Latest / The Last Observer: Science and Beyond / An AI Just Finished Math’s Sponge Problem

Transcript

An AI Just Finished Math’s Sponge Problem · The Last Observer: Science and Beyond

• 0:00
  Speaker 1: Welcome to The Last Observer, a journey through science, history,
• 0:04
  Speaker 1: and the universe's greatest mysteries. From ancient enigmas to the
• 0:08
  Speaker 1: deepest questions at the frontiers of reality, we explore the discoveries, ideas,
• 0:13
  Speaker 1: and mysteries that challenge what we think we know. Stay curious,
• 0:18
  Speaker 1: keep questioning, and let's search for the truth beyond the known.
• 0:27
  Speaker 2: Welcome to The Last Observer Podcast. Today we will talk
• 0:31
  Speaker 2: about a strange moment in mathematics, when an artificial intelligence
• 0:35
  Speaker 2: proved a decades-old idea about how randomness suddenly turns into connection.
• 0:41
  Speaker 3: I keep turning that over. Not a new gadget, a proof.
• 0:45
  Speaker 3: A machine just closed a problem that a Fields medalist
• 0:48
  Speaker 3: said might be worth another. Fields medal.
• 0:52
  Speaker 2: And the timing is almost unfair. On August 30, 2026,
• 0:57
  Speaker 2: Hugo Dumanil Copin wrote that it was only a matter
• 0:59
  Speaker 2: of time before the famous conjecture in his field fell
• 1:03
  Speaker 2: to the bulldozers. Days later, Anthropic released a proof generated
• 1:07
  Speaker 2: by a large language model.
• 1:09
  Speaker 3: Dumanil Copin already has a Fields Medal, from 2022, for
• 1:14
  Speaker 3: work on phase transitions. He tried this one himself and
• 1:18
  Speaker 3: did not get it. Then he basically predicted the machine
• 1:21
  Speaker 3: would arrive first.
• 1:23
  Speaker 2: Scientific American published the account on September 30th. Manon Bischoff
• 1:28
  Speaker 2: reported it, originally for Spectrum der Wissenschaft. The mathematician she
• 1:33
  Speaker 2: spoke with sound genuinely split. Joy that the thing is
• 1:36
  Speaker 2: finally settled. A quieter disappointment that the last step was
• 1:40
  Speaker 2: not human.
• 1:42
  Speaker 3: Before we get lost in the feelings, I want the picture.
• 1:45
  Speaker 3: What is the actual puzzle?
• 1:47
  Speaker 2: It lives in percolation theory. That sounds like coffee. And
• 1:51
  Speaker 2: the coffee image is honestly the best way in. Picture
• 1:54
  Speaker 2: a sponge, or a filter, or a cliff of porous rock.
• 1:58
  Speaker 2: At what point are there enough open holes that water
• 2:01
  Speaker 2: can find a path all the way through?
• 2:03
  Speaker 3: So it is not about one crack. It is about
• 2:06
  Speaker 3: whether a bunch of small random openings line up into
• 2:10
  Speaker 3: one long road.
• 2:11
  Speaker 2: Exactly. In 1957, Simon Broadbent and John Hammersley formalized that.
• 2:18
  Speaker 2: They turned the material into a network. Nodes are holes.
• 2:22
  Speaker 2: Edges are the little passages between them. Each passage is
• 2:25
  Speaker 2: open or closed with some probability.
• 2:28
  Speaker 3: A city of pipes, and each pipe is rolled like
• 2:31
  Speaker 3: a die. If the chance a pipe is open is tiny,
• 2:34
  Speaker 3: you get puddles, isolated clusters. Nothing reaches forever.
• 2:39
  Speaker 2: Raise that chance, and at some critical value, the character
• 2:42
  Speaker 2: of the whole network changes. Above it, there is a
• 2:46
  Speaker 2: real chance of an infinite connected cluster. Below it, that
• 2:50
  Speaker 2: chance is zero. Mathematicians call the threshold PC.
