a16z Frontier show

Mathematicians discuss AI's impact on discovery versus understanding, formal verification, and the future of the field.

1h 30m Listen to the original Episode page
Artificial IntelligenceMathematicsEducationFormal VerificationResearch MethodologyAI AlignmentPost-Quantum Cryptography

Summary

Speakers Scott Duke Kominers, Dan Boneh, and Ittai Abraham explore how AI is separating the 'discovery frontier' from the 'understanding frontier' in mathematics. They note that while AI can generate formal proofs and solve problems rapidly, humans remain essential for interpreting results, developing 'taste' for problem selection, and ensuring logical consistency. The conversation covers the role of Lean formalization, the potential for 'billion-page' proofs, and the psychological impact on researchers. Educators are advised to shift focus toward higher-level abstractions and communication skills as AI handles routine computation.

Key takeaways

  1. 1
    Society depends on systems like post-quantum cryptography, which AI is not yet good enough to solve, creating a near-term demand for human expertise in critical security domains.

    Unidentified speaker (speaker_2) · 1:21:25

  2. 2
    Speaker_1 recalls stating in January that graduate students would be the first generation to use LLMs for substantive research, noting this prediction came true within seven to eight months.

    Unidentified speaker (speaker_1) · 1:17:19

  3. 3
    Speaker_3 identifies AI alignment as a huge challenge, suggesting that advancing human mathematical understanding is crucial for ensuring AI remains aligned with human values and goals.

    Unidentified speaker (speaker_3) · 1:26:22

Episode sections

  1. 0:00 – 0:53

    Intro: Academic System Strain and Human Judgment

    Speakers discuss whether the academic system is collapsing due to a tripling of paper rates, noting pressure on knowledge consolidation institutions. They argue that while AI can provide answers, human judgment in selecting problems and deriving understanding remains essential.

  2. 0:53 – 10:38

    The Andrew Wiles Story and the 'Discovery vs. Understanding' Frontier

    Scott Duke Kominers contextualizes Andrew Wiles' proof of Fermat's Last Theorem as part of a larger mathematical edifice. Dan Boneh introduces the concept of two frontiers: the discovery frontier (AI solving problems) and the human understanding frontier (mathematicians comprehending the results), which are currently separating.

  3. 10:38 – 18:40

    AI Progress and the 'Phase Transition' in Capability

    Ittai Abraham recounts spending two years on metric space embedding problems, only for an AI to make significant progress in one hour. Dan Boneh describes a 'phase transition' in June 2026 with GPT-5/5.6, where AI suddenly outperformed him in finding open questions and generating conference-level papers.

  4. 18:40 – 22:10

    Lean Formalization and Verification

    Dan Boneh explains Lean as a programming language for logic that compiles proofs, ensuring logical consistency. He notes that while formalizing math was previously prohibitive, AI can now generate Lean proofs for theorems, allowing for machine-checked verification of AI-assisted papers.

  5. 22:10 – 24:45

    AI Output Quality and the Need for Human Reverse Engineering

    Scott Duke Kominers observes that AI models struggle to write high-quality human-readable papers, often skipping difficult steps while over-explaining obvious ones. Consequently, mathematicians must spend effort reverse-engineering the logic behind AI-generated results, even when a Lean proof confirms correctness.

  6. 24:45 – 27:00

    AI writing assistance vs. human understanding

    Speakers discuss whether AI can improve mathematical writing through reinforcement learning, noting that current AI suggestions for editing are often poor. They contrast AI's ability to generate formal proofs with its inability to provide human-understandable explanations due to a lack of 'theory of mind'.

  7. 27:00 – 31:45

    The gap between discovery and understanding

    Ittai Abraham predicts a short-term gap where AI excels at discovery but humans must still drive understanding. He suggests math may shift towards an 'astrophysics' model, where humans observe and interpret results generated by AI, similar to how physicists study galaxies.

  8. 31:45 – 41:00

    The billion-page proof and P vs NP

    Dan Boneh expresses concern that AI might solve P vs NP with a proof too long for humans to comprehend. Scott Duke Kominers counters that math has always involved 'telescopes' and structures we don't fully understand, suggesting long proofs provide new data points for mathematicians to analyze.

  9. 41:00 – 47:29

    Human impact and career displacement

    The conversation shifts to the psychological impact on mathematicians, citing a Substack post by Kerwin Hampshire about a 'spiritual crisis.' Speakers discuss how AI solves problems researchers planned to spend lifetimes on, forcing a reevaluation of identity and aspirations in the field.

  10. 47:29 – 50:02

    The Shift from Problem-Solving to Research Taste

    Speaker_1 argues that as AI handles technical execution, the value of human researchers shifts to developing 'taste' for identifying interesting problems. Faculty evaluation criteria must adapt to distinguish between candidates who prompt AI for publications and those who solve problems independently.

  11. 50:02 – 53:02

    Urgency in Applied Research and Persuasion

    Speaker_0 notes that while AI struggles with some theoretical questions, applied areas like post-quantum cryptography remain urgent and challenging. He emphasizes that convincing the community of new philosophies requires years of gradual progress and effective communication, which remains a human-centric skill.

  12. 53:02 – 56:04

    Acknowledging Loss and the Discovery Frontier

    Speaker_3 describes the current era as the biggest change in mathematics history, advising those who dislike change to avoid theoretical research. He suggests focusing on the 'human understanding frontier'—explaining results to elementary students—rather than just pushing the discovery frontier.

  13. 56:04 – 1:01:12

    The Social Joy of Discovery and Community

    Speaker_1 and Speaker_2 discuss the social nature of mathematics, where sharing ideas and building on each other's work creates community. Speaker_1 shares personal anecdotes of solving long-standing problems and the excitement generated by recent OpenAI math papers, highlighting that discovery joy can persist even with AI assistance.

  14. 1:01:12 – 1:30:16

    Compression, Elegance, and Democratization

    Speaker_3 interprets the joy of discovery as stemming from 'compression'—the satisfaction of succinctly understanding complex ideas. Speaker_1 argues that AI may democratize discovery, allowing more people to experience the thrill of finding new knowledge, rather than eliminating it.

Claims worth verifying

AI can solve the Millennium problems and provide a Lean proof but cannot provide a human-understandable explanation.

Unidentified speaker (UNKNOWN) ·25:52

How to verify: Check the official announcement and published proofs for the Millennium Prize Problems (e.g., Poincaré Conjecture, Navier-Stokes) for any recent AI-generated formalizations in Lean or Coq, and review the accompanying human-readable expositions. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Mathematicians have formal barriers proving P does not equal NP, such as relativization and natural proofs.

Unidentified speaker (UNKNOWN) ·34:24

How to verify: Consult academic literature on complexity theory, specifically papers by Razborov and Rudich on 'Natural Proofs' and Baker-Gill-Solovay on 'Relativization'. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

A young mathematician named Kerwin Hampshire wrote a Substack post titled 'The Dark Night of Mathematics' describing a spiritual crisis due to AI.

Unidentified speaker (UNKNOWN) ·41:27

How to verify: Search for 'Kerwin Hampshire' and 'The Dark Night of Mathematics' on Substack or Google Scholar to locate the specific essay and author. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

AI models are currently limited to the convex hull of human-written knowledge and may not break conceptual barriers.

Unidentified speaker (speaker_1) ·40:23

How to verify: Review research on AI interpretability and generalization limits, such as studies on whether LLMs can discover new mathematical theorems outside their training distribution. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Ittai Abraham told graduate students in January 2026 to be excited about using generative AI for substantive research tasks.

Unidentified speaker (UNKNOWN) ·44:22

How to verify: Check Ittai Abraham's public communications, blog posts, or university announcements from January 2026 regarding AI in education. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Faculty candidates may submit hundreds of publications that were entirely prompted by AI models rather than written or invented by the candidate.

Unidentified speaker (UNKNOWN) ·47:51

How to verify: Review faculty CVs and publication acknowledgments for AI usage disclosures; compare publication volume against known AI prompting capabilities. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

The speaker has worked on a specific math problem since 2004 and another since 2007, solving the latter recently using AI.

Unidentified speaker (speaker_1) ·58:14

How to verify: Check the speaker's (Dan Boneh) publication history and arXiv preprints for papers published after 2026 that cite AI assistance or solve long-standing problems in cryptography. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

OpenAI released over 400 papers on various math problems.

Unidentified speaker (speaker_1) ·1:06:21

How to verify: Search OpenAI's official blog or arXiv for papers authored by OpenAI researchers or attributed to their models in the relevant timeframe. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

A mathematician described OpenAI's math results as 'another trophy in OpenAI's bag, one of hundreds presented like severed rhino heads.'

Unidentified speaker (speaker_2) ·1:06:38

How to verify: Locate the referenced article in The Atlantic by Steven Strogatz discussing OpenAI's math contributions and verify the quote. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Recent work reduced the proof regarding Landau's zeros to a three-page argument based on a determinant construction.

Unidentified speaker (speaker_1) ·59:15

How to verify: Search arXiv for recent papers on 'Landau zeros' or 'Landau-Siegel zeros' that discuss reductions to short proofs or determinant constructions. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

The speaker solved a lemma after three months without notes, experiencing the same satisfaction as the first time.

Unidentified speaker (speaker_3) ·1:01:59

How to verify: This is a personal anecdote; verification is limited to the transcript itself unless the speaker (Scott Kominers) has published or spoken about this specific experience elsewhere. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Speaker_2 states that AI models have only recently reached the level where educators must rethink what to teach, a change occurring in the last couple of months.

Unidentified speaker (speaker_2) ·1:15:30

How to verify: Cross-reference with the podcast release date of 2026-11-18 and the verification window 2026-08-18 through 2026-11-18 to confirm the timeline of model capability shifts discussed. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Speaker_1 stated in January that graduate students would be the first generation to use LLMs and generative AI for substantive, substitutive research tasks.

Unidentified speaker (speaker_1) ·1:17:19

How to verify: Search for public statements, blog posts, or course syllabi from Speaker_1 dated January 2026 containing this prediction. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Speaker_1 claims that within seven to eight months of their January statement, students could write end-to-end papers using AI.

Unidentified speaker (speaker_1) ·1:17:38

How to verify: Verify the timeline against the podcast release date (November 2026) and look for student submissions or academic case studies from mid-2026 demonstrating AI-generated end-to-end papers. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Speaker_2 asserts that AI is not yet good enough to solve critical societal problems like the post-quantum cryptography transition.

Unidentified speaker (UNKNOWN) ·1:21:25

How to verify: Review current cryptographic standards (e.g., NIST PQC standards) and AI research benchmarks to assess if AI has independently solved or finalized these complex cryptographic transitions. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Speaker_2 quotes Andrew Wiles stating that the definition of a good mathematical problem is the mathematics it generates rather than the problem itself.

Unidentified speaker (speaker_2) ·1:29:15

How to verify: Search Andrew Wiles' published interviews, lectures, or writings for this specific quote regarding the definition of a good mathematical problem. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Speaker_1 mentions receiving outreach from high school and middle school students globally who are doing serious mathematics using AI tooling.

Unidentified speaker (UNKNOWN) ·1:28:23

How to verify: Check Speaker_1's public correspondence logs, social media, or university office hours records for evidence of such outreach during the 2026 verification window. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Do your own research

