AI as a learning partner, not a replacement for thinking
Joe Liemandt uses AI tutors for mastery learning, Terry Sejnowski calls AI an idea pump and diagnostic partner, and Ted Gibson explains what language models do and don't understand. The best use of AI in learning is to speed practice and feedback while keeping the human work of understanding.
The Short Version
Three very different experts point to the same way of using AI for learning. Let it speed up practice, feedback and idea generation, and keep the actual understanding with the person. Joe Liemandt of Alpha School uses AI tutors, which he stresses are not chatbots, to find each student's gaps and drill them to mastery. That frees human adults for motivation and coaching [1]. Salk Institute neuroscientist Terry Sejnowski calls language models an "idea pump" and a diagnostic partner, and he warns that learning only sticks when it is active, because being handed the answer leaves a fact you can't apply to new problems [2]. MIT psycholinguist Ted Gibson explains why the human has to stay in charge: language models are superb at the form of language, but "meaning is tough still," so they can follow the familiar shape of a problem into a wrong answer [3]. Put together, AI works best when it makes practice faster and feedback more precise, not when it does the thinking for you.
Table of contents
What They Said
Joe Liemandt: AI tutors for mastery, humans for motivation
Liemandt is a technology entrepreneur and the principal of Alpha School. There, students spend about two hours a day on AI-driven, individualized academics and the rest of the day on life skills [1]. His argument starts with history. A century ago, societies adopted an industrial classroom model from Prussia, with a teacher at the front and kids in rows, and it gave the world "universal access, but not universal learning" [1]. One-on-one tutoring was always the best model, he says, but it was never affordable, so schools settled on 30-to-1 classrooms [1].
Liemandt draws on Bloom's 2 Sigma finding from the 1980s: tutoring each student to mastery can bring them to roughly top-2% performance. He adds that mastery is half of that effect, the half people forget [1]. For him, AI's job is to make mastery-based individual tutoring cheap. He is firm that this does not mean chatbots, which he says most students use to cheat. Alpha's AI assesses what a child knows and generates lessons at that level [1]. In his words: "The average kid spent 121 minutes last year in the apps and learned more than 2x what they would have in sitting in class for six and doing homework. So when you just look at those standards, it's, you know, depending on the class and stuff, between 5 and 10x, normal speed" [1]. The summary notes these are Alpha's own reported results, and independent trials with MIT's Blueprint Labs are still underway [1].
He also values AI as a research instrument, comparing it to a microscope or telescope. In a normal classroom, he says, the data are "so dirty": "Your effect size is driven by random classroom elements, not the learning science" [1]. With students working one-on-one in an app, "I can measure their engagement, I can see if that lesson was good because I know how they did on the questions right afterward. I can then test them later because I'm getting spaced repetition. I'm getting recall later. I can say whether that really was transferred to long-term memory" [1]. That feedback loop turned up concrete findings. "In eighth through 12th, we crushed it with just worked examples and spaced repetition. Because once you get into algebra, the multi-line problems, worked examples really matter. But in fourth through seventh, it wasn't as effective." When scores dipped, "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]. His conclusion: "Initial mastery always matters. Just it always does" [1].
Liemandt's approach doesn't skip the learner's own work; it demands more of it. His sports analogy: if a player loses the ball 20% of the time, the coach works on ball handling, not dunks. In school, by contrast, an 80% grade in fourth grade means moving on to fifth-grade material, and the gaps compound [1]. He points to working memory as the reason. Anything memorized to fluency doesn't use up a slot, so a student who has to calculate seven times eight loses capacity she needs for a multi-step algebra problem [1]. His example is a student stuck in the low 700s on SAT math who kept making "careless" errors. "She was just running out of working memory slots. And so she went back, memorized her multiplication tables," and scored a 790 [1]. The AI finds and drills the gaps, but the student still has to do the reps.
He is also clear that AI doesn't replace the human side. Alpha trains its guides on David Yeager's "high standards, high support" framework, and Liemandt sums up child development as: "The key to developing a child is that they struggle and fail on their road to success supported by a caring adult" [1]. He calls motivation "90%" of learning, and he says the adults' job is to find and remove each child's blocks, such as "I'm not a math person" [1]. AI, he argues, frees the guides to do what teachers originally signed up for: develop kids [1]. He also sees AI helping where plain reading falls short. Reading widely teaches the boundaries of words ("what's the difference between happy and ecstatic?"), and "in our apps now, we can just be like, oh, we can give you vocabulary words and show you the limits" [1].
