Grant Sanderson (@3blue1brown) — Past, present, & future of mathematics
What this covers
Grant Sanderson (@3blue1brown), the mathematician and educational video creator, sits with Dwarkesh Patel to sketch how mathematics is actually learned and created—and how it differs from popular narratives about genius and threat. The conversation spans the history and future of mathematical knowledge, the social structures that make learning stick, the role of Tools for Thought in how we think, and whether AI will render human mathematicians obsolete. Sanderson argues that mathematical achievement in domains like the IMO represents a discrete creative milestone—akin to mastering Go or chess—that is orthogonal to general intelligence or job replacement. The core thread running through is that education depends far less on explanation quality than on motivation, peer influence, and the presence of a mentor or compelling project.
The conversation touches several distinct territories. On learning, Sanderson contends that social factors—friend groups and real projects—matter vastly more than the proliferation of online explanations since the printing press. He traces this through the Curse of Knowledge, the difficulty of constructing explanations that don't assume prior context, and his view that educational technologies stratify rather than democratize because they fail to address the motivational barriers that cause underperformance in traditional schools. On mathematics itself, he examines how axiom systems need a Motivating Problem to guide which mathematics gets discovered, why notation functions as a medium for thought itself, and how mathematical progress is more collaborative and world-responsive than the lone-genius myth suggests. On AI, he separates the question of whether machines can win the IMO—which would be historic but narrow—from whether they will displace human workers, which depends on entirely different capabilities. Throughout, Sensitivity to Initial Conditions emerges as his frame: how a teacher's thirty-second intervention can pivot a student's entire trajectory, and how constraints like deadlines in Summer of Math Exposition unlock latent creative potential.
Sanderson argues that mathematical achievement and creativity (like solving IMO problems) are domain-specific milestones distinct from general intelligence or job replacement, and that effective education depends far more on social motivation, mentorship, and personal projects than on the quality or availability of explanations.
- Getting gold in the IMO is analogous to mastering Go or chess—a discrete creative domain—not a sign of imminent general capability or job replacement
- The sensitivity of a student's trajectory to small mentor interventions shows education is fundamentally social, not informational
- Since the printing press, explanations have been widely available; what's scarce is motivation, peer groups, and compelling projects
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
There are two types of math proofs: ones where you morally understand why the result should be true, and ones where you can follow each step but are left wondering why anyone would have come up with that approach—the latter leaving you wanting something more.
“There are the ones where you morally understand why the result should be true and then there are the ones where you're like, "I can follow the steps. Why would you have come up with that?"”
Anthropological studies show unschooled people naturally place the number halfway between 1 and 9 at 3 rather than 5, indicating humans natively think about quantity logarithmically rather than linearly, which makes sense because increases in social complexity scale roughly equally across multiplicative jumps; acquiring formal numeracy may have caused us to lose this intuition, which is why logarithms are now hard to teach.
“if you ask what number is halfway between one and nine, those groups will answer three whereas you or I or people in our world would probably answer five”
What most determines whether people actually learn is not the quality or availability of explanations—which has been abundant since the printing press—but social factors: whether your friend group pushes you up or pulls you down, and whether you have a compelling reason like a job or personal project; compelling projects and encouraging peers do far more than the average online explanation, which only lubricates the process once motivation exists.
“what's probably most important to getting people to actually learn something is not the explanation or the quality of explanations available because since the printing press that has not been true”
Because the space of possible axiom systems and resulting mathematics is unfathomably huge, you need a process—usually a motivating real-world or curiosity-driven problem—to cull which axioms and questions are worth pursuing; pure math textbooks fail by omitting why a given set of axioms was found useful rather than some other.
“the space of possible math is unfathomably huge. So you need some process that culls down what are the useful things to maybe pursue.”
The word education derives from 'educe' (to bring out), reminding us that an educator's job is not to shove knowledge into students' heads but to bring it out—providing an environment that elicits as much as possible through inspiration, projects, mentorship, eye contact, and being a real figure in students' lives—which a video fundamentally cannot do; videos provide explanation, which is distinct from education.
