personal feedback for programmers

Michael Truell, Sualeh Asif, Arvid Lunnemark and Aman Sanger, the founders of the AI code editor Cursor, speak about professional programming, not schooling [3]. It is important to be plain: the source contains no explicit discussion of onboarding new programmers or of mastery learning. What it offers is a detailed picture of an AI tool designed around fast, individual feedback and human judgment, which is the same design principle Liemandt and Sejnowski describe [3].

A colleague looking over your shoulder. Michael describes what Cursor does well as acting like a fast colleague looking over your shoulder who predicts what you'll do next, and letting you jump ahead by telling the AI what to do [3]. When the founders first used GitHub Copilot, it felt like a friend finishing your sentences, and even when it was wrong you just typed another character [3]. Cursor Tab extends this to predicting your next edit and the next place you will jump, and Arvid frames pressing Tab as injecting intent: programming becomes higher- and higher-bandwidth communication with the computer [3].

Instant results to react to. The team thinks agents are really cool but not yet super useful for most tasks, because most programming is iterative, so you mostly want instant results to react to [3]. That preference for short cycles of try, see, correct is the same loop Sejnowski describes as prediction and update, applied to a working professional [2][3].

Learning a codebase. The closest the source comes to onboarding is its treatment of context. Cursor suggests relevant files to add based on your history, and a model could ask a clarifying question or show several options when intent is unclear [3]. The team is interested in teaching models a specific codebase directly, through continued pre-training and synthetic question-answer data about a repository [3]. Arvid's habit of labeling dangerous code with comments on every line helps both humans, who forget, and models, which pay more attention [3]. And models lack the calibration a staff engineer has from seeing sketchy code take down a server, the kind of cultural knowledge that only comes from experience [3].

Review as guided understanding. Aman says "Code review kind of sucks," and suggests models could highlight the important parts of a diff, gray out the boring ones, flag likely bugs and walk reviewers through changes in logical order [3]. That is essentially a personalized explanation of unfamiliar code, delivered at the moment it's needed.

Humans stay in control. Michael argues good engineering is made of countless micro-decisions and trade-offs, and handing them off loses control; the team wants a future that emphasizes the programmer's speed, agency and control rather than chatting with a bot like an engineering department over Slack [3]. Cursor's manifesto describes a hybrid engineer who will "iterate at the speed of their judgment," [3] and Michael tells young people "I actually think this is a really really exciting time to be building software," because "people's taste and creative ideas will be magnified." [3] Arvid describes a migration that took him and Sualeh five days and could someday take ten minutes by showing a few examples, letting programmers try things first instead of planning everything at the blackboard [3]. They also note it is sometimes easier to show an example than to explain [3], which echoes Liemandt's finding that worked examples matter for multi-step problems [1].

Where They Agree, and Where They Don't

Where they agree

Personal feedback at the learner's level is the gift. Liemandt's AI tutor assesses exactly what a child knows and serves the next lesson at that level [1]. Sejnowski's learning rule is built on comparing an expected result with the actual one and updating, and his model explained an abstract to a ten-year-old better than he could [2]. Cursor predicts each programmer's next edit from their own context and history [3]. All three treat AI as a way to give one person feedback tailored to them, quickly and cheaply.

Doing beats being told. Liemandt rejects chatbots as cheating tools and insists on a loop to mastery [1]. Sejnowski says "it has to be active learning" and that being told the answer stores a fact that won't transfer [2]. Cursor keeps the human making micro-decisions and judging every suggestion rather than handing work off [3]. The shared warning is that an AI that simply produces answers can undercut learning or control; an AI that makes you act, check and correct speeds it up.

Practice and fundamentals are not optional. Liemandt's ball-handling analogy and multiplication-table story [1] line up closely with Sejnowski's two learning systems and his criticism of schools that remove practice [2]. Both men lean on sports: Liemandt's coaching model [1], Sejnowski's tennis serve and Huberman's swimming textbook [2].

AI amplifies humans rather than replacing them. Liemandt says AI frees guides to develop kids [1]; Sejnowski's dermatology example shows doctors plus AI beating either alone [2]; the Cursor team wants programmers' taste "magnified" and themselves in the driver's seat [3].

Education is worth it for everyone, at every age. Liemandt wants to reach a billion kids [1]; Sejnowski's biggest audience is adults aged 25 to 35, and he links education to later cognitive reserve [2]; Cursor sees faster iteration as freeing programmers to try more ideas [3].