Huberman challenged Liemandt on the role of struggle. From neuroplasticity, Huberman argued, "the error signal is what actually drives the opportunity for change. And so without errors, there's no struggle, and therefore, without struggle, there's no learning" [1]. Liemandt's view is that if something can be taught without suffering it should be, and that people willingly struggle when they pursue goals they care about [1]. Huberman also noted that individualized learning removes shame: nobody knows an eighth grader is working on third-grade material [1].
Terry Sejnowski: active learning, and AI as a partner
Sejnowski directs the computational neurobiology laboratory at the Salk Institute and helped pioneer both neural-network AI and the math of how the brain learns [2]. His account of learning is where the "keep the human work" point comes from.
He separates two systems. Cognitive learning is cortical, step by step and slower. Procedural learning runs through the basal ganglia and becomes automatic [2]. You need both. "If you want to play tennis... you have to practice, practice, practice," he said, and the same is true of thinking: "If you want to become good in any area, if you want to become a good financier, if you want to become a good doctor or a neuroscientist, right? You have to be practicing, practicing, practicing" [2]. As a physics PhD student, he found that working through problem sets, not memorizing equations, is what made him a physicist [2]. He criticizes a push in California schools to cut back on practice so students don't feel stress [2].
His key rule is that "it has to be active learning" [2]. Wrestling with a problem yourself uses trial and error, whereas simply being told the answer stores a fact that won't come to mind when a similar but different problem appears [2]. That is the main risk in using AI to learn: if the model hands you the answer, you've skipped the part that builds skill. He ties this to the brain's learning rule. The basal ganglia learn by predicting the next reward and updating on the difference between what you expected and what you got, gradually building a "value function" [2]. That is the same temporal difference algorithm DeepMind used in AlphaGo [2]. The update only happens when you act and get feedback.
Sejnowski connects generalization in brains and in machines. Large language models work, he says, because "they're not parrots. They just, they're not, they just don't memorize what they, what they've, the data that they've taken in. They have to generalize" [2]. "The brain is really, really good at generalizing. In fact, in many cases, you only need one example to generalize," like learning how restaurants work after a single visit [2]. Human learners get that generalizing ability through active work, not by reading answers.
On AI as a partner, he is optimistic. His colleague Rusty Gage, who discovered adult neurogenesis in the hippocampus, feeds his lab's experiments and the literature into a language model, and "we use it as an idea pump" [2]. Sejnowski cites a dermatology study in Nature in which expert doctors and an AI each classified skin lesions with about 90 percent accuracy, but doctors using the AI reached 98 percent, because the AI knew rare lesions while doctors knew common ones in depth [2]. "It's not going to be adversarial. It's going to be a partnership" [2]. He adds that AI trained on past hurricanes now predicts landfall more accurately on a laptop in 10 minutes than supercomputer simulations that take days [2].
He also has practical advice for working with these tools. Try several chatbots and see which feels comfortable, noting that some are better at math. He says Google's Gemini went from 20 to 80 percent correct on math problems after fine-tuning with chain-of-reasoning [2]. He relays a New York Times account of a technical writer who was worn out after a month of working with ChatGPT until she started treating it politely, like a human assistant: "You treat it like a machine. It will treat you like a machine" [2]. Once she did, she got better answers and stopped feeling fatigued, because she was using brain circuits honed by a lifetime of human interaction [2]. Huberman described a tool into which you drop PDFs and websites and ask questions, which answers with citations and even critiques statistics [2]. He also said it's unfair to compare language models with the best humans rather than the average one, since most people couldn't pass the LSAT or MCAT [2].
Sejnowski is candid about what isn't known. He calls understanding and intelligence "weasel words because you can't pin them down," and says the debate over whether language models understand language reveals that "we don't understand what understanding is" [2]. He notes that models show in-context learning, getting better over a conversation "even though there's no plasticity. That's a mystery. We don't know the answer to that question yet. But we also don't know what the answer it is, what the answer is for humans, either" [2]. The biggest difference from humans, he says, is that a model goes blank between conversations while the human brain keeps generating its own thoughts [2].
That difference points to what learners should keep doing themselves. Sejnowski's best ideas come while walking or jogging without music or podcasts, so he records voice memos. His free course teaches students stuck on a problem to walk away, do the dishes, or walk around the block [2]. "The key is to think about it before you go to sleep," he says, because the brain works on it overnight [2]. His group found that sleep spindles, traveling waves during non-REM sleep, help consolidate the day's learning into the cortex without overwriting what's already there [2]. He also warns that shortcuts have costs. The sleep drug zolpidem doubled spindles and next-day recall in Sara Mednick's experiment, but it also erases memories of what happens after you take it, and both he and Huberman conclude that you never get anything for free [2]. The course he built with Barbara Oakley, Learning How to Learn, has been taken by about 4 million people in 200 countries. It doesn't teach subject matter. It teaches how to acquire knowledge, handle exam anxiety and avoid procrastination, and Huberman highlights self-testing as a way to find and correct errors [2].