• 2:55
  Speaker 3: Infinite is the part that always snags me. We are
• 2:58
  Speaker 3: not talking about a big sponge on a kitchen counter.
• 3:01
  Speaker 3: We are talking about an endless grid, the kind of
• 3:04
  Speaker 3: object only a proof can hold.
• 3:06
  Speaker 2: Right. And the Holy Grail was not even the exact
• 3:09
  Speaker 2: number for that threshold. Exact values are brutal. Harry Keston
• 3:14
  Speaker 2: found one famous case in 1980. On a two-dimensional square lattice,
• 3:19
  Speaker 2: the critical probability for bond percolation is one-half.
• 3:23
  Speaker 3: One-half feels clean, like the universe left the answer on
• 3:26
  Speaker 3: the table.
• 3:28
  Speaker 2: It is clean, and it is rare. Most geometries do
• 3:31
  Speaker 2: not hand you a nice fraction. The deeper question was
• 3:35
  Speaker 2: about the shape of the change. When you cross the threshold,
• 3:38
  Speaker 2: does the chance of belonging to an infinite cluster jump,
• 3:42
  Speaker 2: or does it rise smoothly from zero?
• 3:44
  Speaker 3: A cliff or a ramp?
• 3:46
  Speaker 1: Yes.
• 3:47
  Speaker 2: In one and two dimensions, for the square lattice, people
• 3:51
  Speaker 2: already knew the transition is continuous. The probability starts at
• 3:54
  Speaker 2: zero and climbs. No sudden leap. In very high dimensions,
• 3:59
  Speaker 2: 11 and above, the same continuous picture holds, partly because
• 4:04
  Speaker 2: each point has so many neighbors that averaging becomes kinder.
• 4:09
  Speaker 3: So the missing middle was dimensions 3 through 10, the
• 4:12
  Speaker 3: worlds we actually live near. Our space is 3. Four
• 4:16
  Speaker 3: shows up all over physics, and those were the ones
• 4:19
  Speaker 3: still open.
• 4:20
  Speaker 2: That is the conjecture Anthropic's model addressed. In those in-between dimensions,
• 4:25
  Speaker 2: the transition is continuous too, what the community had suspected
• 4:29
  Speaker 2: for decades and had not proved.
• 4:32
  Speaker 3: I want to sit with how ordinary the metaphor is.
• 4:35
  Speaker 3: A wet sponge, a coffee filter, forest fires jumping between trees,
• 4:40
  Speaker 3: disease moving through a crowd if each contact is a
• 4:43
  Speaker 3: coin flip. the same skeleton of a question.
• 4:47
  Speaker 2: And the proof is not a simulation of one sponge.
• 4:50
  Speaker 2: It is a statement about every large enough network of
• 4:53
  Speaker 2: that type. That is why mathematicians care. A computer can
• 4:57
  Speaker 2: run a million grids and still miss the infinite case.
• 5:00
  Speaker 2: A proof has to cover the thing you cannot simulate.
• 5:04
  Speaker 3: Which makes the ambivalence sharper. Benedict Giano, at the Technical
• 5:08
  Speaker 3: University of Braunschweig, told Scientific American that if a person
• 5:13
  Speaker 3: solved this, they would probably receive a Fields Medal. Then
• 5:17
  Speaker 3: the solver was a model.
• 5:19
  Speaker 2: He described the feeling as joy and disillusionment in the
• 5:22
  Speaker 2: same breath. I think that is the honest reaction. The
• 5:26
  Speaker 2: landscape of the problem got clearer. The story of who
• 5:29
  Speaker 2: gets to be the one who clarifies it got stranger.
• 5:33
  Speaker 3: I am still not sure I believe a language model
• 5:35
  Speaker 3: can carry a proof that careful people failed to finish.
• 5:39
  Speaker 3: We should look at what was actually released. and what
• 5:42
  Speaker 3: continuous really means when you cannot draw the ramp.