  • Verify the specific capabilities and release dates of GPT-5/5.6 mentioned by Dan Boneh against official OpenAI documentation.
  • Review the Lean theorem prover documentation and examples of AI-generated proofs to assess current verification reliability.
  • Check recent publications from OpenAI and other AI labs regarding their mathematical proof capabilities, specifically the Landau zeros proof reduction.
  • Assess the current state of post-quantum cryptography standards (e.g., NIST selections) to evaluate the urgency and solvability of these problems by AI.
  • Read Kerwin Hampshire's Substack post on the 'spiritual crisis' in mathematics to understand the psychological impact on researchers.
  • Review recent curriculum changes in top-tier mathematics and CS PhD programs to see how they are adapting to AI capabilities.
  • Investigate the definition and evaluation criteria for 'taste' in mathematical research as discussed by Scott Duke Kominers.
  • Look for counter-arguments regarding the 'convex hull' limitation of AI models in generating fundamentally new ideas outside training distributions.
Full transcriptexpand
0:00
Unidentified speaker (speaker_0): The whole academic system is actually collapsing because the rate of papers has like double, tripled.
0:05
Unidentified speaker (UNKNOWN): On
0:05
Unidentified speaker (speaker_1): the one hand, you know, it's putting pressure on a lot of existing institutions for circulating and consolidating knowledge. On the other hand, certain branches of academia are more productive than they've been in decades. We're
0:16
Unidentified speaker (UNKNOWN): gonna live in a world with a math truth oracle. You can
0:18
Unidentified speaker (speaker_1): ask it any question you
0:20
Unidentified speaker (UNKNOWN): want. It'll tell you here's
0:21
Unidentified speaker (speaker_1): the
0:21
Unidentified speaker (UNKNOWN): answer.
0:22
Unidentified speaker (speaker_0): Whether a computer can solve a problem X that I'm working on is not gonna take the joy. that I have when I work on problem X. There
0:29
Unidentified speaker (speaker_2): will always
0:30
Unidentified speaker (UNKNOWN): be judgment
0:31
Unidentified speaker (speaker_2): calls and
0:31
Unidentified speaker (UNKNOWN): value
0:32
Unidentified speaker (speaker_0): judgments
0:33
Unidentified speaker (speaker_2): that
0:33
Unidentified speaker (UNKNOWN): I
0:33
Unidentified speaker (speaker_2): expect
0:34
Unidentified speaker (speaker_0): humans will want
0:35
Unidentified speaker (UNKNOWN): to
0:35
Unidentified speaker (speaker_2): make themselves.
0:36
Unidentified speaker (speaker_1): We will need a lot more mathematicians in the future, and with any luck, the opportunities to explore will energize some people to do so.
0:43
Unidentified speaker (speaker_2): How are you adapting what you want to instill in the next generation of mathematicians?
0:53
Unidentified speaker (speaker_2): When a lot of people think about math discovery, um, I think maybe an image that comes to mind, you know, aside from maybe the goodwill hunting kind of Hollywood version of it, is like the story of Andrew Wiles, who solves uh he comes up with a proof for Fermat's last theorem. He works in isolation. Um, he's this sort of lone genius, and he has this brilliant discovery
1:12
Unidentified speaker (speaker_2): breakthrough, solves a really open long-standing problem that's been there for centuries.
1:17
Unidentified speaker (UNKNOWN): And
1:17
Unidentified speaker (speaker_2): it's kind of a story of human ingenuity and passion and triumph. And I wonder at this moment, when AI is mowing down problem after problem, is that story no longer one that can be told? Like, is that is the is that era over? Is that done now?
1:38
Unidentified speaker (speaker_2): The kind of interesting
1:38
Unidentified speaker (UNKNOWN): thing.
1:38
Unidentified speaker (speaker_2): Andrew Wiles era of math?
1:41
Unidentified speaker (UNKNOWN): I
1:41
Unidentified speaker (speaker_3): think there's a little bit of a sort of important
1:45
Unidentified speaker (speaker_3): wrapper to put on your version of the Andrew Wiles story. You know, the Wiles proof of Fermat's last theorem, which then
1:53
Unidentified speaker (speaker_3): had a subtle flaw that was later repaired in work with Taylor, itself built on many, many, many layers of mathematics that people have. found over large long periods of time. And so there's this combination of the soul, like Wiles is a absolute genius, right? There's this sort of combination of the image of the human passion, like, you know, sort of
2:13
Unidentified speaker (speaker_3): uh there's a there's even a musical sort of mocked up around the idea called Fermat's last tango, and has this line, you know, seven years he spent shut up in his attic. Something like curves and complex functions was, you know, uh, you
2:26
Unidentified speaker (UNKNOWN): know,
2:26
Unidentified speaker (speaker_3): sort of all the time. I don't remember the exact lyrics, but whatever.
2:29
Unidentified speaker (UNKNOWN): Did you
2:29
Unidentified speaker (speaker_2): understudy Scott?
2:31
Unidentified speaker (speaker_3): Oh gosh. If only, if only.
2:45
Unidentified speaker (speaker_3): Of edifices of knowledge that
2:49
Unidentified speaker (speaker_3): no individual contribution sort of
2:53
Unidentified speaker (speaker_3): these days maybe ever sort of happens without all of this sort of superstructure of math around it. Aaron
2:59
Unidentified speaker (UNKNOWN): Powell
2:59
Unidentified speaker (speaker_0): What's hard to predict is where things are headed. And I think Robert, your question is about where things are headed. Yeah? So if the models continue to improve at the rate that they have improved over the last three or four years, then yeah, probably you know, I I guess my the way to think about this is with the analogy to chess, right? That I kind of like
3:19
Unidentified speaker (speaker_0): like to tell the story that, you know, back, what is it, 1997,
3:23
Unidentified speaker (UNKNOWN): the
3:23
Unidentified speaker (speaker_0): first time that was the first time a machine beat a grandmaster, actually the world champion, Kasparov at at uh at chess.
3:29
Unidentified speaker (UNKNOWN): Um
3:30
Unidentified speaker (speaker_0): and then over time it became very clear that there's like no point for a human to to play against a computer in chess chess 'cause because the human because the human will never will never win. Um and so we're not at that point yet. But if they continue to accel accelerate the way they have been accelerating, we might get to a point where there's no point in a human um kind of
3:50
Unidentified speaker (speaker_0): working on a hard problem 'cause a computer can do it much, much faster. And I should caveat that there are other reasons to f work on a problem than to just get a solution.
3:59
Unidentified speaker (UNKNOWN): Yeah. Yeah, I mean do you
4:01
Unidentified speaker (speaker_2): do you believe that, that there could be a that like we might reach a point where there is there is no point kind of as you framed it for a human to be
4:08
Unidentified speaker (UNKNOWN): working
4:08
Unidentified speaker (speaker_2): on this stuff.
4:09
Unidentified speaker (speaker_0): So maybe I'll say that and then I want to hear Scott too in that and and and and Justin and Itali. But uh
4:15
Unidentified speaker (UNKNOWN): uh
4:15
Unidentified speaker (speaker_0): you know part of the to to be honest, uh the reason uh for example I like to work on questions is not so much because
4:22
Unidentified speaker (UNKNOWN): o
4:22
Unidentified speaker (speaker_0): obviously I want to solve them, but the reason to be honest I like to work on them is because it enhances my understanding of the of the problem, the techniques that are involved.
4:30
Unidentified speaker (UNKNOWN): So
4:30
Unidentified speaker (speaker_0): I I d I don't think that's gonna go away. So whether a computer can solve a problem X that I'm working on is not gonna take the joy that I have when I work on problem X. You know, it doesn't matter to me that a computer can solve it. I still enjoy working on the problem, getting an understanding, uh, you know, enhancing my knowledge of the space just by thinking about the problem.
4:49
Unidentified speaker (UNKNOWN): And
4:49
Unidentified speaker (speaker_0): I have to say also we are part of part of the
4:53
Unidentified speaker (UNKNOWN): a
4:53
Unidentified speaker (speaker_0): big chunk of our of our work actually is educating the public, educating our students. And um
5:00
Unidentified speaker (UNKNOWN): I
5:01
Unidentified speaker (speaker_0): you know I firmly believe you can't be a good educator unless you're kind of at the cutting edge. And the way to get to the cutting edge is is by thinking about problems yourself and figuring out things for yourself.
5:11
Unidentified speaker (UNKNOWN): Um
5:12
Unidentified speaker (speaker_0): so to me it doesn't really even if we kind of kinda so again I think today we're at a point where they're just a tool. They help us um accelerate what we've done, what what we're what we're working on. If we are if we do extrapolate and it is true that they are gonna have super extra superhuman abilities, that they can solve any problem better than us, it doesn't matter. I
5:32
Unidentified speaker (speaker_0): will still enjoy working on problems. I will would still enjoy figuring out things for myself and just gaining an unders under better understanding by doing that. Yeah.
5:39
Unidentified speaker (speaker_3): I I agree with what you're saying. I I think even now they're more than a tool. So at least the way I think about this, uh, there are kind of two frontiers. Uh one frontier, I'll call it the discovery frontier. This is where kind of this state-of-the-art new results are being produced. And then the second one is the human understanding frontier. This is kind of where, you know, as you
5:59
Unidentified speaker (speaker_3): you mentioned, human understanding is kind of the your motivation. So I think this is kind of the, you know, one of the core ideas in mathematics is that the goal is to is to get to human understanding. Well, this second frontier, by the way, this is uh related to uh Bill Thurston's uh 1994 uh essay on on you know proof on
6:19
Unidentified speaker (speaker_3): on the progress of math, um, where where kind of he you know really put human understanding as a as an important goal of of math. And I think there was kind of a different line of thought in math, which said that only the only thing important is to push this discover this discovery frontier. I think what we're seeing now, at least in the last half year, is a separation between these two frontiers.
6:39
Unidentified speaker (speaker_3): So we're seeing, for example, you know, AI solve uh millennium problems or you know, have a lean proof of them. And mathematicians are now struggling to finally get to a human understanding. of that. So we're seeing kind of, you know, a separation between these two frontiers. And I think it's going to be, you know, fascinating in the future. Maybe,
6:59
Unidentified speaker (speaker_3): maybe we'll talk about that later, but you know, how will that how will these two frontiers move away from each other or stay together?
7:05
Unidentified speaker (speaker_1): Or stay together. Right. I mean, like I sort of,
7:08
Unidentified speaker (UNKNOWN): you know,
7:09
Unidentified speaker (speaker_1): I think it's an open question. There's there are many different branches of the open question here, right? Like pun not intended. I
7:14
Unidentified speaker (UNKNOWN): guess
7:14
Unidentified speaker (speaker_1): we're talking about open questions in mathematics. There's an open question about the open questions in mathematics. There's a question of whether, as Dan said, the models will sort of continue to accelerate, right? Whether it's sort of their ability to solve math problems on their own or or you know in in complementary
7:30
Unidentified speaker (speaker_1): infrastructure with humans improves at the rate it's been improving over the last, you know, year or so.
7:37
Unidentified speaker (UNKNOWN): And then
7:37
Unidentified speaker (speaker_1): there's a separate question is. you know, sort of conditional on your different hypotheses about what that might look like, what that sort of changes about the product the production of math. And that employ relies on some additional hypothesis about how the
7:51
Unidentified speaker (speaker_1): topology of math looks, right? Like, you know, uh you could imagine that, you know, five years from now we're, you know, back in, you know, sort of we're back in our attics, right? It's sort of it's it's you know sort of a replica of the Andrew Wiles story, except that we have sort of a you know a a very fancy like you know sort of proving assistant and an extremely fancy like you know sort of uh you know
8:11
Unidentified speaker (speaker_1): AI system helping us push a frontier that is you know just as hard as what the old frontiers felt like just with the new tools, right? It's like we might sort of you know, speed, speed, speed, speed, speed, and then you know you actually get you know concavity, you get sort of you know, math gets hard again, and now it's hard up to the tool set we have, you know, the
8:31
Unidentified speaker (speaker_1): new tool set, which again, this is lots of computational revolutions in the history of math, especially in the last half century, have had a version of that. You sort of you're a bunch of new things you can do with a computational tool that changes where the frontier of math is. And then you know, sometimes you find objects you don't quite understand, and then eventually humans like catch. up
8:51
Unidentified speaker (speaker_1): to sort of, you know, the tool slows down, humans catch up, you get to a point where you're now again on the margin, sort of all at the frontier together. There's another universe in which the models sort of stay ahead in solve a discovery space, and he ties framework. And you know, humans, you know, and and human understanding as a
9:11
Unidentified speaker (speaker_1): project of math becomes more and more a project of of understanding sort of like, you know, sort of the the things that are happening in this in this discovery space, um, rather than discovery and understanding sort of moving together as a single object. I don't know. Like I tend to believe that sort of
9:29
Unidentified speaker (UNKNOWN): to
9:29
Unidentified speaker (speaker_1): the extent that understanding math as a human project, that will always be the case. right? It's sort of but even there, you know, AI models help a lot. these days too, right? You know, sort of certainly when I'm trying to understand a new mathematical object, uh, you know, I spend some time reading textbooks and I spend some time, you know, sort of, you know,
9:49
Unidentified speaker (speaker_1): using some sort of AI interpretation sort of architecture.
9:54
Unidentified speaker (speaker_3): So so I'll I'll give you my my short-term uh first first this uh uh go kind of hypothesis, which is in the short term, I think what we're seeing is that uh
10:03
Unidentified speaker (speaker_3): uh you know, large language models with all this, you know, reinforcement learning are making them really, really good at solving math problems. Um to the extent that they're basically at the frontier of
10:15
Unidentified speaker (UNKNOWN): the thing.
10:17
Unidentified speaker (speaker_1): I don't think I would say
10:18
Unidentified speaker (UNKNOWN): to
10:18
Unidentified speaker (speaker_1): that generality.
10:21
Unidentified speaker (speaker_2): Let's get into that after after
10:22
Unidentified speaker (UNKNOWN): you've got that. a lot of things. Yeah but let's
10:24
Unidentified speaker (speaker_3): actually I'd I'd love to I'd love to get into that, Scott. Um but But I would say they're definitely very good at solving math. Or, you know, they they at least surprise
10:32
Unidentified speaker (UNKNOWN): me
10:32
Unidentified speaker (speaker_3): to how good they are in solving math problems. And it seems
10:36
Unidentified speaker (UNKNOWN): like,
10:37
Unidentified speaker (speaker_3): you know, we
10:37
Unidentified speaker (UNKNOWN): can force it.
10:39
Unidentified speaker (speaker_2): What was the moment? What was what was the result that made you have your eyes opened to, you know, your you're you're mind blown that this stuff really truly works? And is more than just kind of like a novelty or something like that. I mean I think it was last year, last summer that um the AI models had reached
10:58
Unidentified speaker (speaker_2): Olympia gold medal level. And then since then, like you said, they've solved apparently solved Millennium Prize challenges. So
11:07
Unidentified speaker (UNKNOWN): yeah,
11:07
Unidentified speaker (speaker_2): what was the moment for you that kind of like opened your eyes to this?
11:12
Unidentified speaker (speaker_3): About 10 years ago, I I spent you know a few years of my life trying to solve some sort of problem around kind of metric space embedding. And we made good progress on some of these. And then, you know, we got stuck on others. And I I spent maybe two years trying to solve some of these problems. And then I think this summer, uh, you know, I just tried one of them one of the
11:31
Unidentified speaker (speaker_3): models, and it in basically one hour solved, made, made, you know, very non-trivial improvements. improvement in one of these problems. Um and then obviously when when OpenAI kind of published these 300 or 400 results, then you know two of those papers were actually kind of solving major improvements in in these uh metric space embedding results.
11:51
Unidentified speaker (speaker_3): So I guess you know, uh problems that my colleagues and I tried to work on for maybe a year or two uh were kind of solved by an effort of, again, OpenAI OpenAI is claiming less than three hours. on average per per problem. So that to me was kind of a when it hit personal personally on a problem that I worked on.
12:11
Unidentified speaker (UNKNOWN): I'd
12:11
Unidentified speaker (speaker_2): love to go round Robin and hear everybody's uh you know story about this.
12:15
Unidentified speaker (speaker_0): For I mean for me this summer, I think it was GPT five six. I I lose track of the numbers, uh five si soul or something. Uh but uh this summer, uh twenty 2026, twenty six, I don't know, d June or something.
12:27
Unidentified speaker (speaker_0): Um, you know, I was like not getting much out of five five. You know, I was I that that was just good enough for me to start messing around with it. Um and then and then five six, um it w it was like a phase transition where suddenly for lots of things it was it was better at math than I was. Um I was actually doing this thing where I I was like, look, you know,
12:47
Unidentified speaker (speaker_0): go go through the literature.
12:50
Unidentified speaker (UNKNOWN): Uh