The GIST of every new summary in your inbox, daily or weekly. Free, and you can unsubscribe any time.
Ted Gibson: what language models do and don't understand
Gibson heads the MIT Language Lab. He explains why AI can't be trusted with understanding on its own, and he gives the models real credit too. He has written that large language models are arguably the best current theories of human language: "the reason it's arguably the best is that it does the best at predicting what's English, for instance. It's incredibly good. It better than any other theory" [3]. His caveat: "It's opaque. Like there's not, you don't know what's going on. It's another black box" [3]. Chris Manning and colleagues at Stanford found evidence that models like BERT build something resembling dependency structure internally [3].
Gibson draws the line at meaning. "They're doing the form, they're doing it really, really well," he says, but "meaning is tough still. And that's why they're not, they're, you know, they don't understand" [3]. His example is a modified Monty Hall problem. If you tell a model you know the car is behind door one and the host opens door three, it still tells you to switch, because it follows the familiar shape of the puzzle. Humans rarely make that mistake [3]. Interestingly, models also struggle with center-embedded sentences the way humans do, without being trained to, which suggests they may be modeling how humans process form [3].
His brain research explains the gap. His wife, MIT neuroscientist Ev Fedorenko, can locate a person's language network with fMRI in about 15 minutes. Tasks such as spatial memory, music, math and reading computer code don't activate it at all; they use other areas, including a "multiple demands" network for hard tasks [3]. Rosemary Varley's patients with global aphasia, whose strokes wiped out their language networks, can still play chess, drive and do arithmetic as well as age-matched controls [3]. Gibson's conclusion: "language is a communication system for the meaning. It's not the meaning" [3]. Fedorenko's group, with first author Kyle Mahowald, has argued that language models lack the thinking and reasoning the brain handles separately from language [3]. For a learner, fluent text from a model is not proof that reasoning happened.
Gibson is not dismissive of AI. He thinks much of language is learned, and that language models support this view [3]. Lex Fridman pushed back, arguing that anecdotes aren't systematic evidence and that form alone gets surprisingly close to apparent understanding [3]. Gibson's point is not that models are useless. It's that fluent form and actual understanding are different things, and they can come apart.
Where They Agree, and Where They Don't
They agree that practice and feedback drive learning. Liemandt's two-hour AI blocks are built around worked examples, initial mastery and spaced repetition, measured question by question [1]. Sejnowski says expertise in any field comes from "practicing, practicing, practicing," with a value function updated on every attempt [2]. Both want more practice in schools, not less. Liemandt insists on fluency in basics, and Sejnowski criticizes cutting practice to reduce stress [1][2].
They agree AI is most valuable as a partner to human judgment. Liemandt's AI handles academics so human guides can handle motivation and struggle [1]. Sejnowski's dermatologists do best with AI, not instead of it [2]. Gibson's research shows why the pairing is needed: models are strong on form and weak on meaning [3].
They agree on being wary of chatbots that just give answers. Liemandt says most students use chatbots to cheat, and he builds structured tutors instead [1]. Sejnowski warns that being told the answer doesn't generalize [2]. Gibson shows a model confidently giving the wrong answer to a familiar-looking problem [3].
They differ on how good general chatbots are. Sejnowski encourages people to try several chatbots, treat them like colleagues, and use them as idea pumps [2]. Liemandt treats open-ended chatbots mainly as a cheating risk for students and prefers AI that assesses and teaches within a structured curriculum [1]. The difference is mostly about who is using them. Sejnowski is talking about expert adults generating hypotheses, while Liemandt is talking about children building fundamentals.
They differ on understanding. Gibson says models "don't understand," pointing to the Monty Hall error and brain evidence that language and thought are separate [3]. Sejnowski is more agnostic. He calls understanding a weasel word and says the debate shows "we don't understand what understanding is" [2]. Fridman argues that form gets surprisingly close to apparent understanding [3]. Each side has a reason. Gibson looks at specific failures and brain anatomy, while Sejnowski looks at generalization and in-context learning that nobody can yet explain. For learners, the practical result is the same: check the reasoning yourself.
They differ on struggle. Huberman argues that without errors and some "mild agitation" there's no learning [1]. Liemandt says if something can be taught without suffering, it should be, though he builds in productive struggle through challenges and high standards [1]. Sejnowski's active-learning principle sides with struggle in the form of wrestling with problems yourself, and he adds that punishment teaches faster than reward [2]. The positions fit together if "struggle" means errors and effort, which AI can make more efficient, rather than pointless suffering, which it can remove.