• 5:46
  Speaker 2: Good, because the claim is specific and the caveats matter.
• 5:50
  Speaker 2: The model did not hand over a napkin sketch. Anthropic
• 5:53
  Speaker 2: posted a formal proof, the kind a proof assistant can
• 5:57
  Speaker 2: check line by line. That is a different standard than
• 6:00
  Speaker 2: a confident paragraph.
• 6:02
  Speaker 3: Then let us stay with the sponge a little longer,
• 6:05
  Speaker 3: and then open the file. All right, I made coffee
• 6:08
  Speaker 3: while you were talking about filters, which feels on theme
• 6:11
  Speaker 3: and also slightly mocking the math.
• 6:13
  Speaker 2: The filter is doing percolation right now. Water finds paths
• 6:17
  Speaker 2: through paper because enough fibers leave gaps. If the paper
• 6:21
  Speaker 2: were tighter, the pot would just sit there.
• 6:24
  Speaker 3: So in the mathematical version, each possible gap is open
• 6:28
  Speaker 3: with probability p. Theta of p is the chance that
• 6:32
  Speaker 3: the origin, some starting point, belongs to an infinite open cluster.
• 6:37
  Speaker 2: Below the critical p, theta is zero. Above it, theta
• 6:41
  Speaker 2: becomes positive. The fight was about how theta leaves zero.
• 6:45
  Speaker 2: A continuous transition means it lifts off gradually. A discontinuous
• 6:49
  Speaker 2: one would mean it snaps to a positive number the
• 6:52
  Speaker 2: moment you cross.
• 6:54
  Speaker 3: Why would anyone expect a snap? It feels unnatural.
• 6:57
  Speaker 2: Because some phase transitions do snap. Think of ice and
• 7:01
  Speaker 2: water in a simplified model, or certain magnets. Order can
• 7:05
  Speaker 2: appear with a jump. Percolation had reasons to be continuous
• 7:09
  Speaker 2: in the easy dimensions, and reasons to worry in the
• 7:12
  Speaker 2: middle ones. The middle is where intuition gets expensive.
• 7:16
  Speaker 3: Keston's one-half result is from 1980. People have had 46
• 7:20
  Speaker 3: years with that lighthouse and still could not finish the
• 7:23
  Speaker 3: middle dimensions. That is a long stare at one question.
• 7:28
  Speaker 2: Dupinil Kopin used that image himself. A mathematical question is
• 7:32
  Speaker 2: a lighthouse in the night. It illuminates and guides, aesthetic
• 7:36
  Speaker 2: and scientific. The point of the conjecture was not only
• 7:40
  Speaker 2: the final line. It organized decades of partial maps.
• 7:44
  Speaker 3: That essay is dated August 30th. He writes that the
• 7:47
  Speaker 3: most famous conjecture in the field will fall to the bulldozers.
• 7:51
  Speaker 3: He is not celebrating. He is bracing.
• 7:54
  Speaker 2: And he is specific about what might be lost. Understanding,
• 7:57
  Speaker 2: for him, is not only theorems. It is the failed routes.
• 8:02
  Speaker 2: His own attempts on this problem produced other insights. If
• 8:05
  Speaker 2: a model jumps to the end, those side paths might
• 8:08
  Speaker 2: never get walked.
• 8:10
  Speaker 3: I feel that in a smaller way. When I look
• 8:12
  Speaker 3: up a route and the app just draws the line,
• 8:15
  Speaker 3: I arrive and I do not know the neighborhood. The
• 8:18
  Speaker 3: proof might be like that, correct and oddly uninhabited.
• 8:22
  Speaker 2: Janel makes a related point. Mathematicians may have to stop
• 8:26
  Speaker 2: basing their confidence on being the ones who prove conjectures.
• 8:30
  Speaker 2: The work also includes forming the conjectures, translating proofs into
• 8:34
  Speaker 2: human language, and folding them into textbooks.
