3Blue1Brown's Grant Sanderson: Why the AI era is the best time to do math
量子位 · wechat · 2026-10-19
Grant Sanderson devoted a video to arguing that even if AI fully surpasses humans at proving theorems, math's value and funding won't collapse — the field may be entering its best era.
Key points
- Math research spans five activities: proving theorems, posing conjectures, building theory, finding applications, communicating. Proofs are not all of math.
- Math's three values: applications (imaginary numbers, non-Euclidean geometry took centuries to pay off), transferable problem-solving skills, and aesthetic joy.
- Funding (teaching, public/private grants, industry) remains justified even if machines do the proving.
What AI changes
- "AI's current ability is overestimated, but its rate of progress never was": from DeepMind's IMO geometry result in 2023 to general models solving IMO-level problems today, in just two years.
- Proofs become "raw data"; research tilts toward theory-building — distilling elegant structures from mass-generated proofs, reshaping what a PhD thesis looks like.
- Education's bottleneck is motivation, not explanation quality; machines solving problems doesn't stop you from growing stronger — "a machine lifting weights for you doesn't build your muscles."
Conclusion: newcomers now can help redefine which mathematical work gets recognized. If Grant were a PhD student in 2026: "do 3Blue1Brown — invest in what you love."
Related event: 3Blue1Brown: There's Never Been a Better Time to Go Into Math(10 posts)→
More from AGI Musings
- Researcher Dan Faggella: negotiation won't save us from AGI — radical self-augmentation might — danfaggella · 2026-10-21
- Vercel's rauchg: AI will compute the perfect delta of your understanding, delivering free ultimate education — andreisavu · 2026-10-21
- Will the Bitter Lesson crush optimization research? Igor Carron thinks so — IgorCarron · 2026-10-21
- John Lam: keep your data estate out of OpenAI/Anthropic silos to stay swappable — john_lam · 2026-10-21
- Founder: consumer AI companies treat privacy and trust as an afterthought — vhanagwal · 2026-10-21
- Do you say please and thank you to your agent? Reddit debates AI politeness — Luvena21 · 2026-10-21