Why It Works
Learning runs on prediction errors. Sejnowski's reward-prediction rule means the brain improves by acting, comparing the result with what it expected, and updating [2]. Huberman makes the same point from neuroplasticity: "the error signal is what actually drives the opportunity for change" [1]. AI tutors help by producing more and faster error signals: immediate scoring on each question and later recall checks [1]. If the AI does the problem for you, there's no error signal to learn from.
Active retrieval builds knowledge you can transfer. Being told an answer stores a fact that won't come to mind for a different problem; wrestling with it builds something that generalizes [2]. Spaced repetition and recall tests, which Liemandt's apps use to confirm transfer to long-term memory, work the same way [1].
Fluency frees working memory. Working memory has limited slots, and anything memorized to fluency doesn't take one [1]. AI-driven drilling of basics like math facts and vocabulary boundaries frees capacity for harder reasoning, as with the SAT student who reached 790 [1].
Feedback loops improve teaching itself. One-on-one data let Alpha test learning science cleanly, finding for example that worked examples plus spaced repetition work for grades 8 to 12 but younger students also need initial mastery [1].
Complementary strengths add up. In the dermatology study, AI knew rare cases and doctors knew common ones in depth, so together they reached 98 percent [2]. That only works if the human keeps their own expertise.
Form isn't meaning. Language and reasoning use separate brain systems [3], and models trained on form can produce fluent answers that follow the wrong pattern [3]. Human understanding is what catches those errors.
Consolidation happens away from the screen. Sleep spindles weave new learning into the cortex [2], and ideas come during mind-wandering and before sleep [2]. A model that goes blank between conversations can't do that for you [2].
Where It Applies, and Where It Doesn't
Where it applies well. Building fundamentals, such as math facts, vocabulary, algebra steps and grade-level gaps, is where AI tutoring shows its biggest reported gains. Liemandt says a grade level in one subject takes 20 to 30 hours for K–8 students, so gaps of years can close in months [1]. Adults benefit too: the peak audience for Learning How to Learn is people aged 25 to 35 already in the workforce who need new skills [2]. For experts, AI works well as an idea generator and a second opinion [2].
Where caution is needed. Open-ended chatbots can let students skip the work [1]. Problems that look like familiar ones but differ in a key detail are exactly where models follow the form and miss the meaning [3]. Claims of 5-10x learning speed come from Alpha's own data, with randomized trials still in progress [1]. Sejnowski stresses how much remains unknown about understanding and about in-context learning [2].
What AI doesn't replace. Motivation, the "90%," still depends on caring adults, aligned interests and breaking through self-limiting beliefs [1]. Physical skills and procedures have to be practiced; you can't learn tennis from a book [2]. And the offline parts of learning, like walking, sleeping and mind-wandering, belong to the learner [2].
What To Do With It
Try it yourself before asking AI. Work the problem first, then use the model to check your answer or explain your mistake. Being told the answer skips the active learning that transfers [2].
Use AI to find your gaps, then fill them. Like Alpha's assessments, have the tool quiz you to find missing prerequisites, then go back and master them before moving on [1]. Don't accept an "80%" grasp of fundamentals [1].
Drill basics to fluency. Memorize the building blocks of your field, like multiplication tables or key vocabulary, so they stop using working memory [1]. Ask AI to show you the boundaries between similar words or concepts [1].
Follow the sequence that worked. Study a worked example, practice until you reach initial mastery, then return to it with spaced repetition and recall tests [1].
Test yourself. Use quizzes as a way to find and fix errors, not just to get a score [2].
Use AI as an idea pump, and check its ideas. Feed it your notes or sources and ask for hypotheses or critiques, as Rusty Gage's lab does [2]. Then check each idea against what you know, especially when a problem looks familiar [3].
Watch for form over meaning. When an answer sounds right, change a key detail of the problem and see whether the model adjusts, as in the modified Monty Hall example [3]. Fluent language doesn't prove the reasoning is right [3].
Treat the tool like a colleague. Ask clearly and politely and have a conversation. Users who did this got better answers and less fatigue [2]. Try more than one model, since some are stronger at math [2].
Keep the offline work. Take walks without audio to let your mind wander, record ideas as voice memos, and think about a hard problem just before sleep [2]. Protect your sleep so it can consolidate what you learned [2].
Keep a human in the loop. For children especially, pair AI tutoring with a caring adult who sets high standards, provides real support, and helps them past blocks like "I'm not a math person" [1].