• 8:38
  Speaker 3: Forming the questions still sounds deeply human. The sponge picture
• 8:42
  Speaker 3: did not arrive from a loss function. Broadbent and Hammersley
• 8:45
  Speaker 3: were watching liquids in porous media.
• 8:48
  Speaker 2: Yes, and the application stayed physical. Oil moving through rock,
• 8:52
  Speaker 2: a wildfire finding connected dry patches, even some models of
• 8:56
  Speaker 2: network robustness where nodes fail at random. And you ask
• 9:00
  Speaker 2: whether a giant component survives. Same threshold logic.
• 9:05
  Speaker 3: So a continuous transition is not a decorative fact. It
• 9:08
  Speaker 3: tells you the change is tunable. You can be just
• 9:11
  Speaker 3: above critical and have a barely-there infinite cluster, thin and fragile,
• 9:16
  Speaker 3: rather than a sudden flood.
• 9:18
  Speaker 2: Near criticality, clusters have this wild, scale-free character. Pieces of
• 9:23
  Speaker 2: every size. That is why the proofs are hard. The
• 9:26
  Speaker 2: system is balanced on a knife edge, and ordinary averaging fails.
• 9:31
  Speaker 3: High dimensions rescue you because neighbors multiply. Eleven and up,
• 9:35
  Speaker 3: there is a result that the transition is continuous. The
• 9:38
  Speaker 3: crowd is large enough that fluctuations calm down.
• 9:42
  Speaker 2: Three through ten refused that comfort. Two connected to be trivial,
• 9:46
  Speaker 2: not connected enough for the high dimension shortcuts. A classic
• 9:50
  Speaker 2: stuck place in mathematical physics.
• 9:53
  Speaker 3: And this is where I want to be careful with
• 9:55
  Speaker 3: the headline. Solved by an AI can mean a chatbot
• 9:58
  Speaker 3: guessed a paragraph. What Anthropic released is a formalization, a
• 10:03
  Speaker 3: proof written so a computer checker can accept every step.
• 10:07
  Speaker 2: That distinction matters. Earlier this year, a clawed system formalized
• 10:11
  Speaker 2: Fermat's last theorem. Millions of lines, checked by a proof assistant.
• 10:16
  Speaker 2: Formalizing a known proof is already enormous. Producing a new
• 10:21
  Speaker 2: proof of an open conjecture is a different claim, and
• 10:24
  Speaker 2: it needs the same checking standard.
• 10:26
  Speaker 3: So the honest version is this. A language model proposed
• 10:30
  Speaker 3: a structured argument, and the argument was put in a
• 10:33
  Speaker 3: form that can be mechanically verified. The credit, if we
• 10:36
  Speaker 3: still use that word, is shared with the checker.
• 10:39
  Speaker 2: I like that better. The romance of the lone genius
• 10:42
  Speaker 2: was already a myth. Now the room just got more crowded.
• 10:46
  Speaker 2: a model, a proof assistant, and a few humans who
• 10:49
  Speaker 2: still have to decide whether the statement matches the conjecture
• 10:52
  Speaker 2: they actually cared about.
• 10:54
  Speaker 3: That last part is easy to skip. A checker can
• 10:57
  Speaker 3: confirm that the lines follow. It cannot, by itself, tell
• 11:01
  Speaker 3: you that you formalized the right question. Someone has to
• 11:04
  Speaker 3: look at the sponge and say, yes, this is the
• 11:07
  Speaker 3: transition we meant.
• 11:09
  Speaker 2: Johnnell's ambivalence lives right there. The conjecture fell immediately. The
• 11:13
  Speaker 2: human job moved from the final blow to the framing,
• 11:16
  Speaker 2: the audit, and the explanation. Whether that feels like a
• 11:20
  Speaker 2: demotion depends on what you thought mathematics was for. Let
• 11:24
  Speaker 2: me try to say what the result actually asserts, without
• 11:27
  Speaker 2: hiding inside notation.