12:50
Unidentified speaker (speaker_0): look
12:50
Unidentified speaker (UNKNOWN): at the literature.
12:50
Unidentified speaker (speaker_0): looking for open questions uh on this topic I'm an expert in, but like, you know, look at adjacent things I haven't looked at myself before. You know, make make sure, you know, look for some signal that this is like a, you know, some people have thought about this, you know, it's an open question in a paper at a serious conference, and just try to solve it. And, you know, I was achieving no success,
13:10
Unidentified speaker (speaker_0): you know, with the earlier model, and then boom, this the summer in June or whatever. you know, I I I I was telling people like, oh, if I wanted to, I could just have like 10, you know, uh I was saying cult papers that's a conference on learning theory. It's like a solid, you know, not the best, but s solid TC uh theoretical computer science conference. And so that
13:31
Unidentified speaker (speaker_0): then it was like, okay, clearly this is now solving problems that some people have thought about uh, you know, a little bit and and actually think are important. right? Um and so that that was suddenly to me like when it can sort of, you know, if you ask it the right question, it'll just produce cult papers for you. Like that, you know, that that's what like a very good grad student uh
13:50
Unidentified speaker (speaker_0): wants to do in their PhD back in in pre-I AI days. Um and so then, you know, since then I've been hammering it on my you actually have to make a decision like what do you want to do with your time and effort now? What do you actually want to publish? Um you know, no longer are you going to put out every publishable result you get if you were doing that before because, you know, uh y you
14:10
Unidentified speaker (speaker_0): gotta make it count. Um
14:12
Unidentified speaker (speaker_2): Scott knows this very well, uh with his uh his restriction on archive now with all the papers he's been generating. Right,
14:18
Unidentified speaker (speaker_1): exactly.
14:19
Unidentified speaker (UNKNOWN): I'm not
14:19
Unidentified speaker (speaker_3): even allowed to put out all the papers that I,
14:21
Unidentified speaker (speaker_1): you know, is wanna want to put out based basically on my
14:22
Unidentified speaker (UNKNOWN): like own. the other thing. The ones that you actually compile are
14:24
Unidentified speaker (speaker_0): worth your time, you're still not allowed to put them out, right? Um and yeah, I mean I d you
14:28
Unidentified speaker (UNKNOWN): know. And just for
14:29
Unidentified speaker (speaker_2): just for listeners, yeah. Scott has actually been restricted, right? Is that the right word on on our
14:34
Unidentified speaker (UNKNOWN): I I was I was I was
14:35
Unidentified speaker (speaker_1): restricted. I was rate limited on the archive before they rate limited everybody on the archive. And their only explanation to be for you know they and I and I tried to get information, you know, sort of did you did I do anything wrong? Did you find any issues with my like papers? And the answer was no, you just have too many of them.
14:50
Unidentified speaker (UNKNOWN): It's too good,
14:50
Unidentified speaker (speaker_2): too prolific.
14:51
Unidentified speaker (speaker_1): You know, it's uh so
14:54
Unidentified speaker (speaker_1): um and now they've rate limited everyone, by the way, because there is such an influx of papers that their moderation system, at least as they've implemented it, it's a whole separate category, you know, category of question whether that's the optimal market design, but their their moderation system as they've implemented it can't r like can't hold up to
15:14
Unidentified speaker (speaker_1): the massive number of new papers, especially by new authors. Trevor
15:18
Unidentified speaker (UNKNOWN): Burrus
15:18
Unidentified speaker (speaker_0): It's probably we're saying that the whole academic system is actually collapsing because the rate of papers has like double, tripled, quadruple
15:26
Unidentified speaker (UNKNOWN): wealth.
15:27
Unidentified speaker (speaker_0): Trevor
15:27
Unidentified speaker (UNKNOWN): Burrus, Jr.:
15:27
Unidentified speaker (speaker_1): Collapsing is sort of a it's a funny statement, right? Like on the one hand, you know, it's putting pressure on a lot of existing institutions for, you know, you know, um
15:37
Unidentified speaker (speaker_1): for circulating and consulting. consolidating knowledge. On the other hand, you know, sort of academia is more or certain branches of academia are more productive than they've been in decades, right? Like it's sort of way more.
15:48
Unidentified speaker (speaker_1): Like
15:49
Unidentified speaker (UNKNOWN): even by orders
15:49
Unidentified speaker (speaker_1): of magnitude, more productive. Trevor
15:50
Unidentified speaker (UNKNOWN): Burrus, Jr.:
15:51
Unidentified speaker (speaker_0): Like I see,
15:51
Unidentified speaker (UNKNOWN): I see the
15:52
Unidentified speaker (speaker_0): my my own group. I mean, the rate, the rate in which uh papers are generated is like dramatically higher than it was before.
15:58
Unidentified speaker (UNKNOWN): I'm not
15:59
Unidentified speaker (speaker_1): sure I would call it a collapse. I would call it like straining the existing institutions, but because of opportunity. not because something's, you know, sort of fundamentally going wrong.
16:05
Unidentified speaker (speaker_2): Just to the brink of collapse, but not quite
16:07
Unidentified speaker (UNKNOWN): to
16:07
Unidentified speaker (speaker_2): the to the point of collection
16:08
Unidentified speaker (UNKNOWN): of
16:08
Unidentified speaker (speaker_2): the
16:08
Unidentified speaker (UNKNOWN): same thing.
16:09
Unidentified speaker (speaker_0): This is a whole other, you know, conversation. I don't know how in scope it is for this discussion here about like how the community, you know, deals with the higher rate of of you know discoveries and productivity coming out. Um but I let me make some points where With this is all very controversial. So uh not not everyone's gonna agree with what I say. But you know, uh
16:29
Unidentified speaker (speaker_0): like our a lot of the conferences I'm talking about, uh you know, cult, stock fox, the theoretical computer science conferences, they've seen like a doubling of submissions or so, which is like a lot. But here's the flip side. Um one, uh you can And um like I'm a big fan of uh like lean
16:49
Unidentified speaker (speaker_0): formalization of of math. um, which means you have like a machine check proof of of the theorems. And you know, of course there are there are limits to this and you know, a human has to make sure the theorem that's been proved matches the actual like English language claim in the paper and all this stuff. But you can actually get a lot more confidence that these you know AI assisted papers are
17:09
Unidentified speaker (speaker_0): correct than we had in in pre pre-I I AI days, number one. And so like one you know goal of review, which it never really succeeded fully in doing, was to like make sure things are right and not wrong.
17:20
Unidentified speaker (UNKNOWN): That was a
17:20
Unidentified speaker (speaker_1): huge bottleneck in the reviewing process. One of the things that made it the slowest was the verification of a mathematical proof.
17:26
Unidentified speaker (speaker_0): Yes. So that that is one way things can actually be better. And then the other thing is
17:33
Unidentified speaker (UNKNOWN): Can I
17:33
Unidentified speaker (speaker_2): just ask uh also um we hear a lot about this lean formalization
17:37
Unidentified speaker (UNKNOWN): now.
17:38
Unidentified speaker (speaker_2): w was it always the case that people would supply a lean formalization of their
17:43
Unidentified speaker (speaker_2): new thing
17:44
Unidentified speaker (UNKNOWN): given
17:44
Unidentified speaker (speaker_2): AI output in order to say, look, I went the extra mile to show that this actually all nets out in a way that's like logically consistent.
17:52
Unidentified speaker (speaker_3): I think this is part of why uh a AI is so successful in in in math, because there's a way for it to evaluate whether it it's correct or not. So by having a formal proof. that allows it to do this reinforcement learning in a very efficient manner. And so in a way, this is kind of why
18:11
Unidentified speaker (speaker_3): these models were super successful in making very long
18:16
Unidentified speaker (speaker_3): lines of thought. Because they
18:18
Unidentified speaker (UNKNOWN): can very furious success.
18:20
Unidentified speaker (speaker_1): I think this is part of it. But like a lot of the success in chains of thought come in models working not. just in lean space, right? Like you know, sort of using lean to convexify across math, I think, is a big part of how some of these models have made progress, but I don't think it's the the the full story. But anyhow, we should we've maybe talked that about that in a second. Let's answer
18:40
Unidentified speaker (speaker_1): Robert's question about lean.
18:42
Unidentified speaker (speaker_0): So so get getting lean like lean formalizations of uh you know math theorems was a like a painstaking you know, like top mathematicians would spend you know months and months trying to get like a lean proof. Uh and by the way, we've only recently converted largely converged on lean. There are other like mechanisms,
19:01
Unidentified speaker (UNKNOWN): other sort of
19:02
Unidentified speaker (speaker_1): different you know, sort
19:03
Unidentified speaker (UNKNOWN): of
19:03
Unidentified speaker (speaker_1): like proof of the
19:04
Unidentified speaker (UNKNOWN): same. But
19:04
Unidentified speaker (speaker_2): lean is kind
19:05
Unidentified speaker (UNKNOWN): of the
19:05
Unidentified speaker (speaker_2): gold standard. It's this programming language based on
19:07
Unidentified speaker (UNKNOWN): like
19:07
Unidentified speaker (speaker_2): what the fundamental axioms of of math. Uh and
19:11
Unidentified speaker (UNKNOWN): so
19:11
Unidentified speaker (speaker_1): I think the the way to think about it is it's like a programming language. that works in logic, and it has a compiler like a programming language. So you write a series of logical statements at sort of some very
19:23
Unidentified speaker (speaker_1): granular level. You sort of reduce like a series of ideas into a chain of logic. And then you compile at the end. And when you compile, like just like a normal compiler will like check for syntax errors or something, here it checks for logical consistency of the chain of statements. And so if your program compiles, you know that the sequence of statements, you know, sort
19:43
Unidentified speaker (speaker_1): of form a complete logical proof. And so you can explicitly verify. You know, if you if you have a mathematical statement at the top and you build a proof that sort of leads to that statement being true, you can explicitly verify, right? If the program compiles, the the sort of the um, you
20:02
Unidentified speaker (UNKNOWN): know,
20:02
Unidentified speaker (speaker_1): it kernel checks.
20:03
Unidentified speaker (speaker_1): then you know that your your sort of proof is is logically complete.
20:08
Unidentified speaker (UNKNOWN): It's
20:08
Unidentified speaker (speaker_1): like you have
20:10
Unidentified speaker (UNKNOWN): your
20:10
Unidentified speaker (speaker_2): checkbook but for uh
20:11
Unidentified speaker (speaker_1): yeah exactly it's that's exactly
20:13
Unidentified speaker (UNKNOWN): right. It's
20:14
Unidentified speaker (speaker_1): balancing your checkbook but for but for math.
20:16
Unidentified speaker (UNKNOWN): Right.
20:16
Unidentified speaker (speaker_0): And I'd say you know the the now it as of this summer, so you know in the last five months or something, uh you know, for like your average theoretical computer science paper, you can just ask the AI uh, you know, to give a lean proof of all the main theorems in the paper, and then you you have to go and carefully check that, you know, uh the theorem
20:37
Unidentified speaker (speaker_0): that pre you know uh it's not pulling the fast one on you.
20:40
Unidentified speaker (UNKNOWN): Um
20:40
Unidentified speaker (speaker_0): and so that's a ga that's uh you know, night and day uh from you know before Four five or months ago. Right
20:46
Unidentified speaker (UNKNOWN): again. I
20:46
Unidentified speaker (speaker_1): mean think about it so think about what the the counterfactual was, right? You're if you're reducing uh a state like a set of statements that's written in words, right? Humans read math papers in words, not in machine code, right? If you're trying to reduce it to its formal logic architecture, much of math doesn't have that format. And moreover, like natively, right? It's it's
21:06
Unidentified speaker (speaker_1): not sort of like written in sort of sequence of formal logical statements. There's all sorts of synecdichies synecdoches that come from you know, sort of using ideas and relationships. and between concepts that are well established. And all the foundations sort of remember we were talking about this edifice of math, right? Like all the foundations of that edifice of math have to also, if you want this thing to be sort of
21:26
Unidentified speaker (speaker_1): completely kernel-checked, it has to sort of, you know, be turtles all the way down until you get to the most basic mathematical foundations. And that was just completely prohibitive for most mathematicians to do. at scale.
21:41
Unidentified speaker (speaker_1): Proving really complicated theorems hard enough, then figuring out how to transform them into this, you know, sort of machine code is a completely different, or
21:51
Unidentified speaker (UNKNOWN): I don't
21:51
Unidentified speaker (speaker_1): know, it's a recursive problem on top of the really hard problem.
21:56
Unidentified speaker (speaker_1): Even as Justin was saying, five months ago, it was pretty hard to do this at really large scale using sort of state-of-the-art AI. It's really been a very recent speed up that this is something you can hope to do. And still there are domains where it doesn't work at scale yet.
22:10
Unidentified speaker (UNKNOWN): Let
22:10
Unidentified speaker (speaker_3): me add to that a little bit in that uh the models seem to be not very good at writing like what we would consider high-quality papers. So when
22:18
Unidentified speaker (UNKNOWN): it generates it
22:19
Unidentified speaker (speaker_3): generates a fascinating result, you know, we're all excited about the result, but when you try to read the paper, it turns out there are like obvious steps where it would spend multiple ch pages trying to explain why this obvious step is correct. And then there are very difficult steps where it would just write like, you know, three lines saying, Oh, you know, this follows from that and you have no idea why that's true. And now you have to spend a
22:39
Unidentified speaker (speaker_3): lot of effort to try and clarify that. So the model is not quite writing like the way we would expect it to write. So the only reason why we would spend the effort to go and understand uh what it what it wrote is because we have to have some evidence that it's correct. And so when the labs, in addition to the written paper, the labs produce a lean proof that uh is machine
22:59
Unidentified speaker (speaker_3): checkable and so we know the result is correct, then it's actually worth the effort to go and kind of reverse engineer what the paper actually uh tries to say. Now so right.
23:09
Unidentified speaker (speaker_2): Very funny about this to just to bring it back to the opening question I had about uh I guess I was saying his name wrong, but Fermaz Fermat's uh last theorem. Wasn't the whole point that he had written in the margin of some book saying that he had a proof or whatever and like he didn't have time to explain it and then
23:23
Unidentified speaker (speaker_2): it was all this
23:24
Unidentified speaker (speaker_2): reverse engineering of what we thought maybe he had found.
23:28
Unidentified speaker (speaker_3): Yeah, so that was not uh that okay, that that uh that would be a model uh paper to an extreme where it just writes I have a wonderful proof of this theorem, you go figure it out. Uh please. At least the models don't do that to us. But imagine imagine Fermat had a lean proof for his uh for his statement. Then he could write in the margin, you know, I have a lean
23:48
Unidentified speaker (speaker_3): proof. Um then that would well I'm
23:50
Unidentified speaker (UNKNOWN): not sure.
23:54
Unidentified speaker (speaker_3): Oh here's a snark, here's a ZK proof that uh the my lean proof is correct. That would kind of give us a lot more confidence that what he was saying is true. Um but regardless, what what what's happening with these models today is that uh we kind of go and spend the effort to write them better, to write the papers better um once we know the result is correct. Now, you know, my friends at the labs
24:14
Unidentified speaker (speaker_3): tell me that uh right now their the models are not writing well, but this is just a matter of time. You know, next next year, probably the models will be generating the results and also we'll be writing them well. So it will be just a matter of us just reading the papers. and just absorbing them. I don't know how much
24:30
Unidentified speaker (UNKNOWN): people
24:30
Unidentified speaker (speaker_3): have to
24:30
Unidentified speaker (UNKNOWN): do.
24:30
Unidentified speaker (speaker_1): It's maybe not so surprising that if there's like something the models are bad at introspecting about, like how a human would process the text of a proof written for humans might be a a really hard frontier at some level.
24:41
Unidentified speaker (UNKNOWN): Um so
24:43
Unidentified speaker (speaker_1): I don't know. Yeah. Yeah.
24:44
Unidentified speaker (UNKNOWN): Now do you
24:45
Unidentified speaker (speaker_3): think that that can that the paper writing, the math paper writing can be improved uh strictly based on more of this reinforcement learning around like uh around math as a domain in particular?
24:57
Unidentified speaker (speaker_2): Or is
24:57
Unidentified speaker (speaker_3): this something that needs to be improved across other domains, like
25:01
Unidentified speaker (UNKNOWN): you
25:01
Unidentified speaker (speaker_3): know, say humanities, just writing in general? I mean, one thing
25:04
Unidentified speaker (speaker_2): that uh
25:04
Unidentified speaker (UNKNOWN): is
25:05
Unidentified speaker (speaker_3): interesting to me is when I use AI as a tool,
25:08
Unidentified speaker (speaker_2): I
25:08
Unidentified speaker (speaker_3): will often ask it like, hey, how can you
25:10
Unidentified speaker (UNKNOWN): how
25:10
Unidentified speaker (speaker_3): can I improve this piece?
25:12