Where they don't, or where the sources don't reach

Struggle and discomfort. Liemandt says that if something can be taught without suffering, it should be, and that people struggle willingly toward goals they care about [1]. Sejnowski is more willing to defend the discomfort of practice, calling efforts to remove it for students' stress "ridiculous," and notes punishment teaches in one trial [2]. The two are not fully opposed: Liemandt also insists kids "struggle and fail on their road to success supported by a caring adult," and holds 80% as unacceptable [1]. The difference is emphasis. Liemandt wants to strip out needless suffering and keep purposeful struggle; Sejnowski stresses that some stress is part of how practice works.

Rewards versus punishment. Liemandt uses positive incentives, even cash, to break through mental blocks [1]. Sejnowski notes small rewards move you steadily closer to a goal, but "the negative punishment is much more effective, one trial learning." [2] Neither suggests punishing students; Sejnowski is describing how the brain learns, not recommending a classroom policy. Still, the sources give different pictures of what drives the fastest learning.

Do we still need the basics if machines can do them? Huberman plays devil's advocate with Sejnowski: calculators and search mean "there is a world where certain skills are no longer required," freeing brain space for other knowledge [2]. The sources don't settle this in one line, but Liemandt's working-memory argument answers it from his side: facts memorized to fluency don't use working-memory slots, so skipping them makes harder work harder [1].

Initial mastery versus spaced repetition. Even within Alpha, the learning-science literature suggested initial mastery could be skipped in favor of spaced repetition; Alpha's data backed that for grades 8–12 but not for grades 4–7 [1]. Liemandt concludes initial mastery "always matters," a conclusion drawn from Alpha's own data rather than from a published trial [1].

How much the machine should do. Cursor's Arvid predicts "I think people will just not write tests anymore," with models proving code meets a spec, but his colleagues push back that intent is hard to specify formally [3]. The team also says agents aren't yet very useful for most tasks [3]. This is a live debate about how much to delegate, not a settled view.

The evidence is uneven. Alpha's figures (twice the learning, five to ten times the speed) are its own reports, with independent trials underway, and Huberman raises selection effects [1]. The Cursor team warns that benchmarks mislead and leans on qualitative vibe checks [3], a reminder that impressive numbers need careful measurement. Sejnowski is candid that "we don't understand what understanding is." [2] And the Cursor source says nothing directly about onboarding or teaching, so its contribution to this insight is by analogy, not direct testimony.

Why It Works

Mastery removes compounding gaps. Knowledge is layered, chemistry on algebra, algebra on fractions, fractions on math facts, so moving on at 80% leaves holes that grow [1]. Bloom's result, as Liemandt reads it, splits the two-sigma gain equally between tutoring and mastery [1]. An AI that assesses and loops to mastery addresses both halves at once [1].

Fluency frees working memory. Working memory has limited slots; facts known to fluency don't take one, leaving room for multi-step reasoning [1]. The SAT student who jumped from the low 700s to 790 after learning her multiplication tables is his illustration [1]. Sejnowski's account of procedural learning becoming automatic through the basal ganglia is the brain-level counterpart [2].

Feedback drives synaptic updates. The brain predicts a result, compares it with what happens and updates, building a lasting value function [2]. Short feedback cycles, whether an AI lesson that checks each answer [1] or an editor suggestion accepted or rejected in a keystroke [3], give that system many clean comparisons.

Active effort produces transfer. Being told an answer stores a fact that won't surface for a similar but different problem; working it out engages trial and error and builds generalization [2]. This is why the same AI can either help or hurt: used to generate answers it bypasses the mechanism; used to drill and diagnose it feeds it [1][2].

Cognitive and procedural learning reinforce each other. Lectures and problem sets, knowledge and practice, help each other [2]. Alpha's combination of worked examples, initial mastery and spaced repetition is a practical recipe that joins explanation with repeated practice [1].

Removing shame and blocks raises motivation. Individual pacing means nobody sees who is on which level, and measuring growth rather than rank keeps attention on progress [1]. Aligning work with what learners care about, like finishing academics to reach sports practice, makes struggle voluntary [1].

Sleep and energy lock learning in. Sleep spindles consolidate the day's learning into the cortex without overwriting it [2], and exercise supports the brain's energy supply [2]. Freed time, which Alpha uses for afternoons of building [1], matters only if learners also sleep and move.

Partnership beats either alone. The dermatology case shows AI and humans covering each other's blind spots, reaching 98 percent together versus about 90 percent separately [2]. Cursor's design, where the model drafts and the programmer judges, relies on the same division of labor [3].