“the word education derives from the same root as the word educe, to bring out ... the job of an educator is not to like take their knowledge and shove it into the heads of someone else the job is to bring it out”
Good explanations are scarce for two reasons: knowing something makes it hard to remember what it was like not to know it (especially at higher abstraction), and the best explanation depends heavily on the individual learner—so the golden challenge is constructing explanations as generally useful as possible without assuming shared context, which professional explainers rarely attempt because their audiences are narrowly targeted.
“there's a difference between knowing something and then remembering what it's like not to know it”
Educational technologies repeatedly promise to disrupt schools and never do; instead they are stratifying—very good for those already motivated or already at the top, while those performing more poorly struggle the most, possibly because the same factors that cause poor performance in traditional systems also prevent good engagement with the plethora of available tools.
“how stratifying these technologies will be in that they actually are very very good for those who are already motivated or kind of already on the top in some way and they end up struggling the most just for those who are performing more poorly”
Self-learners shoot themselves in the foot by skipping calculations believing they're incidental to core understanding, when in fact putting in the reps of calculations is where underlying intuitions solidify and where you discover what the substance of a result actually came from; without homework's forcing function, people skip this, so a notebook and pencil should be considered part of the reading process.
“where a lot of self learners shoot themselves in the foot is by skipping calculations by thinking that that's incidental to the core understanding. But actually, I do think you build a lot of intuition just by putting in the reps of certain calculations”
A deadline is what actually makes the difference between people completing a creative project and procrastinating on it—participants in SoME often already had the project idea and the potential energy, and the event's main function was to provide a deadline and a promise of reach.
“give you a deadline, which let's be honest is what actually makes the difference between people doing something and procrastinating on it”
Contrary to the popular misconception that math is a field of lone geniuses at blackboards, it is highly collaborative—the typical number of authors per paper is much larger now than before, and progress in niche fields requires travel and conferences to work with the few others in that subfield.
“people have this misconception that math is a field with lone geniuses who are coming up with great insights alone next to a blackboard. The reality is that it's a highly collaborative field”
Good notation works like fluency with a programming interface: you write something with few symbols and then discover non-trivial implications because the constrained rules of algebra let you run an almost game-like process of reduction and expansion—a similar fluid medium for simulation and programming could one day shift how we even think about problems.
“you write something with a small number of symbols but then you discover a lot about the implication of that ... you can go through an almost game-like process to see how it reduces and expands”
Mathematics is developed more in the service of the world one lives in and the adjacent problems it solves than is typically recognized; information theory could have been defined purely much earlier but wasn't until error correction over real information channels at Bell Labs prompted it, and chaos theory required computers to even observe the phenomenon of sensitivity to initial conditions before the math could be discovered.
“maybe this is a sign that the math that's developed is more in the service of the world that you live in and the adjacent problems that it's used to solve than we typically think of it”
Math is uniquely suited to synthetic data generation for training AI in ways other domains are not, because you can produce many proofs in a proof-checking language like Lean and train on whether each is valid, similar to how AlphaGo was trained by self-play.
“Math lends itself to synthetic data in the ways that a lot of other domains don't. You could have it produce a lot of proofs in a proof checking language like Lean”
Gödel's incompleteness theorem practically doesn't come up for working mathematicians because the constructed contradiction has the same self-referential flavor as 'this statement is a lie'—a pathology rarely encountered in natural questions; for problems like the Riemann hypothesis or twin primes, mathematicians strongly expect there is an answer, and the theorem's relevance sits only at the edges of theory (e.g. Paris–Harrington).
“From what I understand, it really doesn't come up.”
An AI achieving gold in the IMO would be a historic creative milestone analogous to mastering Go or chess, but it is orthogonal to replacing human jobs, because the impediments to job replacement are different things—like longer context windows, building relationships, and understanding where someone is coming from—rather than the discrete problem-solving creativity math competitions reward.
