3Blue1Brown: Why There's Never Been a Better Time to Go Into Math

Grant Sanderson of 3Blue1Brown argues that even if AI becomes better than mathematicians at proving theorems, math research, math funding and math careers survive. In There’s never been a better time to go into math, he answers a PhD student who wrote “just tell me what to do tomorrow” by listing what mathematicians actually do, where the value of math comes from and who pays for it. Then he asks what happens to each piece as machines get better at proofs, at conjectures, at theory building and finally at everything. His conclusion is that the biggest change is to how academic credit is assigned, not to whether the field exists, and that a field this unsettled is the best possible place for a newcomer.

TLDW

Sanderson splits math into a research stack (proving theorems, posing questions, building theory, applications, communication), a value stack (downstream applications, transferable problem-solving skill, aesthetic joy) and a funding stack (teaching, which is the biggest, then public grants, private foundations and industry). He walks through four escalating scenarios of AI capability and finds that proof automation barely touches value or funding but upends how credit is assigned. Proof should now be treated as raw data for theory building, and education stays human because it is limited by motivation, not by clear explanations. His advice for tomorrow: be social, give lots of talks, lean into whatever makes you love math, and go on a long walk with a fellow student to work out what you want to contribute.

Thoughts

The title claim turns out to have nothing to do with AI unlocking new research, and that is what makes it convincing. Sanderson says the reason now is a great time to enter math is that the community knows it has to change and does not know how, so senior people are looking around for examples. Institutions with settled norms reward people who fit the norms. Institutions in flux reward people who show what the new norm could be. His own example lands: ten years ago a visual-first explanation project would have been laughed out of almost any department as a PhD thesis, and today it might not be. That is a career argument that applies well beyond math. When incumbents are confused about what counts, a newcomer’s work can set the standard instead of being measured against it.

The sharpest intellectual move in the video is reframing proof as data. For centuries a proof was the proxy for understanding: a famous problem counted as “solved” the moment a valid proof existed. Sanderson argues machine-generated proofs should be treated like numerical calculation, useful input for theory building and not an end in itself. His line that a forced blow-up result for Navier-Stokes means nothing to him unless someone understands the ideas and grows a theory from them is a clean statement of what math was always supposed to be for. It also explains much of the current anger. If credit has been attached to proofs for decades, then a lab dropping hundreds of proofs at once feels like theft of credit, even when no understanding was taken. Change what earns the credit and much of that grievance goes away.

The point about teaching deserves more attention than it will get. Sanderson notes that teaching, not grants, is where most math money comes from, and then argues that education “is not rate limited by the clarity of explanation, it’s rate limited by motivation.” That is a strong claim from a man whose whole career is making explanations clearer, and it matches what anyone who has watched a free online course go unfinished already knows. People learn because someone they respect visibly cares. If that is right, the jobs most exposed to AI are not the human-facing ones but the ones that were only ever about producing an output. It also gives his advice to “give lots of talks” a practical edge: oral presentation becomes the way to show what you actually understand once written work can be machine-assisted.

His most honest stretch is the “anything you can do I can do better” scenario, where machines beat humans even at applying math to science. Here he admits that if funding is justified purely by downstream applications, the rational move would be to pour it into “some big blob of compute that just does math.” He does not think policymakers will flip that switch, but he does expect newly rich AI founders to set up private foundations that do exactly that. That is a believable prediction and a slightly unsettling one. It suggests a split future where machine math is privately funded for results and human math is publicly funded the way the arts are, for flourishing and for the problem solvers it trains.

The gym analogy runs through the whole argument and holds up in every scenario. Machines lift heavier weights than you, and that has never made lifting pointless. Training problem-solving skill and feeling the click of a solution both require the human to do the work, which is why Sanderson keeps finding those parts of the value stack untouched. The weak spot, which he flags himself, is assessment. If problem sets can be done by a model, assigning and grading them stops working as training, so the institution has to find new ways to make people do the reps. That is the practical problem every math department faces this semester, and it is more urgent than any question about the far future.

Key Takeaways

Sanderson’s thesis is that there has never been a better time to go into math or to fund it, and that the biggest change AI brings is to how academic credit is assigned.
Math research breaks into proving theorems, deciding which questions are worth asking, theory building, applications, and communicating all of it clearly.
The value of math comes from downstream applications (complex numbers to quantum mechanics, non-Euclidean geometry to general relativity, elliptic curves to cryptography), transferable problem-solving skill, and aesthetic joy.
Teaching is the biggest source of math funding, ahead of public grants like the NSF, private foundations like the Simons Foundation, and industry.
AI progress in math follows “position overrated, velocity not”: from barely adding numbers four years ago to olympiad problems to serious Erdős problems that mathematicians had tried hard to prove.
Applications almost always come from theory, not from single theorems. Group theory mattered more than the unsolvability of the quintic.
Proof automation leaves the value and funding stacks nearly untouched but makes each mathematician more leveraged in building theory.
Mathematicians’ reactions split: some feel the joy is gone, others are exhilarated, and he sees more negativity online and more excitement in person.
Machine proofs often reuse ideas from human papers without attribution, so researchers have to add that credit back themselves.
Proof should be treated like numerical calculation: raw data for theory building, not the end goal.
If machines beat humans at everything, application-driven funding would rationally go to compute, and he expects wealthy AI founders to start private foundations that do just that.
Public funding would likely shift toward the arts model: math done by humans, for human flourishing and for training problem solvers.
Education is limited by motivation, not by how clear the explanation is, which makes teaching one of the safest jobs from AI displacement.
Grading problem sets on their own no longer makes sense as training, and oral presentation will matter more and more as the way to measure understanding.
His concrete advice: lean into what you love about math, give lots of talks, and go on a long walk with a fellow student to talk about why you each chose math.