Unidentified speaker (speaker_2): Like, what would make
25:12
Unidentified speaker (UNKNOWN): this
25:13
Unidentified speaker (speaker_2): post better? And
25:13
Unidentified speaker (speaker_3): it provides
25:15
Unidentified speaker (speaker_2): a
25:15
Unidentified speaker (speaker_3): multitude of su very precise suggestions about change change this,
25:19
Unidentified speaker (speaker_2): delete
25:19
Unidentified speaker (speaker_3): that, move this over
25:20
Unidentified speaker (speaker_2): here.
25:21
Unidentified speaker (speaker_3): And invariably,
25:22
Unidentified speaker (UNKNOWN): they're
25:23
Unidentified speaker (speaker_2): all terrible suggestions. And not taking any of them. Um so
25:27
Unidentified speaker (UNKNOWN): I
25:27
Unidentified speaker (speaker_2): I wonder how much of this is like a writing thing, and maybe the labs just haven't gotten to perfect that yet. Maybe it is perfectible, it'll get a lot better. Um although it's
25:35
Unidentified speaker (UNKNOWN): so
25:35
Unidentified speaker (speaker_2): Robert, this
25:36
Unidentified speaker (UNKNOWN): goes
25:36
Unidentified speaker (speaker_2): back. I I at least the way I think about this to you know, I my previous point about kind of the short-term prediction, which is, you know, it it it's kind of surprising that AI can can solve the millennial problem and and provide a lean proof, but cannot provide a a human
25:50
Unidentified speaker (speaker_1): understandable explanation.
25:52
Unidentified speaker (UNKNOWN): And
25:52
Unidentified speaker (speaker_1): this goes back to kind of you know what Dan talked about, human understanding. And and
25:56
Unidentified speaker (UNKNOWN): to
25:57
Unidentified speaker (speaker_1): me, the the reason is the following math is kind of completely universal and has no relation
26:03
Unidentified speaker (UNKNOWN): to
26:04
Unidentified speaker (speaker_1): humans. It's just a lean proof of derivations of mathematics. While human understanding, right writing something that is human understandable requires you to have a mental
26:14
Unidentified speaker (UNKNOWN): model, um,
26:16
Unidentified speaker (speaker_1): a theory of brain,
26:17
Unidentified speaker (UNKNOWN): uh of how how the listener, what what it under underst what what the listener understands, what the reader understands, um, what are their intuitions, how it learns. And so it is much easier to give uh a reinforcement task where you have a proof and you have to say, is that a a correct proof or an incorrect proof? It
26:36
Unidentified speaker (speaker_1): is much harder
26:37
Unidentified speaker (UNKNOWN): to show a explanation of a let's say a a proof or a lemma and ask, is this a good explanation or is this better than the
26:44
Unidentified speaker (speaker_3): previous explanation?
26:45
Unidentified speaker (UNKNOWN): You
26:45
Unidentified speaker (speaker_3): know, I think developing this theory of mind is actually not a limit only on math. This is kind of a fundamental limit on many other domains.
26:54
Unidentified speaker (speaker_1): And
26:54
Unidentified speaker (speaker_3): so, you know, I I kind of share with you know d Dan's kind of long-term view that that this will be finished, fixed.
27:00
Unidentified speaker (UNKNOWN): But
27:00
Unidentified speaker (speaker_3): I think there's kind of a fundamental difference. And that is why we're kind of seeing this amazing gap where it's actually not great at at this
27:07
Unidentified speaker (UNKNOWN): theory,
27:07
Unidentified speaker (speaker_3): I'll call this theory of mind of explaining
27:10
Unidentified speaker (UNKNOWN): um
27:10
Unidentified speaker (speaker_3): and human understanding, while it is actually amazing uh relatively. at at making math derivations. And and therefore, kind of my again, my short-term prediction is that this might take a while, where we'll have AI be very, very good at pushing the uh discovery frontier, but we'll still have humans do a lot of
27:30
Unidentified speaker (speaker_3): work on kind of pushing the human understanding frontier. Now obviously AI is going to help you that task, um, but but a lot of the work will have to actually be uh um done by by humans.
27:39
Unidentified speaker (UNKNOWN): And
27:40
Unidentified speaker (speaker_3): we're seeing that already, at least I've seen that in many of the recent uh
27:43
Unidentified speaker (UNKNOWN): uh
27:43
Unidentified speaker (speaker_3): theory of computer science works,
27:45
Unidentified speaker (UNKNOWN): right? Where
27:45
Unidentified speaker (speaker_3): where there were uh these amazing advances and then we had human experts that actually gave really good
27:51
Unidentified speaker (UNKNOWN): uh uh
27:52
Unidentified speaker (speaker_3): explanations of these proofs.
27:55
Unidentified speaker (speaker_2): Itai, you've got a great metaphor that you've used to describe uh this frontier as it's unfolding, where uh you've you you've talked about it as sort of like a shift into the astrophysics kind of era of math, I believe, is sort of the way you put it, where these things are sort of like telescopes and we have to kind of interpret all of their uh their data.
28:15
Unidentified speaker (speaker_2): I'm probably butchering it, but I thought maybe you could uh share this uh this year. Yeah. So
28:19
Unidentified speaker (speaker_0): I guess this is my my medium-term prediction. And and there I think there are basically two paths. So one path is that this discovery frontier.
28:27
Unidentified speaker (speaker_0): and the human understanding frontier will kind of move
28:30
Unidentified speaker (UNKNOWN): uh
28:31
Unidentified speaker (speaker_0): step by step. You know, we humans
28:34
Unidentified speaker (UNKNOWN): will develop these really
28:35
Unidentified speaker (speaker_1): great
28:35
Unidentified speaker (UNKNOWN): abstractions that we can explain
28:37
Unidentified speaker (speaker_1): all
28:38
Unidentified speaker (UNKNOWN): these
28:38
Unidentified speaker (speaker_1): very complicated
28:39
Unidentified speaker (UNKNOWN): new
28:39
Unidentified speaker (speaker_1): results that
28:40
Unidentified speaker (UNKNOWN): that the
28:40
Unidentified speaker (speaker_1): discovery frontier will
28:41
Unidentified speaker (UNKNOWN): make. And
28:42
Unidentified speaker (speaker_1): AI will help us do that as well, right? As Dan suggested.
28:45
Unidentified speaker (UNKNOWN): But then there's
28:45
Unidentified speaker (speaker_1): kind of another
28:46
Unidentified speaker (UNKNOWN): uh alternative
28:47
Unidentified speaker (speaker_1): universe
28:48
Unidentified speaker (UNKNOWN): where actually the the, even though we're gonna make
28:50
Unidentified speaker (speaker_1): it
28:51
Unidentified speaker (UNKNOWN): these extreme strides in in in human understanding, uh
28:54
Unidentified speaker (speaker_1): moving that frontier
28:54
Unidentified speaker (UNKNOWN): forward, the frontier of discovery will move even faster than that. And so there will be this huge
29:00
Unidentified speaker (speaker_1): gap,
29:01
Unidentified speaker (UNKNOWN): you know,
29:02
Unidentified speaker (speaker_1): uncountable gap between the discovery frontier of what the It'll definitely be a countable gap.
29:07
Unidentified speaker (speaker_1): Okay, countable gap. Yeah. So there will be a countable gap. Uh I know my canter diagonalization arguments here like
29:15
Unidentified speaker (speaker_1): Thank you.
29:22
Unidentified speaker (speaker_1): The most dramatic statement could be some sort of like uncountable of uncountable of uncountable sort of like you know grams function uncountable gaps. But
29:29
Unidentified speaker (speaker_2): There you go. Anyhow,
29:30
Unidentified speaker (UNKNOWN): sorry.
29:33
Unidentified speaker (UNKNOWN): Second path, there's going to be this, let's say, a large distance between these two frontiers. And when there's such a large distance, I think then the part that the human understanding math becomes much more similar to science because these things already exist. Discovery already happened. So basically now mathematicians are
29:52
Unidentified speaker (UNKNOWN): basically going to uh examine
29:55
Unidentified speaker (speaker_0): the real world, which is what science does, Do and kind of observe uh what you know what has AI already solved.
30:02
Unidentified speaker (UNKNOWN): And so
30:02
Unidentified speaker (speaker_1): the
30:03
Unidentified speaker (speaker_1): the
30:03
Unidentified speaker (speaker_0): analogy here, right, is is kind of you know maybe around uh what what a physicist would do, right? It would study the galaxies through through a telescope,
30:11
Unidentified speaker (UNKNOWN): or
30:12
Unidentified speaker (speaker_0): or what you know what an economist would do. It wouldn't look at what each and every entity is doing, but it would kind of make you know observations and try to learn from those observations. And I just want
30:22
Unidentified speaker (speaker_1): to say that you know, this this idea of um studying something that we don't understand and trying to and striving to make progress in human understanding, I I think that that is something that will you know always continue. Um whether it's for the frontier of math, right, or or you know, even for like understanding how our brain works, right? So so that's like an another great example where
30:41
Unidentified speaker (UNKNOWN): we don't really
30:41
Unidentified speaker (speaker_1): understand how the brain works. You know,
30:43
Unidentified speaker (speaker_3): Itali,
30:43
Unidentified speaker (UNKNOWN): I
30:43
Unidentified speaker (speaker_3): have to I have to push back on this just a little bit in in in that uh you know there's infinitely many truths, right? And we're not gonna have a book. of all possible truths because there are infinitely many of them. And even though the the machines are likely, you know, if we extrapolate ahead, maybe the machines will actually get much further ahead of us in in their in
31:03
Unidentified speaker (speaker_3): their knowledge, there will still be questions that we can ask that they just didn't bother to ask and didn't bother
31:08
Unidentified speaker (UNKNOWN): to
31:08
Unidentified speaker (speaker_3): answer. So then there are two paths, right? You we as a human, as uh us as humans, we can go to the machine and ask the machine to solve the problem we just raised, or we can try to r solve it ourselves. And I claim that we will always kind of enjoy kind of the uh the process of solving for our own understanding, even though the even though we know the machine can do it instant
31:28
Unidentified speaker (speaker_3): instantly and far faster than us.
31:31
Unidentified speaker (speaker_1): Okay, but I'm so so that that's fine. But again, the the way I think about that is that we're gonna live in a world with an with a math math-truth truth oracle.
31:39
Unidentified speaker (UNKNOWN): So
31:39
Unidentified speaker (speaker_1): you know you can ask it any question you want, and it'll it'll it'll tell you, you know, here's
31:44
Unidentified speaker (speaker_2): the answer.
31:44
Unidentified speaker (speaker_3): Okay, so here's what scares me. Here's what scares me. Here's this here's the thing that I'm really quite quite worried about. So like let's say let's pick one of our most important questions. Let's say P versus NP question. We would love to have a solution to a proof that P is not equal to NP. But what happens if the models actually come up with a solution, except the solution is now a billion. pages. Yeah, and
32:04
Unidentified speaker (speaker_3): no human will ever understand ever understand it. So what happens? I don't know what happens to mathematics when we get to a point where the right now we're nowhere near there. This is why I consider these these as tools. They have all the proofs they have produced, we can quite easily understand. But what happens when we move beyond that and the proofs are just too long,
32:23
Unidentified speaker (UNKNOWN): you can see that we're going to be able to do that. Okay, let me let me
32:24
Unidentified speaker (speaker_1): let me make a controversial claim. So
32:26
Unidentified speaker (UNKNOWN): I want
32:27
Unidentified speaker (speaker_1): to claim that like math has always had telescopes. Right. That like, you know, in math, we you know, what what is an open question but a you know sort of look at a pattern or an idea or a structure that we don't entirely understand, but that someone has spotted through some lens. Right.
32:44
Unidentified speaker (UNKNOWN): And
32:44
Unidentified speaker (speaker_1): so, Dan, to your example of the like billion page you know, p not equals np proof, like that feels to me very similar to the idea of, you know, here we have this quadratic form where we think we know exactly which integers it represents and we don't know how to explain why. Right. Like that's,
33:01
Unidentified speaker (UNKNOWN): you
33:01
Unidentified speaker (speaker_1): know, kind of the same, except in one of them you have you could generate sort of you have oracle access to generating new information about it, and the other, you know, at some, you know, some fidelity, and the other you have oracle access to opening a page. You know, I can open page 700,321. 700,321,416. Just made that number up. It's not like a, I don't know, a password or whatever. Um,
33:21
Unidentified speaker (speaker_1): not no seed phrases hidden in this podcast. Um
33:26
Unidentified speaker (speaker_1): like to me, like those are you know, that that's that's very much what math is, right? Like math has data, math has structure that we sort of can see bits of and not understand. And so I don't know, like producing another, you know, another billion pages of it, you know, gives us more toeholds to sort of I realize you
33:45
Unidentified speaker (UNKNOWN): have to say
33:45
Unidentified speaker (speaker_1): toeholds and grab. I think
33:47
Unidentified speaker (UNKNOWN): that's a really I
33:47
Unidentified speaker (speaker_3): think that's a really good vision. In some sense,
33:49
Unidentified speaker (UNKNOWN): if there is
33:49
Unidentified speaker (speaker_2): a forecast, by the way, do you do you expect uh P versus N P to be solved in this
33:53
Unidentified speaker (speaker_3): fashion? We're kind of stuck a little bit. We we kind of don't even have an approach to proving it. So it's gonna definitely uh
34:00
Unidentified speaker (UNKNOWN): it's
34:01
Unidentified speaker (speaker_3): very likely gonna be quite a complicated proof. Um but actually uh the Scott's vision I think is actually quite cool in in that maybe it's a blessing that it's a billion page proof. Because now an army of mathematicians will be will b or computer scientists will be will be happily employed understanding it for the next uh couple of centuries.
34:21
Unidentified speaker (UNKNOWN): It's
34:21
Unidentified speaker (speaker_0): It's worth mentioning briefly,
34:23
Unidentified speaker (speaker_3): right?
34:24
Unidentified speaker (UNKNOWN): We
34:24
Unidentified speaker (speaker_3): actually
34:24
Unidentified speaker (speaker_0): know like formal
34:25
Unidentified speaker (speaker_3): barriers to
34:27
Unidentified speaker (speaker_0): uh proving P does not equal NP, assuming that's the way the answer goes, as uh I d I I maybe I should hesitate to say most of us believe now because some people seem to have flipped their views. But anyway.
34:44
Unidentified speaker (speaker_0): Anyway, so th there are actual barriers of the form like you know, no proof technique in this classic. class of techniques, which feels like it captures like almost anything you could imagine, although, you know, surely there are ways around it, but we haven't come up with any yet. So there are actual barriers, right? And, you know, the the AIs today, uh, I'd
35:03
Unidentified speaker (speaker_0): say like interpretability of of what they're doing, how they're reasoning is a pretty still nascent, you know, uh scientific area. So they're doing some reasoning
35:14
Unidentified speaker (UNKNOWN): that
35:14
Unidentified speaker (speaker_0): we don't fully understand. Um and
35:18
Unidentified speaker (UNKNOWN): i
35:19
Unidentified speaker (speaker_0): I think there's a huge question of will the reasoning that they're doing, you know, uh like really ever even be able to break through these barriers we have? Because we know, you know, it needs
35:29
Unidentified speaker (UNKNOWN): like
35:30
Unidentified speaker (speaker_0): a brand new idea and approach, completely new. Yeah. And so, you know, um the
35:36
Unidentified speaker (speaker_0): it's obviously very powerful. Like no one knows how powerful it's gonna get. Uh it's it but it you know, th there are these s certain things it like maybe just can't ever do, you know, or like, you know, the models would have to completely change their nature before it could do it. I don't know. So you
35:53
Unidentified speaker (UNKNOWN): know
35:53
Unidentified speaker (speaker_0): uh I never want to bet against what these models are gonna be capable of. But also if you're just like, oh, they'll be able to solve every math problem at you know, ever you could ever think of, then you quickly I find, you know, like why does anything matter? You know, you you know, just wait for
36:07
Unidentified speaker (UNKNOWN): the other. Well and
36:08
Unidentified speaker (speaker_1): and and and that's a really hard, yeah, it's a really hard claim to make, right? We sort of we don't even know the full description of all math problems that could ever be described. And so imagining that a system that exists today could in principle solve, or you know, a system with suitable upgrades based on today could in principle solve categories of math. that we don't even you know
36:28
Unidentified speaker (speaker_1): that that could be ex whose existence could be exposed by these systems, but you know, but who knows what they look like. Seems like a hard generalization to reason about.
36:38
Unidentified speaker (speaker_3): So I I I
36:39
Unidentified speaker (UNKNOWN): I
36:39
Unidentified speaker (speaker_3): think about this from from the lens of uh of compression as
36:43
Unidentified speaker (speaker_1): one way to think about this problem. So you know you could think about all of AI as kind of spending a lot of compute and a lot of data in order to compress it. into some you know s relatively small model that that then you know can can give these amazing results. And the human brain is also similar. I mean, we kind of use abstractions and and you know mathematical
37:03
Unidentified speaker (speaker_1): reasoning to have a compressed uh you know brain model of of of say mathematics. Uh and and the okay, so to me the difference is that uh
37:13
Unidentified speaker (UNKNOWN): the way
37:14
Unidentified speaker (speaker_1): we're kind of biologically wired, there there's some some fundamental limits to the amount of compute and storage and networking that that our brains can do. Um so so you know, it and it could be that that uh you know that that limits us to be able to do some sort of sorts of of of mathematics. And maybe not, maybe it can do all, but it definitely is some sort of finite amount of resources. While, you know,