“That feels orthogonal because getting a gold in the IMO feels a lot more like being really, really good at Go or Chess.”
The decisive factor in whether a good video reaches its audience is not flaws in YouTube's recommendation algorithm—which is quite good—but the mismatch between the proxies used to nominate videos into the machine-learning-driven recommender, because running a full neural network over all videos is computationally infeasible, so proxies like the co-watch graph determine which ~1000 candidates even enter the running.
“It's not the flaws in the algorithm. The algorithm is probably quite good. It's the mismatch between the proxies being used to nominate stuff”
Below Dunbar's number, a tit-for-tat reciprocity structure makes more sense than formal accounting, so it would be obnoxious to track everything numerically—which may explain why early small communities didn't develop commerce, money, or numeracy.
“when you're below Dunbar's number in your communities, a tit for tat structure just makes a lot more sense and actually works well and it would just be obnoxious to actually account for everything”
Much of mathematics—even high-school-level content—is surprisingly recent because for most of history almost no one was a pure mathematician; people were mathematicians plus something else, and only as population grew and more people gained the economic freedom for academia did the man-hours devoted to developing new theorems spike.
“How many people did pure math for most of history? For most of history, no one. No one was a pure mathematician.”
The concept of AGI implies a discrete change—a benchmark after which a system becomes 'generally intelligent'—but this discrete step doesn't actually exist; intelligence in AI is continuous, with GPT-4 already being general in that one training algorithm applies across many tasks while still not exceeding the best humans at any given domain.
“it seems like what people want to get at is a discrete change that I don't think actually exists”
The scale on which one person is two standard deviations above another may not be the scale that matters; contributions to a field don't always come from genius flashes of insight but often from becoming an expert in two different things and finding connections between them, so a static g-factor mental model may be the wrong frame.
“contributions to these fields don't always look like genius insights and that sometimes there's fruit to be born from say becoming kind of an expert in two different things and then finding connections between them”
A teacher's tiny intervention—a 30-second comment that a student is good at math, or accidentally handing them a research problem—can have monumental rippling effects that completely shift a student's life trajectory, demonstrating that students sit at a nexus of extreme sensitivity to initial conditions where an educator can swing their entire future either way.
“the thing that takes at most 30 minutes of the teacher's time, maybe even 30 seconds, has these completely monumental rippling effects for the life of the student”
Grant does the mundane details of video production himself rather than delegating because for him the seemingly mundane editing and animation-coding details are the very medium through which he thinks about constructing the piece—the cuts are his way of figuring out what the final piece should be, and articulating that in words to a collaborator would be harder than just doing it.
“the seemingly mundane details are sometimes just how I even think about constructing it in the first place”
Creators who jump in aiming to maximize audience end up making content that is best for no one because they compete with all other communicators while still unskilled, whereas doing something unreasonably niche with no expectation of blowing up produces the best possible version of that content for its small audience, since no one else is serving that narrow target.
“when you jump in with that as a goal you kind of aim for too large an audience and end up making the content which is best for no one”
The high quality of Summer of Math Exposition entries is not driven by the modest cash prize pool—people aren't motivated by an expected-value calculation, and a very large prize could actually harm the event by inducing delusional over-investment and disputes; the real draw is the deadline, the chance of reach via peer review, and tooling being good enough for individuals to make incredible things.
“the impression I get is that people aren't fundamentally motivated by winning some cash prize”
The world would be better if more people savvy with STEM spent some of their time—not necessarily their whole career—teaching, because it is one of the highest-leverage ways to get more people to engage with math; ad hoc on-ramps like emergency credentials and Teach for America make this feasible, and Grant intends to do a high school teaching stint himself to avoid hypocrisy and preserve his empathy with learners.
