ExpertsFMGrant Sanderson – AI and the future of math
Dwarkesh Podcast · June 30, 2026
5 Selected Lessons:
heuristic · 2 reasoning_pattern · 1 case · 1 framework · 1
Know where the money comes from before choosing a career
Source
Dwarkesh Podcast — Grant Sanderson – AI and the future of math (June 30, 2026) · at 1:22:35 ▸
When this applies
when anyone is choosing a career path and worried a skill they love might be automated or devalued
The takeaway
Before committing to a field, explicitly identify the actual economic mechanism funding that role (prestige/brand value, public-good subsidy, direct service value) and ask whether your contribution maps to it, independent of whether you're intrinsically good at or interested in the underlying skill.
Why it works
People who are good at a subject tend to default into careers in that subject because they've been rewarded for proceeding through its hoops, without separately asking who is paying for their work and why — brand value, public-good funding (grants), and direct teaching/service value are each distinct revenue sources with different vulnerability to automation.
One level up
Separate 'what I'm good at / enjoy' from 'why someone pays for this' — career resilience depends on understanding the latter, especially as automation changes which skills are scarce.
Where it stops holding
This stops working when someone is deliberately choosing a non-economic path (Grant's 'gentleman scientist' caveat) where value-creation/money isn't the relevant frame at all.
Detail
procedure:
- Identify what you're actually being paid for in a prospective role — is it brand value to an institution, public-good belief funding (e.g., NSF grants), direct teaching/service, or something else.
- Check whether AI automation is likely to erode or enhance that specific mechanism (not just 'the skill' in the abstract).
- Choose/adjust career strategy based on that economic mapping rather than only on intrinsic interest or aptitude.
rationale: Grant notes this question is 'surprisingly' rarely asked, even pre-AI, and that explicitly asking it already puts a student 'steps ahead' of peers.
In their words
You should understand where the money is coming from and what value you're actually adding and the connection between those two. I think often a surprisingly small amount of thought is put towards that.
Three archetypes of a hypothetical AI proof (lightning bolt, mountain, brute hustle)
Source
Dwarkesh Podcast — Grant Sanderson – AI and the future of math (June 30, 2026) · at 4:06 ▸ · at 22:01 ▸ · at 25:16 ▸ · at 28:40 ▸
When this applies
when trying to anticipate whether an AI-generated solution to any hard problem will be human-legible or an opaque black box
The takeaway
The interpretability cost of a discovery depends on its structural type — reused existing concepts bridge easily, wholly new concepts require costly digestion, and sheer length defeats human parsing regardless of correctness.
Why it works
A solution formed by connecting two known expert domains ('lightning bolt') is cheap for humans to verify since only endpoints need checking; a solution requiring an entirely new conceptual framework ('mountain building') demands humans rebuild understanding from scratch; a solution that's just a very long chain of reasoning with no new theory ('brute hustle') is verifiable but possibly unintelligible due to sheer length.
One level up
When assessing whether a powerful system's output will be usable, classify it by structural type, not just by whether it is 'correct.'
Where it stops holding
This stops working when the three archetypes blend together in a given solution, making the classification itself ambiguous or when verifiers disagree on which type a result represents (as with the contested ABC conjecture solution).
Detail
approach: Classify a candidate proof/discovery as (1) a lightning-bolt connection between known expert fields, (2) new mountain-building requiring fresh theory, or (3) raw brute-force long-chain reasoning with no new ideas.
signals:
- Can the idea be summarized to an expert in one sentence and they 'get it' immediately (lightning bolt)
- Does understanding require learning an entirely new framework of abstractions (mountain)
- Is the proof just extremely long with no compressible insight (hustle)
rationale: Each type has a different cost/benefit for human understanding and a different verification/trust profile.
In their words
If the character of it is mountain building, you have to put in a lot more time to understand that new mountain that was built because it's a new thread, that it's not just lightning bolt between them.
Benchmarks equal training environments
Source
Dwarkesh Podcast — Grant Sanderson – AI and the future of math (June 30, 2026) · at 7:13 ▸ · at 11:32 ▸
When this applies
when assessing whether any capability can be rapidly automated by current ML methods
The takeaway
A capability becomes automatable roughly as fast as it becomes measurable; if you can't score it, you can't currently train it.
Why it works
If a task's success can be clearly scored (a benchmark), that same scoring signal can be turned into a training reward, so clearly-scoreable tasks fall quickly while fuzzy, subjective tasks resist training.
One level up
The boundary of what a technology can learn to do is set by what can be verified, not by raw difficulty.