37:34
Unidentified speaker (speaker_1): what are the one of the things that that we're seeing with AI is that we're we can add more and more computation. compute to it, right? And and m more storage. And that can and we're seeing that these models are creating stronger and stronger capabilities. But at the same time, we don't know we don't know whether they're breaking sort of conceptual capability barriers, right? To Justin's point
37:54
Unidentified speaker (speaker_1): earlier, like you know, you could imagine that maybe with enough compute you could find all convex combinations of the set of things that's in the methods that we know cannot prove. P is not equal to NP, and still not make it outside of that bound boundary, right? There's still like it's not obvious that that scaling compute also gets you outside of
38:12
Unidentified speaker (UNKNOWN): the R.
38:12
Unidentified speaker (speaker_1): Right,
38:12
Unidentified speaker (speaker_3): right. But
38:13
Unidentified speaker (UNKNOWN): but I don't
38:13
Unidentified speaker (speaker_3): think there's any
38:14
Unidentified speaker (UNKNOWN): reason
38:14
Unidentified speaker (speaker_3): to
38:14
Unidentified speaker (speaker_0): believe that there's like a fundamental theorem of mathematics that
38:18
Unidentified speaker (UNKNOWN): all
38:18
Unidentified speaker (speaker_0): mathematical results have to fit into a
38:20
Unidentified speaker (speaker_3): human brain.
38:21
Unidentified speaker (UNKNOWN): Um,
38:22
Unidentified speaker (speaker_0): m maybe, maybe it is the case,
38:23
Unidentified speaker (speaker_3): right? That
38:24
Unidentified speaker (speaker_0): there's there is some sort of fundamental compression,
38:26
Unidentified speaker (UNKNOWN): um, but
38:26
Unidentified speaker (speaker_3): maybe
38:26
Unidentified speaker (UNKNOWN): not.
38:27
Unidentified speaker (speaker_0): Maybe there
38:27
Unidentified speaker (UNKNOWN): are
38:27
Unidentified speaker (speaker_0): kind of
38:27
Unidentified speaker (speaker_3): mathematical results.
38:29
Unidentified speaker (UNKNOWN): That was literally the
38:30
Unidentified speaker (speaker_3): P versus NP question.
38:32
Unidentified speaker (UNKNOWN): Exactly. Like I I
38:34
Unidentified speaker (speaker_1): I don't I don't think and I'm I'm not sure I was claiming that like all math results have to fit into a human brain. Like, you know, it's it's just not obvious how
38:43
Unidentified speaker (speaker_1): I mean
38:43
Unidentified speaker (UNKNOWN): let me
38:44
Unidentified speaker (speaker_1): even go go one step you know more in more.
38:47
Unidentified speaker (UNKNOWN): The point you're
38:47
Unidentified speaker (speaker_2): making, right? The point you're making now is that maybe human brains are just too puny for uh you know what
38:54
Unidentified speaker (UNKNOWN): math
38:54
Unidentified speaker (speaker_2): is.
38:55
Unidentified speaker (UNKNOWN): No, on the contrary, I claim he's talking about scale. I claim
38:57
Unidentified speaker (speaker_1): that human brains today are using
38:59
Unidentified speaker (speaker_3): you know pen and paper and computers and phones to augment themselves, um, to
39:04
Unidentified speaker (UNKNOWN): kind of
39:05
Unidentified speaker (speaker_3): improve their computation and storage and communication, to kind of do amazing things that people, humans, could not have done, say a thousand years ago or five 500 hundred or 100 a hundred years ago. And you know, actually I'm I'm very optimistic that that uh you know even more extreme ways to connect uh kind of the human brain to to uh to AI will will allow us to get you know e
39:24
Unidentified speaker (speaker_3): even further than that. Okay, so
39:26
Unidentified speaker (speaker_1): that might be kind of the long-term vision of of where where you know we will be able to deal with you know much, much more complicated problems than than we believe we can deal with today.
39:36
Unidentified speaker (speaker_2): Okay, yeah. So maybe um maybe some of these technical advances that we'll see, we'll be able to like hook ourselves into these AI systems and all of a sudden
39:45
Unidentified speaker (UNKNOWN): get it. I mean I cannot
39:46
Unidentified speaker (speaker_1): wait. Mega
39:47
Unidentified speaker (UNKNOWN): I would I
39:47
Unidentified speaker (speaker_1): would love to uh you know sort of what what is the what is the phrase to to jack into the math universe or something. But um you know it's like it's it's neuromancer, but all number theory is like my dream, right? That's like just going gonna to be amazing. Um But I I I still think
40:03
Unidentified speaker (speaker_1): this question about the you know, sort of what happens in in capability space is very open. Right. Like, you know, if you know some people have said that they think the the what the AI models are doing is sort of like bounded within the convex hull of things that have, you know, you know, that humans have done and and more precisely, I guess been written down, things that sort of are like
40:23
Unidentified speaker (speaker_1): live inside the, you know, sort of the the hull of the training data or something.
40:27
Unidentified speaker (speaker_1): And if that's the case, and something like the, you know, sort of this this no-go theorem for P naught equals an NP, you know, covers that convex hull, right? You know, we could, you know, uh
40:40
Unidentified speaker (UNKNOWN): I
40:40
Unidentified speaker (speaker_1): I don't know whether the models could get past it. And I also don't know, you know, and and who knows whether human plus model could get past it, right? You know, once we've once we've yacked acted into to the uh into the oh way we into the math net. Sorry.
40:53
Unidentified speaker (UNKNOWN): Math
40:53
Unidentified speaker (speaker_1): net you'
40:54
Unidentified speaker (UNKNOWN): M
40:54
Unidentified speaker (speaker_1): MathNet fans here, like square one.
40:57
Unidentified speaker (speaker_1): Dan
40:57
Unidentified speaker (speaker_2): knows what I'm talking about.
40:59
Unidentified speaker (UNKNOWN): Yes.
41:00
Unidentified speaker (speaker_3): Robert, if you want to talk also about the human aspect of it, because there is a pretty significant human aspect too to what
41:05
Unidentified speaker (UNKNOWN): happened
41:06
Unidentified speaker (speaker_3): what happened to the P to PhD students, what happened to what happens to junior researchers.
41:12
Unidentified speaker (speaker_2): I do. Yeah.
41:14
Unidentified speaker (UNKNOWN): Yeah.
41:14
Unidentified speaker (speaker_2): Let's
41:16
Unidentified speaker (UNKNOWN): so
41:16
Unidentified speaker (speaker_2): let's
41:16
Unidentified speaker (speaker_3): take this conversation back to
41:18
Unidentified speaker (UNKNOWN): uh
41:19
Unidentified speaker (speaker_3): back to the human element.
41:20
Unidentified speaker (UNKNOWN): Um
41:21
Unidentified speaker (speaker_3): so we'll depart from the machine for a second and focus on on us, uh, the apes out here. Um
41:27
Unidentified speaker (UNKNOWN): I
41:27
Unidentified speaker (speaker_3): I wanna I want to quote something I read which really struck me. It was a Substack post by a young mathematician, uh a guy, Kerwin Hampshire. Uh he wrote this essay, The Dark Mathematics. Night of Mathematics. Maybe some of you have read it. But he described himself in a really dark place. And he
41:47
Unidentified speaker (speaker_3): said, I'm suffering a profound spiritual crisis due to these developments. I've been screaming internally for days. He said a lot of other things too. Um But it it was it was quite a dark piece. And I think he was articulating. something that a lot of people who engage in math are feeling. You
42:07
Unidentified speaker (speaker_3): know, at the same time that all of this is all very exhilarating, there's all of these incredible advances happening. There's also this feeling that something is lost and um and this feeling of displacement. And and where do I fit in this world that people are having to grapple with? What would you say? to
42:27
Unidentified speaker (speaker_3): somebody like Kerwin who is having these questions about his meaning? There
42:35
Unidentified speaker (UNKNOWN): are
42:35
Unidentified speaker (speaker_1): many
42:35
Unidentified speaker (speaker_3): different
42:36
Unidentified speaker (UNKNOWN): reasons
42:36
Unidentified speaker (speaker_3): why
42:39
Unidentified speaker (speaker_3): many different people are struggling with this. And again, it's it this actually is it's emphatically not limited to math and TCS and sort of like these math-based fields, right? Like many, many people are struggling with
42:50
Unidentified speaker (UNKNOWN): the other thing.
42:51
Unidentified speaker (speaker_3): with the way that AI is shifting the form, generative AI is shifting the form of work in whatever their field is, right? Like, you know, UX design. Uh, you know, I don't I don't know, like, you know, Robert, you talked about writing. You know, there's sort of been a lot of you know, there's been a lot of question about like, you know, are opinion articles assisted by AI really the opinion of the author?
43:11
Unidentified speaker (speaker_3): Like, you know, sort of um
43:13
Unidentified speaker (speaker_3): and and a lot of it, I think also, you know, it it ties into people's identities, right? You know, if you're someone who did plan to spend your entire life thinking about one problem, right? You know, sort of I know people who had expected to spend their entire life thinking about a problem that was solved in the most recent OpenAI drop, right? Like, you know, and
43:33
Unidentified speaker (speaker_3): it's hard to, you know, it's like, well, okay, you know, you're you're
43:37
Unidentified speaker (UNKNOWN): you
43:38
Unidentified speaker (speaker_3): were expecting to spend 50 years on this, it's now been solved. And you have to think about like what is, you know, what is your position of that. Um,
43:48
Unidentified speaker (speaker_3): I
43:50
Unidentified speaker (UNKNOWN): for
43:50
Unidentified speaker (speaker_3): me at least, and and certainly as I've been coaching my graduate students, right, you're sort of people
43:55
Unidentified speaker (UNKNOWN): you
43:55
Unidentified speaker (speaker_3): know, sort of people I've been teaching about this. Um
44:01
Unidentified speaker (speaker_3): first of all,
44:02
Unidentified speaker (UNKNOWN): I
44:02
Unidentified speaker (speaker_3): think
44:03
Unidentified speaker (speaker_3): there's a real sense in which the technology should sort of change it, you know, sort of any big technological shift in in your field of work changes your capabilities, your opportunities. And I think in this case, really, it should very much change our aspirations.
44:22
Unidentified speaker (UNKNOWN): You know,
44:22
Unidentified speaker (speaker_3): I really at the start of this year, right? So not predicting the like level of speed up we we've seen, but like in January of 2026, I told you a couple of my graduate students that they should, you
44:33
Unidentified speaker (UNKNOWN): know,
44:34
Unidentified speaker (speaker_3): that they should be super excited that they were in the first generation of scholars that would ever be able to use generative AI to help with substantive research tasks. And I was thinking like, you know, look for a thing that's gonna like plug, you know, look for that one missing step in this proof you're working on, right? I was not, you know, dreaming like, you know, sort of write papers end day. to end on topics that you know, but I was just like, like this is super
44:53
Unidentified speaker (speaker_3): exciting, right? And you can change the types of theorems you might want to prove because you can go and you because you have this additional input source. And now it's like I think we can change the types of mathematical edifice we build, right? Like there's an opportunity to maybe, you know, sort of discover something fundamental through our telescope to continue Etai's
45:13
Unidentified speaker (speaker_3): metaphor in a way.
45:15
Unidentified speaker (speaker_3): that previously, like, you know, in in you know, was was something only only a handful of people could hope to do at any given point. Whereas now, you know, sort of there's all this new discovery opportunity, and then there's tons and tons of opportunity to sort of consolidate and sort of like distill the knowledge out of those discovers. discoveries.
45:35
Unidentified speaker (UNKNOWN): That
45:35
Unidentified speaker (speaker_3): doesn't solve right, like it that that that's
45:38
Unidentified speaker (UNKNOWN): that's That
45:38
Unidentified speaker (speaker_3): can look like a huge
45:40
Unidentified speaker (speaker_3): opportunity to me, right? I'm personally extremely excited about this, but like it doesn't solve the question if that was not how you saw yourself as a mathematician, or how you saw the work of a mathematician.
45:52
Unidentified speaker (UNKNOWN): But
45:52
Unidentified speaker (speaker_3): the one thing I'll say is that for me, and I know for many math people I know, right, what really got me into math was the love of this knowledge. and opportunity to explore it, right? This like, you know, sort of massive, like sort of truly infinite space. Here, here we do mean uncountable to the uncountable to the grams number or
46:12
Unidentified speaker (speaker_3): whatever. Um, this truly infinite space where there's always more to discover.
46:17
Unidentified speaker (UNKNOWN): And
46:17
Unidentified speaker (speaker_3): so, like to me, like seeing bits of it, right? No matter how big the bit is, just it always opens more opportunity of discovery than it closes, right? And I think, and that's very literal with a lot of these OpenAI and other sort of like Frontier Lab results, right? Like, you know, every time, you know, like if you disprove a conjecture that many other things
46:37
Unidentified speaker (speaker_3): hinged upon, right? If sort of like it was a conjecture that was believed to be true and people had written many papers assuming it was true, suddenly there's like tons of old math that we thought was settled that actually lives in parts of the cave we didn't quite understand. And similarly, if you discover like a totally new method or a totally new sort of like class of ideas, you know,
46:57
Unidentified speaker (speaker_3): often you could do many more things with it. There's more to discover with the sort of new tool that's been discovered on the way to this one particular waypoint. So I hope at least that there's a way for people to sort of find, you know, sort of like find a joy of discovery in this, in this,
47:13
Unidentified speaker (UNKNOWN): I
47:13
Unidentified speaker (speaker_3): don't know,
47:13
Unidentified speaker (UNKNOWN): uh
47:14
Unidentified speaker (speaker_3): discovery. avalanche. Okay, sorry. And and mod you
47:17
Unidentified speaker (UNKNOWN): know,
47:17
Unidentified speaker (speaker_3): that would backslash end sorry, in the order orientation, backslash end open set brace, monologue,
47:22
Unidentified speaker (UNKNOWN): closed set brace.
47:23
Unidentified speaker (speaker_1): Look, the the the the truth of the matter is that we are in a period of transition.
47:28
Unidentified speaker (speaker_3): Uh
47:29
Unidentified speaker (speaker_1): you know, we don't quite know where things are gonna land once the transition happens.
47:34
Unidentified speaker (UNKNOWN): Um
47:35
Unidentified speaker (speaker_1): I'm optimistic that it will land at a at a good place. Uh but not everybody is optimistic.
47:41
Unidentified speaker (UNKNOWN): And
47:42
Unidentified speaker (speaker_1): so, yes, there is there is a human problem here in that obviously, for example, when we when we evaluate new faculty candidates,
47:51
Unidentified speaker (UNKNOWN): you
47:51
Unidentified speaker (speaker_1): know, a faculty candidate could come in and they might have a hundred publications,
47:57
Unidentified speaker (UNKNOWN): but
47:57
Unidentified speaker (speaker_1): they never they didn't write any of those hundred publications and they didn't even invent a hundred publications, they just prompted them all. the model. Whereas somebody else may have spent their four years doing a PhD working really hard. They maybe solved two hard problems, but they solved them themselves. And how do you compare such two such candidates? Yeah, what what's um
48:16
Unidentified speaker (UNKNOWN): what
48:17
Unidentified speaker (speaker_1): should we even be optimizing for? So that that's actually one of the challenges that our students don't really know what to optimize for at this moment. Yeah, because we're in a we're in a period of transition.
48:27
Unidentified speaker (UNKNOWN): Um
48:27
Unidentified speaker (speaker_1): so it's it's it's a little it's obviously hard to tell.
48:30
Unidentified speaker (UNKNOWN): Um the
48:31
Unidentified speaker (speaker_1): advice that I try to give is that you know, we are moving into a world where just sort of uh technical prowess is not kind of the point because the machines are so technically strong in what they do. In some sense, our contribution as humans is coming up with this notion of taste. Like what are the interesting problems? there should be
48:51
Unidentified speaker (speaker_1): we should be working on.
48:53
Unidentified speaker (UNKNOWN): So
48:53
Unidentified speaker (speaker_1): in some sense, um I would say that uh like a PhD student who's going through a PhD program, the point of the PhD program now is not to solve two hard problems and get a PhD. It's more to develop a research research taste for what are good pro how do you come up with good problems? How do you tell a good problem for from but a not s not such a good problem?
49:12
Unidentified speaker (UNKNOWN): So
49:13
Unidentified speaker (speaker_1): kind of the goal is to go kind of more develop a framework of doing research, do it, you know, developing good taste in doing research. And that's actually also something that we can uh look for when we uh when we interview candidates, you know, who has good vision for where the field is going and versus not.
49:30
Unidentified speaker (UNKNOWN): Um so,
49:30
Unidentified speaker (speaker_1): in some sense, the the
49:32
Unidentified speaker (UNKNOWN): uh uh
49:32
Unidentified speaker (speaker_1): advice right now, and again we're in a period of transition, so this could change at any moment,
49:36
Unidentified speaker (UNKNOWN): but
49:36
Unidentified speaker (speaker_1): the advice right now is maybe kind of focus on this. developing more a vision for your field rather than focus on one hard problem that your whole life's goal is just to solve that one hard problem.
49:47
Unidentified speaker (speaker_0): I definitely want to, you know,
49:49