Where It Applies, and Where It Doesn't

Applies strongly to layered, checkable skills. Math, reading vocabulary and test-based subjects, where an assessment can say exactly what you know, are Alpha's core [1]. Liemandt notes that reading comprehension "is a function of vocabulary" and that the best way to do well on SAT English is to "just read lots of books. And look up the words you don't understand." [1] Programming, where a suggestion can be accepted or rejected immediately, is another natural fit [3].

Applies to adults, not just children. Sejnowski's course draws most heavily on working adults aged 25 to 35 [2], and the Cursor team's whole product serves working professionals [3]. The principles of feedback and active practice are not age-bound, though Sejnowski notes adults may learn some things less quickly [2].

Varies by age and problem type. Worked examples plus spaced repetition sufficed for grades 8–12 but not for grades 4–7, where initial mastery had to come back [1]. The prefrontal-basal ganglia loop is among the last to mature in early adulthood, which is why adolescents often struggle to stop themselves [2], and younger children are driven more by pleasing parents while teenagers are driven by social belonging [1]. One design does not fit every age.

Does not apply to answer machines. Chatbots used to get answers are, in Liemandt's view, mostly a cheating tool [1], and Sejnowski's account of active learning explains why passive answers don't transfer [2]. The benefit belongs to AI that assesses, drills and demands mastery.

Not a replacement for the human side. Life skills, motivation, grit and the caring adult who supports struggle stay human at Alpha [1]. Cursor's founders keep humans in control of micro-decisions and trade-offs [3]. Many things, like tennis or clearing a flooded scuba mask, can only be learned by doing [2].

Hard to specify tasks remain hard. Where intent is fuzzy, AI struggles; Cursor's team says formal specs are hard for real-world intent and that frontier models are bad at finding bugs when simply asked [3]. The analogous limit in education is anything that can't be cleanly assessed.

Evidence and access caveats. Alpha's results are self-reported and its students are partly self-selected, with trials underway, and Liemandt himself says system-wide change will take about 20 years [1]. Its full tuition model costs far more than public schools, though he plans voucher and public-school versions at roughly $20,000 per student [1].

No shortcuts through drugs. Sejnowski's Ambien example shows a doubling of recall with a hidden cost [2]; anyone considering sleep medication or stimulants should get a doctor's advice rather than using them as learning tools.

What To Do With It

1. Use AI to diagnose, not to answer. Ask an AI tool to assess what you or your child actually know and to generate practice at that level, then keep working until the step is mastered, the loop Liemandt describes [1]. Avoid using it to produce finished answers, which skips the active learning Sejnowski says builds transfer [2].

2. Set a mastery bar before moving on. Treat 80% like a coach treats losing the ball 20% of the time: go back to fundamentals before advancing [1]. Fill basic gaps, such as multiplication tables, before pushing into harder material [1].

3. Build fluency in the basics to free working memory. Memorize the facts that sit under multi-step work so they don't take up working-memory slots, as in the SAT student's jump to 790 [1].

4. Pair explanation with practice. For multi-step problems, study worked examples, check for initial mastery, then use spaced repetition [1]. Then do the problem sets yourself; knowledge and practice help each other [2].

5. Test yourself. Use quizzes and self-tests to find and correct errors, not only to grade yourself [2]. Adults who need new skills can take the free Learning How to Learn course with Barbara Oakley and Sejnowski [2].

6. Measure growth, not rank. Track how fast a learner is improving rather than where they stand against others, and look for the blocks, like "I'm not a math person," that stall them [1].

7. Ask for explanations at your level. Sejnowski's model explained a technical abstract for a ten-year-old better than he could [2]; request explanations pitched to your current understanding, and treat the AI as a partner rather than a machine [2].

8. Keep humans making the decisions. In professional work, use AI for fast suggestions you accept or reject, stay in control of the micro-decisions, and use it to walk you through unfamiliar changes in logical order [3].

9. Protect the biology of learning. Think about a hard problem before sleep so the brain works on it overnight, let your mind wander on walks, and exercise, including short intense intervals [2].

10. Use saved time for building. If AI tutoring compresses academic work, spend the recovered hours on projects that demand real struggle with support, as Alpha does with its afternoons [1].

Sources

1. Joe Liemandt on Alpha School, AI Tutoring, and Learning Faster Through Mastery and Motivation · Huberman Lab
2. Dr Terry Sejnowski How To Improve At Learning Using Neuroscience Ai · Huberman Lab
3. Cursor Team: Future of Programming with AI · Lex Fridman Podcast