Where it stops holding
This stops working when a task's true value is inherently subjective (e.g. the quality of a conjecture or a piece of writing) and no proxy benchmark captures it, or when future training paradigms evolve beyond reward-only methods.
Detail
rule: Benchmark-able tasks become trainable tasks; non-benchmarkable, subjective tasks (like conjecture generation) resist current training paradigms.
threshold: If you can write a clear pass/fail headline ('AI solved X'), it's benchmarkable and will likely be trained for soon; if the best you can say is 'the tone of conversations shifted,' it isn't.
rationale: There's no fundamental difference between a benchmark and a training environment in current ML practice.
In their words
there's really no fundamental difference between a benchmark and a training environment.
Formalization (Lean) matters less for current progress than people assume, but has a distinct future value
Source
Dwarkesh Podcast — Grant Sanderson – AI and the future of math (June 30, 2026) · at 52:50 ▸ · at 56:59 ▸ · at 1:02:38 ▸
When this applies
when choosing between a rigorous but labor-intensive formal verification system and a faster, approximate natural-language/human-judgment verification system
The takeaway
Don't assume the most rigorous tool is the most load-bearing one for current progress; but also don't discount rigorous tools entirely, since they enable a qualitatively different mode (indefinite autonomous exploration without human check-ins) that approximate methods can't match.
Why it works
Natural-language verifiers (even approximate, meta-verified ones) have proven sufficient for near-term progress because they're easier and faster to generate signal from, but formal systems offer a unique unbounded, unsupervised exploration capability (like AlphaZero self-play) that natural language math cannot provide since it always needs human review to avoid value drift.
One level up
Separate 'what's necessary for current incremental progress' from 'what's necessary for a qualitatively different future capability' — the answer can differ for each.
Where it stops holding
This stops working when natural-language proofs' error rates compound and erode trust (making every unverified paper 'insufferable' to review, as Grant notes), at which point formal verification becomes necessary again even for everyday progress.
Detail
rule: Natural language + a trained verifier (validated by a meta-verifier) can substitute for formal Lean proofs for most near-term progress, but Lean/Mathlib retains unique value for autonomous, unsupervised, open-ended exploration and for giving mathematicians trustworthy 'green check marks' on error-prone natural-language outputs.
threshold: If initial attempts succeed in Lean but subsequent versions succeed equally well in natural language (as happened with IMO year-over-year), formalization wasn't the bottleneck.
rationale: DeepSeek math's verifier-plus-meta-verifier approach worked in practice; and the chain-of-thought for the unit distance disproof had no Lean in it at all.
In their words
I think you are right that Lean is maybe overrated on the side of the importance of it being used as a VR environment for any kind of, like, just progress in math generally. But I I definitely wouldn't write it out of the story.
The 100-year verification loop for Galois theory
Source
Dwarkesh Podcast — Grant Sanderson – AI and the future of math (June 30, 2026) · at 11:32 ▸ · at 13:41 ▸ · at 18:40 ▸ · at 22:01 ▸
When this applies
when evaluating any genuinely novel idea, framework, or technology whose usefulness can't yet be benchmarked and might take decades to prove out
The takeaway
The time between a concept's invention and its recognized value can be extremely long and is not correlated with the rigor of early peer review; rejection by contemporary verifiers does not mean an idea lacks value.
Why it works
Truly novel conceptual frameworks often precede any application by decades because recognizing their value requires other fields (cryptography, physics) to mature enough to need them; early human reviewers (even expert ones) frequently reject or ignore them for lack of immediate payoff.
One level up
Don't mistake short-term verifiability for long-term value — the most transformative ideas may fail every near-term test.
Where it stops holding
This stops working when an idea truly is low-value and its rejection is correct, or when markets/fields demand faster feedback loops that don't tolerate century-long uncertainty (e.g., most commercial ventures).
Detail
situation: Lagrange proposed that symmetry of polynomial roots might be the key to solving/disproving a quintic formula; Abel built on this to prove impossibility; Galois pushed abstraction further but his papers were rejected as incoherent and he died young.
decision: Galois's notes were passed on by friends/brother rather than abandoned, eventually reaching Liouville two decades later, who tried to clean them up.
outcome: It took another ~20 years after Liouville for Jordan to formalize what we now call group theory — roughly a century after Lagrange's original insight, and decades after the quintic question was even 'solved' by Abel.
lesson: The verification loop for genuinely new conceptual frameworks can span a century and pass through repeated rejection before being recognized as valuable.
In their words
between the time of Lagrange, having this inkling of maybe symmetries of roots is the right way to go, to where it all looks like modern group theory. You've got this long span, a lot of the time it's not even passing the verified reward of human reviewers, because it gets on someone's desk, they say, I don't really know if there's anything here.