Unidentified speaker (UNKNOWN): sort of
49:49
Unidentified speaker (speaker_0): second uh a lot a lot of what Dan said. Now the raw pro raw problem solving skills maybe get devalued because that's exactly what the AI is so good at. But uh you know
50:01
Unidentified speaker (UNKNOWN): uh for
50:02
Unidentified speaker (speaker_0): for me. personally, it it's been, you know, uh just uh all around, you know, uh terrific, right? Like uh because, you know, I've spent years developing like a a vision um i you know in the various areas I I work in and and this is just a way to
50:22
Unidentified speaker (speaker_0): to execute it uh you know faster and more effectively um but uh the the other thing I'll say is um right, and by the way, like the the AI is like total able to make, you know, uh
50:35
Unidentified speaker (UNKNOWN): uh
50:36
Unidentified speaker (speaker_0): great progress on some of the the questions like I've been stuck on, and
50:40
Unidentified speaker (UNKNOWN): then
50:40
Unidentified speaker (speaker_0): others is just totally and completely still
50:41
Unidentified speaker (UNKNOWN): stuck
50:42
Unidentified speaker (speaker_0): and I go nowhere. Right. Um but uh a lot of what my personal work has has been well, I'll say two things. So one is uh I I think i life is a little simpler in like the more applied areas because you just have like very tangible and urgent goals and achieving those goals are still is still challenging
51:02
Unidentified speaker (speaker_0): even with today's tools. And so like obviously the thing to do is just like work as hard as you can to achieve those goals using the existing tools.
51:09
Unidentified speaker (UNKNOWN): And So when your productivity goes up, right,
51:11
Unidentified speaker (speaker_1): like the shadow value of your work go gets higher, you can actually work even harder, right? Right.
51:16
Unidentified speaker (speaker_0): Right. And so yeah, with like advanced cryptography and stuff, it's like, you know, society urgently needs these protocols. There's a you know post-quantum transition, there's, you know, uh concerns about AI, you know, cryptanalysis. And so it's urgent to, you know, to alleviate those concerns to the extent possible or, you know, just see them through
51:36
Unidentified speaker (speaker_0): to the end. Um so everything, you know, feels urgent and like clearly you just gotta work and get it done. Um, and so that that that's one thing that can be clarifying, but that that, you know, there are different flavors of research. Um, the other thing I'll say is like a lot of my work now,
51:51
Unidentified speaker (UNKNOWN): you know,
51:51
Unidentified speaker (speaker_0): I just spend like years on years trying to convince people that like a certain general approach or philosophy is the right way to think about things and do things. And, you know, uh you make very gradual progress. People are just, I'm saying there's a very human problem, right? Like, uh, yes, by getting years and years and years of results building my case, you
52:11
Unidentified speaker (speaker_0): help, you know, convince people, but they're they'll never be fully convinced uh without like, you know, the vision being, you know, communicated clearly and the results being packaged correctly. And people, you know, can only hold so much nuance in their heads. So you have to give that, you know, you have to not lie to them. but you have to give them like, you know, the nuance anyway. And so all of these things are still
52:31
Unidentified speaker (speaker_0): challenging, maybe even more challenging now with AI putting out these inscrutable proofs. So there's plenty to do, there's lots of opportunity. I think this is, you know, uh like some PhD students are going to hear me say this and be like, well, easy for him to say, right? Uh but you know, i we're just in a transitionary time, like like Dan said, and the, you know, the thing to
52:51
Unidentified speaker (speaker_0): do is there are opportunities and there's displacement, and uh you have to do the best you can to make the most of the opportunities, and that's all anyone can do.
53:02
Unidentified speaker (UNKNOWN): Uh
53:02
Unidentified speaker (speaker_3): maybe I'll quickly add, I I to me at least, you know, this is the biggest change in mathematics in the history of mathematics. Um and so I think it's okay to uh acknowledge the loss. The loss of you know, the world three years ago will not,
53:19
Unidentified speaker (UNKNOWN): you
53:19
Unidentified speaker (speaker_3): know, the the world will not stay as it was three years ago. And I think for for many people, this is a a huge change. I think acknowledging this loss and you know letting some, you know, handling some of that, uh understanding that there's going gonna to be grief, there's gonna be a period of of transition i i is important. Um
53:39
Unidentified speaker (speaker_3): the next thing is about you know a lot of change. So some people do not like a lot of change. Um then I would kind of say this is not the right time uh to do research in in in theoretical computer science or or or in mathematics. Um and and some people actually enjoy change more. Uh so
53:59
Unidentified speaker (speaker_3): I think this is for for them, it's it's an opportunity.
54:02
Unidentified speaker (UNKNOWN): Um
54:02
Unidentified speaker (speaker_3): Uh specifically, you know, we talked about these two frontiers. I think the the discovery frontier right now is is still such that you can you can work with the the the frontier models to kind of co uh create new results. That might change very quickly. Um
54:22
Unidentified speaker (speaker_3): so you know, to the extent that you can do it now, you should do it now. But in a sense, you know, the advice I give is to focus on this second part, on moving the human understanding frontier. And this goes back to what Dan said about asking the right questions, you know, focusing on
54:39
Unidentified speaker (speaker_3): where should we, you know, what parts of the results should we actually build human understanding in? And human understanding is not just getting the most recent results from the AI model. models and writing a proof that is clear to mathematicians. There's such a a large kind of ecosystem of
54:58
Unidentified speaker (speaker_3): explanation that in the end go all the way down to
55:05
Unidentified speaker (speaker_3): elementary school.
55:06
Unidentified speaker (speaker_3): And really, if you think about it,
55:09
Unidentified speaker (speaker_3): The w the way that we learn math today um i i in in school, in in in high school, in elementary school. in college, is so much better um because we have such a better uh human understanding. And so, you know, I I I say if you're passionate about improving human understanding of mathematics, I think this is a really great time um to to work in this field.
55:29
Unidentified speaker (speaker_3): If the only thing you're passionate about is solving problems and competing and showing that you can solve problems um faster or or better than other people. then I think this is going to be a challenging time for you. Um there there are still things you can do. And I think they're, you know, going back to again Dan's analogy of of chess, you know, there might there might be kind of a an
55:49
Unidentified speaker (speaker_3): ecosystem. Uh maybe I'll I'll let Dan talk about about that that option. But but but in a sense, um, you know, I think if if your goal is to kind of um push the frontier, then then I think, you know, that will be very, very hard to do going forward.
56:04
Unidentified speaker (UNKNOWN): Yeah,
56:04
Unidentified speaker (speaker_1): I like to
56:04
Unidentified speaker (UNKNOWN): say that even though even though
56:05
Unidentified speaker (speaker_1): a machine can always beat us in chess, we still enjoy p playing chess with one another.
56:10
Unidentified speaker (UNKNOWN): So it's more
56:11
Unidentified speaker (speaker_1): popular than ever. It doesn't take the joy of the
56:13
Unidentified speaker (speaker_2): I will say I've been playing a lot of chess
56:16
Unidentified speaker (speaker_2): band over the past years.
56:17
Unidentified speaker (speaker_1): But you
56:17
Unidentified speaker (UNKNOWN): know,
56:17
Unidentified speaker (speaker_1): I do want I do want to sharpen a little bit what Itai said in that um uh there's a very human aspect of what has been what is being lost here. And I think that's maybe why worth uh explaining.
56:29
Unidentified speaker (UNKNOWN): Um
56:29
Unidentified speaker (speaker_1): and so I can tell you personally, when I solve a problem, I just feel happy. It's like uh I'm sure all of everybody uh has felt that. When you solve a new problem for the first time,
56:39
Unidentified speaker (speaker_2): you've been working on it for days and days and days, and all of a sudden the uh the solution kind of pops into your head. But I'll tell you even more when then I go explain to my friend,
56:48
Unidentified speaker (UNKNOWN): look
56:48
Unidentified speaker (speaker_2): what I look what I discovered, they get super. excited. They go, Oh wow, this is a beautiful idea. They start to build on it. And this is why there's a mathematical community, right? We share ideas and also a theoretical computer science community. We share ideas, we build on each other's ideas. That's what exactly what forms a community. It's a very social
57:07
Unidentified speaker (UNKNOWN): uh
57:08
Unidentified speaker (speaker_2): science, it's a very social activity that we kind of work uh on things together, build on each other's work and so on. And a lot of that, I think, is what Itai was saying. Basically, a lot of that is being lost.
57:21
Unidentified speaker (UNKNOWN): Yeah,
57:21
Unidentified speaker (speaker_2): this feeling of happiness is not going to be as sharp given that you know that a machine already knows the answer. You didn't
57:28
Unidentified speaker (UNKNOWN): really discover
57:28
Unidentified speaker (speaker_2): it yourself.
57:30
Unidentified speaker (speaker_2): Oh, I I like
57:31
Unidentified speaker (UNKNOWN): I
57:31
Unidentified speaker (speaker_2): hope
57:32
Unidentified speaker (UNKNOWN): please,
57:32
Unidentified speaker (speaker_2): Scott, prove me wrong. Yes, please.
57:34
Unidentified speaker (UNKNOWN): Yeah, yeah. So we can put in
57:35
Unidentified speaker (speaker_1): the show notes a couple of these. Like like I have
57:38
Unidentified speaker (UNKNOWN): read.
57:38
Unidentified speaker (speaker_1): papers that have AI disclosure acknowledgments at
57:42
Unidentified speaker (UNKNOWN): the end
57:42
Unidentified speaker (speaker_1): that say things like that say, like, I've thought about this for a decade. I have always wanted to know the answer. You know, I spent a weekend doing this, you know, sort of with the help of GPT, whatever. And, you know, and
57:54
Unidentified speaker (UNKNOWN): I
57:54
Unidentified speaker (speaker_1): don't know whether I deserve credit for this, but I don't care. Right. I know the answer. It's so cool. And like, you know, what I found, you know, I there's a problem, interestingly. so to the point about, you know, sort of the pro Itai, your problem that you'd worked on for, you know, sort of on and off for for 10 years, right? There's there's a problem I've worked on and off on at on and off for over 20 years, starting in
58:14
Unidentified speaker (speaker_1): 2004.
58:16
Unidentified speaker (speaker_1): And there's a second such problem that I've worked at on and off since like about 2007. I solved the 2007 one recently. I have made essentially no progress on the 2004 one. right? Like and not, you know, sort of, you know, same same model,
58:29
Unidentified speaker (speaker_3): same test. Every time a new model comes out, I like, you know, throw more of it at the 2004 problem. But I recently solved the
58:33
Unidentified speaker (speaker_1): 2007 one.
58:35
Unidentified speaker (UNKNOWN): And
58:35
Unidentified speaker (speaker_3): I wrote
58:35
Unidentified speaker (UNKNOWN): to
58:36
Unidentified speaker (speaker_1): some
58:36
Unidentified speaker (speaker_3): mathematicians that I had actually never met, right? That these are mathematicians whose work I'd followed for years, but I'd never had a reason to communicate with them because I wasn't really working in this area super actively at the time. And sent them, you know, here's this thing, you know, that I that I've you know just and got back this super exciting. excited, like, oh my gosh, that's so cool. Like,
58:53
Unidentified speaker (speaker_1): like
58:53
Unidentified speaker (speaker_3): it was exactly the experience. Um,
58:55
Unidentified speaker (speaker_1): you know, and and it was all over the the email and in the text that like, you know, and I react the same way when people discover things and share them with me, right? Like, like so much of the joy of knowledge is is in like, oh my gosh, like, you know, could never have believed that. Or then, by And the way, towards the human understanding and distillation point, right? Recently, there was this whole, you know, series of
59:15
Unidentified speaker (speaker_1): reductions of, you know, sort of uh the OpenAI proof that there are no Landau's equal zeros. There was this whole like series of discoveries that you could reduce this more and more and more and more until it eventually got down to like about a three-page argument. Um, you know, that that really hinges off of like a very, very elegant determinant construction. Um, and
59:35
Unidentified speaker (speaker_1): then there's another, like, you know, sort of longer argument.
59:38
Unidentified speaker (UNKNOWN): Uh
59:38
Unidentified speaker (speaker_1): again, we could put all of these in the show notes. There, there's like a you know a panopoly of this point. There's another slightly longer, but like, you
59:43
Unidentified speaker (UNKNOWN): know,
59:44
Unidentified speaker (speaker_1): structural argument that gives you more of a sort of like more of a bound. And like all the mathematicians who've thought about this in like rapid fire, they're literally like we're talking about it on Twitter. Like, there's all this excitement about the beauty of the idea that is being discovered and sort of explored in real time in a very social way that I think is. much more
1:00:04
Unidentified speaker (speaker_1): social than the sort of imagined like mathematicians working, you
1:00:08
Unidentified speaker (UNKNOWN): know,
1:00:08
Unidentified speaker (speaker_1): sort of locked in their attics and then like seeing each other at conferences and talking.
1:00:11
Unidentified speaker (UNKNOWN): So I don't know. Scott, Scott, I love
1:00:13
Unidentified speaker (speaker_2): I love what you're saying. I love what you're saying, and I really hope, I really hope this is kind of discontinues. My analogy for this, I've been, by the way, I've been uh having a blast reading the AI-generated AI generated proofs. They're so pretty. I mean, it's really, really enjoyable. I kind of have this mental model of my in my mind that there's like this huge tree with wonderful fruit on it and the the
1:00:33
Unidentified speaker (speaker_2): that's the AI. And the kind of these fruits for oh from time to time they kind of fall on us and we really enjoy examining and and and um and looking at them. And and so I I really uh hope that what you're saying will continue. I do worry that uh in the long run the what you said uh you know. there was a social process of improving
1:00:53
Unidentified speaker (speaker_2): improving what the AI said. If the AI is if if in fact the models actually get sufficiently strong, that social um process of improvement would also be done by the by the machines. And
1:01:06
Unidentified speaker (UNKNOWN): maybe
1:01:06
Unidentified speaker (speaker_1): we're we're maybe like you know.
1:01:07
Unidentified speaker (UNKNOWN): I really hope
1:01:08
Unidentified speaker (speaker_1): you're right. I really
1:01:08
Unidentified speaker (UNKNOWN): hope
1:01:08
Unidentified speaker (speaker_2): you're
1:01:09
Unidentified speaker (UNKNOWN): right. Well, but I
1:01:09
Unidentified speaker (speaker_3): want to try
1:01:11
Unidentified speaker (speaker_2): go ahead.
1:01:12
Unidentified speaker (speaker_3): I I I want to try to split uh uh
1:01:19
Unidentified speaker (speaker_3): So I guess you know one of them is going to be related to uh to to the fact that you figure out something and understand how it works. You have this eureka moment. And the way I interpret this is that this happiness comes from basically compression. You know, going back to to to this discussion about compression. That you you can
1:01:39
Unidentified speaker (speaker_3): finally kind of understand this in a in a succinct succinct manner, uh, and that kind of gives you a lot of satisfaction. There's a very simple way to understand things. And and here's you know something amazing that happened to me once. I I worked on on some problem and and I I solved some some lemma and then and I had to, you know, I had some other deadlines. I really had to work on other things. And then I
1:01:59
Unidentified speaker (speaker_3): came back to it after three months and I didn't make any notes. And I lost, you know, I lost all record of this time. I only remembered that I could prove it back then. And I worked on it again and I solved it again. And I had this amazing satisfaction the second time I, even though I knew I already
1:02:14
Unidentified speaker (UNKNOWN): Oh, yeah,
1:02:14
Unidentified speaker (speaker_1): absolutely.
1:02:15
Unidentified speaker (speaker_3): So
1:02:16
Unidentified speaker (UNKNOWN): I was a
1:02:16
Unidentified speaker (speaker_1): semi-professional puzzle designer. Like when I resolved the puzzles I created after forgetting the solution, you get that same level of joy. Sorry, go ahead.
1:02:23
Unidentified speaker (UNKNOWN): There you go. No,
1:02:24
Unidentified speaker (speaker_3): no, there you go. So this to me is like one part of of the how. the happiness that I think will will um will survive, right? The the the enjoyment of elegance, which is again this ability to compress things. At least, at least this is how I interpret
1:02:37
Unidentified speaker (UNKNOWN): it.
1:02:37
Unidentified speaker (speaker_3): Um the other part is the feeling I am the first one ever in the history of the world um to have actually solved this problem. I think that that will uh have to change, either to change to saying, I'm the first human who ever understood this problem. Um but but
1:02:57
Unidentified speaker (speaker_3): I think that this the the feeling that that uh that uh mathematicians had again for thousands of years where I'm the first person in the world you know I of all the machines existing in
1:03:07
Unidentified speaker (UNKNOWN): in
1:03:07
Unidentified speaker (speaker_3): in planet Earth, I'm the first who who understands uh why this is true. Uh you know that feeling
1:03:15
Unidentified speaker (UNKNOWN): probably not confused. Why do you