“I would like to live in a world where more people who are savvy with STEM spend some of their time teaching. I just think that's one of the highest leverage ways that you can think of to actually get more people to engage with math”
Because students who fall in love with math are by default funneled into apprenticeships toward academia, the three most common destinations—math academia, finance, and computer science/data science—almost certainly have an over-allocation of mathematical talent, even though all three are valuable.
“as a result, they almost certainly have an over allocation of talent. All three of those are valuable”
Building your own tool (like Manim) rather than using existing frameworks may be inefficient, but the level of ownership unlocks a sense of creative freedom—the ability to change the tool to do whatever you want—that may be necessary to take on a wide variety of topics.
“the level of ownership that you feel once it is your own thing that just unlocks a sense of creativity”
The most valuable asset for an online explainer is empathy for what viewers don't understand, and that empathy erodes as life circumstances diverge from the audience's; therefore even the best online educators should periodically return to classrooms to stay sharp rather than being permanently removed from teaching.
“The more distance that you put between yourself and them in terms of life circumstances ... just makes it more and more of an uphill battle to make the content good for them”
There hasn't been a rise of pedagogical video games comparable to explanatory videos because games are very hard and resource-intensive to make; pedagogy-motivated games tend not to be fun, the genuinely effective ones aren't directly educational, and unlike short videos where it's fine that most fail, each game costs the user many hours so the rate of production is too low for a hit to emerge.
“games are very hard to make. It takes a lot of resources for a given game and whenever people seem to try to do it with pedagogy as a motive it seems to be the case that they are not fun”
The biggest constraint on a working mathematician's progress is likely the number of available analogies they can draw on—exposure to disparate fields lets them transfer a problem-solving approach from one area to another—which also explains why math has become highly collaborative with rising author counts and heavy travel, since exposure to others' fields maximizes available analogies.
“the short answer to your question, what is nitrogen for mathematicians, is the analogy”
Inserting a forcing function—such as an NSF grant requirement that 10% of a pure mathematician's time go to collaboration with another department—could beneficially distribute mathematicians' specialized problem-solving talent into non-pure-math settings.
“NSF grants coming with a requirement that 10% of your time goes towards a collaboration with another department or something like that.”
A scientist's 'miracle year' is the exhalation following many years of inhalation—accumulated potential energy from prior thinking finally released at one opportune time—combined with youth's freedom from obligations and the absence of success-driven distractions that later eat into creative flow.
“One thing could be that the miracle year is like the exhalation and there's been many, many years of inhalation.”
Image-generating AI shows something beyond the literal is understood: asking for a landscape made of quartz and gemstones yields something with the essence of a landscape that isn't literally one—the same lateral-thinking creativity required to solve a math problem you can't derive straight from definitions, which is why we find joy in finding such solutions.
“it gives you this thing which has all of the essence of a landscape, but it's not literally a landscape. And so you realize that there's something beyond the literal that's understood here”
It's acceptable that most math videos on YouTube are bad because the medium only requires that enough good ones exist for a searcher to find one that scratches their itch or to be recommended something that sparks a new interest—a tolerance for failure that the high time-cost of games cannot afford.
“It's okay that most of them suck. It's okay because you just need enough that when someone searches for the term that they want they get one that is good”
GPT-4 is general in the sense that one training algorithm applies to a very large set of different kinds of tasks, which is an invention people in the 1960s might not have expected to be true about the nature of how artificial intelligence could be programmed.
“GPT-4 feels general in the sense that you have one training algorithm that applies to a very, very large set of different kinds of tasks”
Despite advances in programming and simulation, paper remains the best medium for actually doing mathematics together; only when the right medium for digging into a problem becomes simulation and programming will the way we think about math fundamentally shift.
“there's still so much more value in paper ... The right medium to do that is still paper”
Projects that later become successful are often started almost on a whim as a low-stakes hobby in college, off-track from one's planned career, because the lack of expectation of growth removes disappointment and lets the work become more in-depth for its early audience.
“these kinds of projects how often something later on ends up being successful is something that was started almost on a whim as a hobby when you're in college”