• 11:29
  Speaker 3: Please, I can hold one formula if you promise to
• 11:32
  Speaker 3: translate it.
• 11:33
  Speaker 2: On the square lattice, in dimensions 3, 4, 5, and
• 11:36
  Speaker 2: on up through 10, the chance that a given point
• 11:41
  Speaker 2: sits in an infinite open cluster is zero at the
• 11:44
  Speaker 2: critical probability, and it rises continuously once you move above it.
• 11:49
  Speaker 2: No jump. The passage from isolated pockets to a system-spanning
• 11:53
  Speaker 2: connection is a ramp.
• 11:55
  Speaker 3: And that was the expected answer. So the surprise is
• 11:58
  Speaker 3: not the shape of the ramp. The surprise is that
• 12:01
  Speaker 3: the ramp now has a proof, and the proof came
• 12:04
  Speaker 3: out of a model within days of a leading human
• 12:06
  Speaker 3: saying the machines were close.
• 12:08
  Speaker 2: Yes. Scientific American is careful on the tone. It says
• 12:12
  Speaker 2: the model has now shown the transition is continuous, confirming
• 12:16
  Speaker 2: the suspicion. It points to Anthropic's formal math repository as
• 12:21
  Speaker 2: the release. This is not a peer-reviewed journal article in
• 12:24
  Speaker 2: the old sense. It is a checked artifact, posted where
• 12:28
  Speaker 2: other formal proofs live.
• 12:30
  Speaker 3: I want to press on that. If I were refereeing
• 12:32
  Speaker 3: the headline... I would ask whether independent mathematicians have read
• 12:36
  Speaker 3: the formalization and agreed it matches the classical conjecture. A
• 12:40
  Speaker 3: green check from a proof assistant is strong. Community reading
• 12:44
  Speaker 3: is the other half.
• 12:45
  Speaker 1: Fair.
• 12:46
  Speaker 2: The reporting quotes Jonnell reacting to the result as a
• 12:49
  Speaker 2: completed proof, with that mix of joy and disillusionment. That
• 12:54
  Speaker 2: is expert reaction, not a press release. Still, early days.
• 12:58
  Speaker 2: Formal libraries move fast, and people will pick at definitions.
• 13:02
  Speaker 3: What would a pick look like? A missing assumption about
• 13:05
  Speaker 3: how the lattice is built? A proof that covers a
• 13:08
  Speaker 3: close cousin of the conjecture rather than the exact statement?
• 13:12
  Speaker 2: Those are the usual ghosts. In formal math, the danger
• 13:15
  Speaker 2: is a subtle mismatch between the human sentence and the
• 13:18
  Speaker 2: encoded sentence. You can prove a true thing that is
• 13:21
  Speaker 2: one adjective away from the famous thing. That is why
• 13:25
  Speaker 2: Dumanil Kopan's lighthouse metaphor still matters. someone has to recognize
• 13:29
  Speaker 2: the light.
• 13:30
  Speaker 3: Assume the match is good because the people closest to
• 13:33
  Speaker 3: it are talking as if it is. What changes in
• 13:36
  Speaker 3: the science?
• 13:38
  Speaker 2: A closed door in the middle dimensions. Techniques that were
• 13:41
  Speaker 2: stuck can be reused on nearby models. Continuity is a platform.
• 13:46
  Speaker 2: Once you know the transition does not jump, you can
• 13:49
  Speaker 2: ask sharper questions about the slope, the size of clusters,
• 13:53
  Speaker 2: the way correlations decay.
• 13:55
  Speaker 3: The next questions, not the trophy question.
• 13:59
  Speaker 2: And those next questions are where Jonnell thinks the human
• 14:02
  Speaker 2: role concentrates, posing them, noticing that a wildfire model and
• 14:07
  Speaker 2: a magnet model are secretly the same shape, writing the
• 14:10
  Speaker 2: chapter a student can enter.