1:03:17
Unidentified speaker (speaker_1): why do you think of this as going away? Like I feel like this opportunity is could you know people like many, many more people can have this experience than used to be able to. Not just that like the individual mathematician might be able to have it more than before, but
1:03:31
Unidentified speaker (speaker_3): oh I mean, you could experience thinking that you're the first. But I I guess you know it's very hard to measure uh who's the first. I
1:03:36
Unidentified speaker (UNKNOWN): mean,
1:03:36
Unidentified speaker (speaker_3): okay, okay. So I guess, you
1:03:37
Unidentified speaker (UNKNOWN): know, in the old. world, I mean, like,
1:03:38
Unidentified speaker (speaker_1): you
1:03:39
Unidentified speaker (UNKNOWN): know,
1:03:39
Unidentified speaker (speaker_3): if you were the first, you could easily you you could just you could publish your result and everybody would be shocked, right? Because that they couldn't repeat this proof. So it would be very clear and obvious that in this competition, right, you are first. While I guess you know, now it's gonna be much harder uh for you to to kind of make that claim. May again maybe
1:03:56
Unidentified speaker (UNKNOWN): you could. I mean, like,
1:03:58
Unidentified speaker (speaker_1): you
1:03:58
Unidentified speaker (UNKNOWN): know,
1:03:58
Unidentified speaker (speaker_1): again, I don't think the importance of the claim, At least for me, you know, the joy of of sort of knowing. a fact that is sort of a new fact has never really been specifically about being the first person in the world to know the fact, right? It's it's much more that you sort of like know something that's like sort of new knowledge that's formed somehow like primordially from through thinking or something like this. But
1:04:18
Unidentified speaker (speaker_1): like, you know, now we're seeing lots of people who have never worked in mathematics before. By the way, Robert, you talked about sort of and when we talked about the identity and human challenges of
1:04:31
Unidentified speaker (speaker_1): people who are seeing math change out from under them, we should we should also talk about people who are seeing math become an opportunity for them, right? Like there are many people who have never done pure mathematics before and are now getting excited about it and discovering things where they're the first person in the world to have discovered them. And like, I I, you know, uh uh certainly, you know, many, many, if not all of
1:04:51
Unidentified speaker (speaker_1): them, get that same type of joy. Like,
1:04:53
Unidentified speaker (UNKNOWN): you
1:04:53
Unidentified speaker (speaker_1): know, what's the like what's the concern, right? Like you, you've you've found your way to a discovery and you don't
1:05:00
Unidentified speaker (UNKNOWN): have to be I have to say I
1:05:01
Unidentified speaker (speaker_2): I love I love your point of view. Actually, I I completely support what you're saying. I love I love your point of view. And I have to say this actually mirrors kind of my experiences that even though uh there's a problem that maybe I couldn't solve, I mean I literally has happened this to me since six months ago, going back to Robert's very first question. A problem I couldn't solve that uh I asked the AI the machines they
1:05:21
Unidentified speaker (speaker_2): solved them. I was so happy to actually see the solution. So
1:05:24
Unidentified speaker (UNKNOWN): the
1:05:24
Unidentified speaker (speaker_2): joy that we the the joy that we get actually from understanding something that we didn't understand a week ago is likely is likely to remain. So I really, I really love the the vision you're you're describing here. The beauty by the way of living through this period is that we get to def we and our students and our community, we get to define where things
1:05:43
Unidentified speaker (UNKNOWN): are
1:05:43
Unidentified speaker (speaker_2): headed. Yeah, what the community will look like after. So this is this decade that we're living through is probably going to be a decade that people r will write history books about. And so you know there's a kind of an important role on you know important role on our shoulders.
1:05:57
Unidentified speaker (UNKNOWN): It
1:05:58
Unidentified speaker (speaker_1): is great, I mean what you guys are describing that there can still be this thrill, this joy of discovery in this.
1:06:03
Unidentified speaker (UNKNOWN): And it can become more
1:06:04
Unidentified speaker (speaker_1): broadly accessible.
1:06:06
Unidentified speaker (UNKNOWN): Right?
1:06:06
Unidentified speaker (speaker_1): Yeah.
1:06:06
Unidentified speaker (UNKNOWN): And
1:06:07
Unidentified speaker (speaker_1): maybe even, yeah, maybe even more democratized. Maybe more people can even experience this and access it.
1:06:21
Unidentified speaker (speaker_1): Mathematics professor he wrote for The Atlantic recently. Um, He was describing the um, you know, the just the salvo of results coming out of OpenAI, the 400 plus papers on uh various math problems. And he just he described one of them, which his friend had labored over for many years and had this bottle
1:06:38
Unidentified speaker (speaker_2): of scotch or whatever that he was going to celebrate once somebody figured it out. Um he described it as uh, you know, when when the result was released in the OpenAI papers, uh he said, quote, it was just another trophy in OpenAI's bag, one of hundreds presented like severed rhino heads.
1:06:57
Unidentified speaker (UNKNOWN): I
1:06:57
Unidentified speaker (speaker_2): thought that imagery was shocking. Um but you know, I it for him, I think that he maybe some of the joy is is gone that he was expecting to have.
1:07:10
Unidentified speaker (UNKNOWN): It certainly certainly
1:07:11
Unidentified speaker (speaker_1): doesn't sound like a joyous like
1:07:13
Unidentified speaker (UNKNOWN): a bottle of
1:07:15
Unidentified speaker (UNKNOWN): thrill of the hunt was
1:07:17
Unidentified speaker (speaker_3): this is exactly why we're focusing on you know what what makes you happy because in a sense if if your expectations right are are that I you know you'll be happy from from let's say again um making progress in this discovery frontier, then you will probably not be very happy in the future. Um
1:07:32
Unidentified speaker (UNKNOWN): I disagree. I
1:07:33
Unidentified speaker (speaker_1): don't think that's obvious at
1:07:35
Unidentified speaker (UNKNOWN): all.
1:07:36
Unidentified speaker (UNKNOWN): Like
1:07:37
Unidentified speaker (speaker_1): I just think it's a different shape
1:07:40
Unidentified speaker (UNKNOWN): of
1:07:40
Unidentified speaker (speaker_3): how we
1:07:40
Unidentified speaker (speaker_1): make
1:07:41
Unidentified speaker (speaker_3): progress. Right?
1:07:42
Unidentified speaker (speaker_1): Like
1:07:43
Unidentified speaker (UNKNOWN): No, no,
1:07:44
Unidentified speaker (speaker_3): but I'm talking about what what you enjoy, right? What makes you happy. So if if what makes you happy, right, is uh is is the feeling that you know you discovered it or or your colleagues discovered it. And then now now AI is gonna write a lean proof saying I proved it, and and now suddenly you're unhappy and and you
1:08:01
Unidentified speaker (UNKNOWN): put
1:08:01
Unidentified speaker (speaker_3): it
1:08:01
Unidentified speaker (UNKNOWN): in. see this as the equivalent
1:08:02
Unidentified speaker (speaker_3): of
1:08:05
Unidentified speaker (UNKNOWN): I wanna I want to push
1:08:05
Unidentified speaker (speaker_1): on this. What if somebody had written a computer program? Like sitting in their at their desk, they wrote a computer program that derived the lean proof through some sort of Turing machine type descent.
1:08:15
Unidentified speaker (UNKNOWN): Right.
1:08:16
Unidentified speaker (speaker_1): So you're trying to prove this thing, right? And this happens sometime in math, sometimes in math versions of this. People like came up with computer computer-aided aided proofs of results for which the the sort of like the fully analytic proof was out of reach. and people were working on the fully analytic proof.
1:08:30
Unidentified speaker (speaker_1): How different is that? And should we like think of that as the same? Or is this just that at larger scale? Or is it is it different because now it has this like sort of distant telescope effect? Like I don't
1:08:42
Unidentified speaker (speaker_1): I don't I don't know that like disc discovery is going to be lost. Like if anything, I continue to think it might get, you know, the surface area. of potential discovery expands, and the opportunities to discover a log that expand, the methods of discovery may change.
1:09:01
Unidentified speaker (UNKNOWN): But I don't
1:09:01
Unidentified speaker (speaker_1): think
1:09:01
Unidentified speaker (UNKNOWN): it's a good idea.
1:09:03
Unidentified speaker (UNKNOWN): So
1:09:03
Unidentified speaker (speaker_1): enjoy discovery. They they're they're they're they're done, right? Like I I I I think that's the wrong answer. The wrong answer.
1:09:09
Unidentified speaker (UNKNOWN): I mean,
1:09:09
Unidentified speaker (speaker_0): your your take is like this reaction um of the mathematician to the OpenAI dump of math results is is sort of like a a wrong reaction,
1:09:18
Unidentified speaker (UNKNOWN): if you
1:09:18
Unidentified speaker (speaker_1): will.
1:09:19
Unidentified speaker (UNKNOWN): I
1:09:19
Unidentified speaker (speaker_1): don't want to I I I don't want to judge it as as right or wrong. Like I I don't I don't you know
1:09:24
Unidentified speaker (UNKNOWN): I
1:09:24
Unidentified speaker (speaker_1): can't speak for that mathematician and what they found sort of motivating and joyous, right? Like I like I don't like I don't want to claim that they're misunderstanding something or that their feelings are are not true, right? Like they are. Um
1:09:37
Unidentified speaker (UNKNOWN): I mean
1:09:38
Unidentified speaker (speaker_1): presumably, I guess. I I I haven't actually talked to them, but but as reported by by Strogatz, presumably, you know, sort of accurate. Um but but I don't think the takeaway is that if you're the type of person who enjoys exploring the discovery frontier, that's going to go away.
1:09:55
Unidentified speaker (UNKNOWN): Like
1:09:56
Unidentified speaker (speaker_1): I think those are very different statements, right? I think it's it's you know uh it's totally possible that discovery in the way we've done it historically may disappear, right? I I I'm not even sure that's gonna be true, but it but it's plausible, I think. But I don't think the message is if you're someone who likes discovery, you know,
1:10:15
Unidentified speaker (speaker_1): per se,
1:10:17
Unidentified speaker (UNKNOWN): you
1:10:17
Unidentified speaker (speaker_2): know, that you're like forget it, like mass, you
1:10:20
Unidentified speaker (UNKNOWN): know,
1:10:20
Unidentified speaker (speaker_2): mass over for you.
1:10:21
Unidentified speaker (speaker_1): Like I just I don't
1:10:21
Unidentified speaker (speaker_2): think that's the But it but it could be that the skill set is gonna going to be quite different, right? Your skill is gonna
1:10:26
Unidentified speaker (UNKNOWN): be. more.
1:10:28
Unidentified speaker (speaker_1): Yes.
1:10:29
Unidentified speaker (speaker_2): Let's talk about those skill sets because a lot of you are teachers, professors, you deal with students. How are you adapting what you teach, the course of your program, what you want to instill in the next
1:10:43
Unidentified speaker (speaker_1): generation of mathematicians? What
1:10:44
Unidentified speaker (UNKNOWN): are
1:10:44
Unidentified speaker (speaker_1): you changing?
1:10:45
Unidentified speaker (speaker_2): Oh my God. That is the I think the hardest problem that we are all facing. Everything
1:10:50
Unidentified speaker (speaker_1): every week.
1:10:51
Unidentified speaker (speaker_2): If I want
1:10:52
Unidentified speaker (UNKNOWN): to
1:10:52
Unidentified speaker (speaker_2): close out on
1:10:52
Unidentified speaker (UNKNOWN): it.
1:10:53
Unidentified speaker (speaker_2): You know, it's very interesting because a lot a lot of the discussion when it comes to uh teaching is how should we teach? Because the students are cheating with AI, supposedly, quotes cheating. That's no the wrong question. Like the mechanics of t the mechanics of teaching, that's not the issue. You know, we'll figure how to teach, we'll figure out.
1:11:10
Unidentified speaker (UNKNOWN): I
1:11:11
Unidentified speaker (speaker_2): think the much bigger question is what should we teach? Yet what in the age of of of these uh amazing machines, what what does it mean for um you know a graduate to be a graduate in mathematics and computer science and physics, what what is it that we need to teach? Um that is a much, much harder uh question to answer. And to be honest, I I I I don't have an answer.
1:11:31
Unidentified speaker (speaker_2): I mean, I w I wish I did. Um I think it's one of these things where again, I'm gonna say it over and over again, we're in we're in a transition period. Uh we are figuring out, you know, we're kind of in a dark room right now. We're kind of blindly uh struggling and figure and kind of uh feeling our way around the room. I think a picture will emerge, um and it's gonna be quite
1:11:50
Unidentified speaker (speaker_2): different. So I think with a what I think I think there are some basic concepts that uh I of course I would say it from a computer science point of view. I think there are basic concepts like algorithms and cryptography would be included in that, that everybody needs to know because it's just important for a computer scientist to know these things. Um but uh uh uh the you know what exactly is
1:12:10
Unidentified speaker (speaker_2): being taught, to what level is being taught, I think that's still, at least for me, this is still something that needs to be figured out. I mean I w I wish I we already knew the answer. I think in a few years we will figure it out, but right now we're still kind of struggling and feeling blindly around this room. Aaron
1:12:25
Unidentified speaker (UNKNOWN): Powell
1:12:25
Unidentified speaker (speaker_2): Have you made any changes? like even in this experimental phase where you're trying to come up
1:12:30
Unidentified speaker (UNKNOWN): with a new
1:12:31
Unidentified speaker (speaker_2): Yeah, what what have you done uh concretely? So again, I don't want to talk about the mechanics of teaching. That of course changes because we have to change how we teach because you have a big button that just you press and it does your your homework for you. So obviously the mechanics of teaching has to change. But I again, I want to say this is not the interesting question. The interesting question is what needs to be taught? What should we be teaching?
1:12:51
Unidentified speaker (speaker_2): And I I would say that like um in the past, uh I try to kind of teach things that I I think that uh computer scientists, for example, would need to know for their jobs. Yeah, that would make them um um
1:13:07
Unidentified speaker (speaker_2): i
1:13:07
Unidentified speaker (UNKNOWN): you
1:13:07
Unidentified speaker (speaker_2): know good computer scientists as I I I would say. Um my thinking now is maybe maybe we need to focus more on uh higher level concepts. So more concepts, less um less detail.
1:13:21
Unidentified speaker (UNKNOWN): Um
1:13:21
Unidentified speaker (speaker_2): so again, these are these are just kind of uh right now these are uh high-level thoughts, but maybe this is kind of the directions where thing where things will go. Again, I want to emphasize we're on transition transition period. We don't quite know where we're gonna Going to land yet.
1:13:35
Unidentified speaker (speaker_3): So yeah, I I teach a relatively small course. Um and and uh I always try to do it very interactive. And and that kind of creates a dynamic inside the course where um you know the there's kind of uh uh human interaction.
1:13:50
Unidentified speaker (speaker_1): Uh
1:13:51
Unidentified speaker (speaker_3): and I I think that is actually going to be more valuable. So you know, being able to interact, uh to verbally express ideas uh is is going to be kind of uh important. The the other thing I I I did was actually stress the importance of human understanding.
1:14:06
Unidentified speaker (speaker_1): So
1:14:07
Unidentified speaker (speaker_3): to the extent that students are spending a lot of efforts going to to university, taking courses, I think if they understand that if if they will basically go to a course and as Dan said, press a button and get results, but don't will not understand anything, then they're wasting their time as well. So I think you know students will also understand
1:14:27
Unidentified speaker (speaker_3): that they're here or they're, you know, they're their studying in order to get human understanding, uh, will kind of uh motivate them to to actually understand. And and I'll just continue last thing that you know around Dan's point. I think we're seeing AI be very good at at a lot of low-level things. Um
1:14:47
Unidentified speaker (speaker_3): and I think understanding higher level abstractions is actually going to be more and more important. So, you know, understanding the found, you know, you don't have to understand all the gory details, but to understand abstractions correctly, um, to be able to kind of, you know, use uh the right framework of things, ask the right questions. These are
1:15:07
Unidentified speaker (speaker_3): really important uh tools. And so I think focusing on these types of of of ideas in a course and also also kind of telling students up front, you know, this is what we're going to focus on. This is what you should focus on. Because otherwise, you
1:15:21
Unidentified speaker (UNKNOWN): know,
1:15:21
Unidentified speaker (speaker_3): you're you're just gonna uh not spend your time efficiently in the course. Um so I think this is kind of what I'm what I'm uh experimenting right now with.
1:15:30
Unidentified speaker (speaker_2): I I think it is important for me though just to say one more point, which is uh we can so all this stuff is has is this like happening in the last year, right? I mean the models have only kind
1:15:38
Unidentified speaker (UNKNOWN): of
1:15:39
Unidentified speaker (speaker_2): become at the level where we have to start thinking about what to teach uh it's only in the last couple of months, right? Um and so we are starting to adapt, like what Itai was saying, what Scott and Justin were saying, we're starting to adapt based on the current state of models. But you
1:15:54
Unidentified speaker (UNKNOWN): know, if
1:15:55
Unidentified speaker (speaker_2): play forward three more years, maybe the models will be in a completely different place. And then all our plans we would have to restart and redesign our entire current curriculum again. This is why I'm saying that this is a kind of a transition period, and it's hard to uh decide or d know exactly what uh what uh what we should be teaching, given that things are changing