Chapters

0:00 Math Student Angst

Sanderson opens with the questions he got from students and professors after asking for input on AI and math, many of them simply asking what to do next. He picks one comment, a math PhD saying “just tell me what to do tomorrow,” as the anchor for the video. He says he has no stake in pushing people into academia and thinks too many students pursue PhDs because school is comfortable. His audience is the people thinking of leaving out of fear that the career will vanish or become joyless.

3:10 How Does Math Work?

He lays out the research stack, citing the saying that good mathematicians prove theorems, great ones make conjectures and the greatest write definitions, with Descartes, Galois and Grothendieck’s “rising sea” as examples of theory building. The value stack covers applications that arrive decades or centuries later, problem-solving skills that carry into computer science, finance and AI safety, and the aesthetic joy that he links to artists’ reactions to AI art. The funding stack is mostly teaching, because math research is cheap and math departments teach far more than their own majors. The individual goods, skill and joy, flow into the money chain through that teaching.

11:24 A Proof Machine

Drawing on interviews he has recorded with mathematicians since mid-2025, Sanderson traces AI from bespoke DeepMind geometry solvers to general language models solving olympiad problems and then serious Erdős problems. Assuming machines become categorically better at proofs, he finds little change to value or funding, since applications come from theory and problem-solving practice works like the gym. He plays a clip of an Institute for Advanced Study mathematician thrilled that a model cracked a major open problem in his field, and gives his first piece of advice: close the laptop and go for walks with fellow students. He then argues proof should be treated as raw data and that oral communication grows in importance.

25:12 A Conjecture Machine

Next he imagines models that are better at choosing which questions are worth asking. If humans still lead on theory building, each mathematician becomes even more leveraged, so funding still makes sense. The split on joy continues: some people will not enjoy a model posing and answering its own questions, while others will feel freed. Sanderson says he will probably spend his life thinking up his own questions regardless of what models can do.

26:59 A Theory Machine

What if machines also build theory, the equivalent of handing a model math up to the 1500s and getting complex numbers back? Humans would still need to understand that theory well enough to apply it to cancer, fusion or superconductors. Here he separates two kinds of joy: the joy anyone feels when a solution clicks, which is unaffected and social by nature, and the frontier researcher’s joy of being first, which he admits would be a real loss. He expects subcurrents of the math community to champion human understanding for its own sake.

31:31 If AI Is Better at Everything

In the “anything you can do I can do better” future, the rational use of application-driven money would be a big blob of compute doing math. He doubts governments would redirect public funds that way but expects private foundations from AI wealth to try it. Public funding would more likely be justified by trained problem solvers and by math as an art. Since most money comes from teaching anyway, and teaching depends on human motivation and relationships, he expects universities to look much as they do now in a future that goes well.

34:58 The Best Time to Enter Math

Sanderson concludes that research must lean on theory building, that proof is losing its value as a proxy, and that training methods like graded problem sets need rethinking. He points to the public appetite for math, from his own channel and Brady Haran’s Numberphile to Martin Gardner’s Scientific American column, and urges the math community to turn outward. The best-time argument is that a field in flux lets newcomers shape it. His advice: build the thesis you love, whether that means using proof machines or rejecting them, give lots of talks, and take a long walk with a friend from your program.

Notable Quotes

“The position is overrated but the velocity is not.”
Grant Sanderson, on his main takeaway from interviewing mathematicians about AI since mid-2025

“The frontier is spiky at a fractal level and it is not at all uniform.”
Grant Sanderson, on how differently AI helps across fields and individual mathematicians

“I think at this point we should consider proof to be kind of the same as numerical calculation, which is to say it’s a form of raw data that is helpful for theory building, but it is certainly not the end in and of itself.”
Grant Sanderson, on rethinking what math research is for

“The proof is just data, it’s like giving me more digits of π.”
Grant Sanderson, on why a machine-generated Navier-Stokes proof matters only if someone builds understanding from it

“Education is not rate limited by the clarity of explanation, it’s rate limited by motivation.”
Grant Sanderson, on why people will still want to learn from humans

“Proof in and of itself has probably been overvalued for decades.”
Grant Sanderson, on why changing how academic contribution is measured is a good thing

“So you, as a new entrant into that field, stand to actually influence what it becomes at one of the most pivotal times in its history.”
Grant Sanderson, on the real reason this is the best time to go into math

Sanderson recommends 100 Reactions to 100 Solutions on Proofs and Prompts for a sense of how mathematicians are feeling right now. Watch the full video here.

3Blue1Brown: Why There’s Never Been a Better Time to Go Into Math