1:16:15
Unidentified speaker (speaker_2): so rapidly. I wish we knew. I like I can't wait until we're kind of at the other end of this transition and we know where the models
1:16:21
Unidentified speaker (UNKNOWN): are
1:16:21
Unidentified speaker (speaker_2): so we can design design things better.
1:16:23
Unidentified speaker (speaker_1): For what it's worth, I'm teaching my students a lot of dynamic programming, like sort of like real meta. dynamic programming, um, you know, sort of around that exact problem, right? Like every I I talk to my students every week about where the technology is and different models of where it not originally intended, sorry, models of where it might go, like models of the future, not not the where the models
1:16:42
Unidentified speaker (speaker_1): models of where the models might go. It's probably a tongue twister, but we're at the end. So
1:16:47
Unidentified speaker (speaker_1): the
1:16:49
Unidentified speaker (speaker_1): you know
1:16:50
Unidentified speaker (UNKNOWN): and
1:16:51
Unidentified speaker (speaker_1): I'm trying very hard to help them think about, you know, sort of how their research, you know, can leverage the tools that exist right now and how it might evolve and change as the tools change, right? Like, you know, like I, you know, and and to impress upon them that this is really a dynamic optimization with a the tremendous uncertainty on
1:17:11
Unidentified speaker (speaker_1): the path of how things change. Um
1:17:16
Unidentified speaker (speaker_1): And then
1:17:19
Unidentified speaker (speaker_1): the other thing, and I said this earlier. I mean, I said in January, like little little did I know in January when I was saying like this is going to be, you know, serious-you current graduate students will be the first generation of scholars that could use LLMs and generative AI tech to do substantive substitutive research tasks. Little did I know that, you know, seven, eight months later, like, you know, you could write any.
1:17:38
Unidentified speaker (speaker_1): end papers end-to-end, and you know, ten months later we'd be doing a podcast about whether the field would even continue to exist.
1:17:46
Unidentified speaker (speaker_1): I really was not expecting any of that.
1:17:52
Unidentified speaker (speaker_1): I have been saying over and over again and throughout this, right? Given like you want to calibrate your aperture
1:18:01
Unidentified speaker (speaker_1): for opportunities to what the state of knowledge is and what your bets about the future of knowledge will look like. Right? Like if you're a if you're a scholar and you're, you know, you this can be a tremendously exciting time because you can think about questions of a depth or complexity or abstraction, or, you know, whatever the thing is that that you
1:18:21
Unidentified speaker (speaker_1): use to shape your questions in a totally different way. But, you know.
1:18:27
Unidentified speaker (speaker_1): if you don't like it, but it's, you know, I've been telling them, like, it's first of all, it's really important to learn all the foundations that you use to build that understanding and perspective. And then you want to take on challenges that that rise to the level of the sort of the new abilities.
1:18:44
Unidentified speaker (UNKNOWN): And
1:18:44
Unidentified speaker (speaker_1): I guess one last point on that in is like, you know, in my classes forever, right? This has just always been a rule in my classes, the
1:18:50
Unidentified speaker (UNKNOWN): the
1:18:50
Unidentified speaker (speaker_1): policy. policy around the use of technology, whatever it is, is that you should always use every tool in existence that you can use in a way that you find productive. Like, you know, open book, open internet, whatever. Like the the objective is to maximize the utilization of tools and to learn how to solve your own optimization around what you should do in your head, what you should do on a computer,
1:19:10
Unidentified speaker (speaker_1): what you should do, like, you know, you know, sort of in in, you know, paired computation with a friend, like whatever.
1:19:15
Unidentified speaker (speaker_1): But my standard rises with the level of tool availability, right? That like, like the, you know, my aspirations for my students get higher and higher and higher every time a new opportunity comes about. And so it's it's incumbent on them to learn how to sort of take those those opportunities, find those opportunities and rise
1:19:35
Unidentified speaker (speaker_1): to them. And I'm here to help and support them to do that. And it's incumbent. on me to teach them how. Right. Like, like that's yeah
1:19:44
Unidentified speaker (speaker_1): that's the objective, right? We when you're out in the world as a researcher, you're gonna have all of this stuff. And so we have to teach people how to maximize and and push the frontiers from that perspective. So it sounds like you guys are all taking a variety of approaches, Scott. You're teaching people to use these tools to to become conversant with them. to apply them uh
1:20:04
Unidentified speaker (speaker_1): constantly, and to up their game with these tools. Ittai and Dan, you both mentioned kind of focusing more on high level concepts and abstractions as uh
1:20:14
Unidentified speaker (UNKNOWN): and communication
1:20:15
Unidentified speaker (speaker_2): skills. Before we close out, I just wanna
1:20:17
Unidentified speaker (UNKNOWN): uh
1:20:18
Unidentified speaker (speaker_2): I just want to ask uh Dan one more question because you you you emphasize that we're in this period of transition right now and that there's uncertainty and maybe we'll have it more. figured out in a few years. But if I'm, you know, a young student who is planning to who was planning to pursue a PhD
1:20:37
Unidentified speaker (speaker_2): in uh, you know, some area of math or cryptography or whatever, what what would you what would you say to them? Because it's not highly encouraging to hear like, you know, oh well, well, it's just like it's kind of messy right now, we'll figure it out. eventually. You
1:20:51
Unidentified speaker (UNKNOWN): know,
1:20:51
Unidentified speaker (speaker_2): if I'm like, I'm about to make a big life decision, you know, I want to decide what I'm gonna do with my life, how I'm gonna spend the next few years. And
1:20:58
Unidentified speaker (UNKNOWN): I
1:20:58
Unidentified speaker (speaker_2): what would you say to that person in in particular? I think it's really important to remember the computer can beat us a chess, but we still enjoy playing chess with one another. Yes, and in s in many ways, um uh math would math and you know theoretical computer science uh would continue to be enjoyable in in that way. I would say like
1:21:18
Unidentified speaker (speaker_2): in in our field, like in theoretical computer computer science, cryptography, there are like real societal questions that we need to solve.
1:21:25
Unidentified speaker (UNKNOWN): Right?
1:21:25
Unidentified speaker (speaker_2): I mean society depends on kind of uh the systems that we develop, like Justin was talking about the post-quantum transition. There's like big open problems that literally the internet is waiting for us to solve. Um and so there there are really important questions to work on. The machines right now are not good enough to solve those questions.
1:21:45
Unidentified speaker (speaker_2): So right now we're in a period where combination of a human and a machine is much better than either one working on their own.
1:21:52
Unidentified speaker (UNKNOWN): Yeah.
1:21:53
Unidentified speaker (speaker_2): Whether that will continue is not clear. But right now there's still a lot for humans to do, to contribute. Um and so in in that sense, you know, it's um uh still
1:22:03
Unidentified speaker (UNKNOWN): uh
1:22:03
Unidentified speaker (speaker_2): qu quite a good quite a good choice to f particularly go into of course the field that we we know and love, the field of the field of cryptography. I'll say one more thing though, that it's it's uh in the thinking more broadly about uh research in general,
1:22:18
Unidentified speaker (UNKNOWN): you know,
1:22:18
Unidentified speaker (speaker_2): research is a human activity. If humans for some reason decide that you know machines are better than us, we don't need to do
1:22:25
Unidentified speaker (UNKNOWN): it,
1:22:25
Unidentified speaker (speaker_2): then the whole research endeavor will die, right? If graduate students don't come in and do get PhDs and get in get uh commit to a life of research, nobody will be do will be doing research and the endeavor will die. So that's actually kind of a an an important point to remember. So there's also uh as a society
1:22:45
Unidentified speaker (speaker_2): we want to make sure uh our students continue to kind of pursue a d to pursue this path uh and uh you know the w work on problems that society needs needs to solve. But I want to end on a kind of a positive note, which is again my basic notes that even though the machines are getting better at s solving the problems that we know
1:23:05
Unidentified speaker (speaker_2): and love, we still enjoy solving these problems. I think Scott's message before was really, really good that uh we still get joy from looking at solutions, we still get joy from explaining solutions to each other, from building on each other's ideas. Um and the fact that the machines can help and run along with us, you know, that's all for
1:23:25
Unidentified speaker (speaker_2): the better.
1:23:27
Unidentified speaker (speaker_1): Well put.
1:23:28
Unidentified speaker (UNKNOWN): Very nice.
1:23:29
Unidentified speaker (speaker_0): I'll just add I don't I don't know how there will no matter how good the machines get at, you know, reasoning and and math, uh there will always be be, you know, judgment calls and value judgments that uh you know I expect humans will want to make themselves. And in order to do that,
1:23:50
Unidentified speaker (speaker_0): uh, we need informed humans. And if you are not trying to understand, you know, the frontier of technology and like all, you know, uh whatever the AIs and humans combined are putting out, then you will not be, you know, in the best position to weigh in on those judgment calls. Um, and you know, I've learned this more and more as I've
1:24:10
Unidentified speaker (speaker_0): moved more and more towards cryptography. Uh, because a lot, you know, there's a lot of you know non-quantifiable judgment calls that um, you know, uh you have to call on all your knowledge, and then even then uh every not everyone's gonna going to agree with you. Um so yeah, so that that's always gonna
1:24:26
Unidentified speaker (UNKNOWN): be a good thing.
1:24:26
Unidentified speaker (speaker_0): going to there. So thought I'd jump in with that.
1:24:29
Unidentified speaker (speaker_2): And this is sort of where uh where you're talking about from the applied side of math, where it kind of rubber meets the road and you actually have to put this stuff into practice. Trevor
1:24:38
Unidentified speaker (UNKNOWN): Burrus
1:24:39
Unidentified speaker (speaker_0): Yeah. Just generally like technology has effects on society, and society has to, you know, control and mitigate and you know uh those effects. And so uh yes, this these judgment calls. happen uh you know not in pure theory land but where like the technology meets society but you know it's all uh spectrum it's not
1:24:59
Unidentified speaker (speaker_0): binary um and you know the theory tools help us under you know make these judgment calls and all of that. So I think there there'll always be a need for people to, you know, understand as far towards the frontier as they can. Um
1:25:12
Unidentified speaker (UNKNOWN): It even
1:25:13
Unidentified speaker (speaker_1): happens in pure theory land when you're trying to understand is this thing truly the explanation for this other thing, right? It's sort of
1:25:20
Unidentified speaker (UNKNOWN): there's
1:25:20
Unidentified speaker (speaker_0): there's like judgment calls and debates about interpretations even in theory land. Yeah. So
1:25:26
Unidentified speaker (speaker_0): I think there'll always be stuff for people to do. And it's just again, it's a it's a hard message to deliver in the face of, you know, so much disruption, um especially to to young people. But uh I do think there's something to hold on to in the in the optimistic view.
1:25:41
Unidentified speaker (UNKNOWN): And
1:25:42
Unidentified speaker (speaker_0): go to
1:25:42
Unidentified speaker (UNKNOWN): Tai.
1:25:43
Unidentified speaker (speaker_3): I mean, one of the huge challenges going forward is how do we make sure that AI is going to stay aligned with
1:25:52
Unidentified speaker (speaker_3): what humans want. And that in particular is going to be important as AI kind of advances even further. And so I think
1:26:00
Unidentified speaker (speaker_3): in pushing the boundary of human understanding. understanding forward is going to be very crucial for our ability, as Justin said, to make good judgments, to have informed values by have having better abstractions that allow us to go closer and closer to this frontier. And so you know I think one of the things I'm excited about
1:26:20
Unidentified speaker (speaker_3): is that
1:26:22
Unidentified speaker (speaker_3): understanding math is actually going to be very important going forward. in kind of making sure that that you know the advance of AI is going to stay aligned with kind of what what humanity wants it to be at.
1:26:36
Unidentified speaker (speaker_1): And maybe let me uh I have no opinion on that, but a a a very practical point is just the more mathematical knowledge that is produced relative to the amount of it that we've digested and consolidated and internalized, going back to
1:26:50
Unidentified speaker (UNKNOWN): this
1:26:50
Unidentified speaker (speaker_1): like you know. billion-page proof hypothesis, like the more math there is to understand, right? You know, sort of if you're looking through your telescope, you see lots and lots of sort of stars and and other sort of celestial bodies. Sorry, I'm not an astronomer, so I'm gonna try to know, you
1:27:04
Unidentified speaker (UNKNOWN): know,
1:27:04
Unidentified speaker (speaker_1): like exoplanets, like in a sort of like, you know, add your add your various celestial objects as as per your preference, the more, you know, the deeper you can sort of see into space. the more there is to discover and to understand. And like I think, you know, sort of wearing my mathematical economist hat, you know, sort of changes in productivity that
1:27:24
Unidentified speaker (speaker_1): create tons more knowledge along, you know, sort of this branch are likely to require tons more mathematicians and sort of mathematician adjacencies to, you know, understand it and build sort of, you know, at least the human knowledge off of it and more and broadly, I I continue to think to to sort of build and map the frontiers of it. So you know
1:27:44
Unidentified speaker (speaker_1): uh
1:27:46
Unidentified speaker (UNKNOWN): I
1:27:46
Unidentified speaker (speaker_1): suspect we will need a lot more mathematicians in the future and and with any luck, you know, sort of the the opportunities to explore will uh will energize some people to to do so. Aaron
1:27:56
Unidentified speaker (UNKNOWN): Ross Powell And
1:27:57
Unidentified speaker (speaker_0): you're saying that if maybe math even but b would be more democratized even
1:28:01
Unidentified speaker (speaker_1): absolutely yeah yeah and and and this we are seeing, right? I'm getting reach outs from lots, you know, I've I've always you know sort of advised a handful of sort of of you know right-tail high school students, uh very, very like I'm I'm very, very lucky to have had the opportunity to do that. Some of the like these are some of the students who've like pushed me to think deeper and harder than than anybody.
1:28:20
Unidentified speaker (speaker_1): Um I'm getting reach outs
1:28:23
Unidentified speaker (UNKNOWN): from
1:28:23
Unidentified speaker (speaker_1): high schoolers and even occasional middle schoolers from all over the world. um, who are doing serious mathematics and at different levels of of their own understanding, but some of them really are like, you know, sort of grokking it and using this tooling to, you know, access the doing of mathematics
1:28:43
Unidentified speaker (speaker_1): independently at an age that I certainly never, you know, would have been able to do. Um
1:28:48
Unidentified speaker (UNKNOWN): and
1:28:48
Unidentified speaker (speaker_1): that's very exciting to me, right? Like you know
1:28:53
Unidentified speaker (speaker_1): any anything that's causing more people to want to spend their weekends like exploring the frontiers of math, I think is uh, you know,
1:28:59
Unidentified speaker (speaker_2): sort of rather
1:29:00
Unidentified speaker (speaker_1): more I I can't make a total statement, but anything that's causing new people to want to spend their weekends exploring the frontiers of math, I think is has has got to be pretty exciting.
1:29:08
Unidentified speaker (speaker_2): And I really like this expansive vision that you've laid out, Scott, about how the frontier really is infinite and uh and makes
1:29:14
Unidentified speaker (UNKNOWN): it.
1:29:15
Unidentified speaker (speaker_2): maybe many sets of
1:29:19
Unidentified speaker (UNKNOWN): we're
1:29:19
Unidentified speaker (speaker_1): talking like you know, the cardinal of the reals to the cardinal of the reels to the e to the grams number to the cardinal of the reels. That's
1:29:25
Unidentified speaker (UNKNOWN): what we're
1:29:25
Unidentified speaker (speaker_1): aiming for here.
1:29:27
Unidentified speaker (speaker_2): Um so I just to bring this conversation all the way back to the top, uh, where you know I introduced this by talking about Andrew Wiles and and his uh his his proof of uh Fermat's last theorem. Um he said something some decades ago that I think is is appropriate here. He said uh the definition of a good mathematical problem is the mathematics
1:29:47
Unidentified speaker (speaker_2): it generates rather than the problem itself. And so I uh I I think maybe that would be a good note to close on. Um that maybe maybe math is infinitely generative and uh there will be there will be lots and lots to do for everybody. who's getting involved. So thank you all for joining. QED.
1:30:06
Unidentified speaker (UNKNOWN): Yeah.
1:30:07
Unidentified speaker (speaker_0): Thank you, Robert. That's a great
1:30:08
Unidentified speaker (UNKNOWN): view.
This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3+sortformer-v2.1 and interpreted by Qwen3.6-35B-A3B-UD-IQ3_S.gguf. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.

More from a16z Frontier show

All insights