• 14:12
  Speaker 3: There is a social piece I cannot shake. Fields medals
• 14:16
  Speaker 3: are awarded to people under 40, often for proofs. If
• 14:20
  Speaker 3: the decisive proofs start arriving from labs, the prize either
• 14:24
  Speaker 3: changes or it starts honoring the questions and the translations.
• 14:28
  Speaker 2: Jono says it plainly. Whether a human being will ever
• 14:31
  Speaker 2: make it onto such a list again is questionable. If
• 14:34
  Speaker 2: the list is only for proving theorems, he does not
• 14:37
  Speaker 2: sound thrilled. He sounds like someone rearranging the furniture while
• 14:41
  Speaker 2: the house is still his.
• 14:42
  Speaker 3: I do not think the prize disappears. I think the
• 14:45
  Speaker 3: citation gets longer. Proved with a model, audited by a group,
• 14:50
  Speaker 3: explained by a person. The person is still in the sentence.
• 14:55
  Speaker 2: Maybe. There is also a risk of a new monopoly.
• 14:58
  Speaker 2: Formal proof at this scale needs compute, tooling, and teams.
• 15:02
  Speaker 2: A university seminar cannot casually match a lab that treats
• 15:06
  Speaker 2: conjectures like nightly jobs. The bulldozers Domenal Copine named are
• 15:11
  Speaker 2: not evenly distributed.
• 15:13
  Speaker 3: That is the part that feels political, even though the
• 15:15
  Speaker 3: theorem is not. Who gets to point the bulldozer, and
• 15:19
  Speaker 3: which conjectures are worth the electricity?
• 15:22
  Speaker 2: Percolation earned it because it sits under so many other stories.
• 15:26
  Speaker 2: It is a clean skeleton for emergence. Many small random events,
• 15:30
  Speaker 2: and then, past a line, a structure that none of
• 15:34
  Speaker 2: the events intended.
• 15:35
  Speaker 3: I like that you said intended. The pipes do not
• 15:39
  Speaker 3: plan the infinite cluster. The cluster is what the probabilities allow.
• 15:43
  Speaker 3: There is something calming in that, and something eerie, when
• 15:46
  Speaker 3: the thing doing the allowing is also a model trained
• 15:49
  Speaker 3: on our sentences.
• 15:51
  Speaker 2: The model did not percolate in the physical sense. It
• 15:54
  Speaker 2: searched a space of arguments. Still, the rhyme is hard
• 15:58
  Speaker 2: to ignore. Connection appearing where separate attempts had stalled.
• 16:02
  Speaker 3: Dumanil Kopin worried that setbacks are part of understanding. I
• 16:06
  Speaker 3: want to keep that. A proof without the scar tissue
• 16:08
  Speaker 3: of failed proofs is a destination without a journey. Useful,
• 16:12
  Speaker 3: a little lonely.
• 16:14
  Speaker 2: The archive of failed human attempts does not vanish because
• 16:18
  Speaker 2: a model finished. Those papers are still the map of
• 16:21
  Speaker 2: the neighborhood. The new artifact is a door at the
• 16:24
  Speaker 2: end of the street. You can use the door and
• 16:26
  Speaker 2: still walk the street.
• 16:28
  Speaker 3: If students are taught only the door, the street goes dark.
• 16:31
  Speaker 3: That is a curriculum problem, not a theorem problem.
• 16:35
  Speaker 2: Agreed. And it is the kind of problem a podcast
• 16:38
  Speaker 2: can actually touch, because it is about what we choose
• 16:41
  Speaker 2: to narrate. The checked lines matter. So does the reason
• 16:45
  Speaker 2: a person cared about a sponge in 1957.
• 16:48
  Speaker 3: I keep coming back to the coffee. If the transition
• 16:52
  Speaker 3: is continuous, then just past the threshold the paths exist,
• 16:56
  Speaker 3: but they are delicate. A. A few more closed fibers
• 17:00
  Speaker 3: and the pot stalls. That picture now has a theorem
• 17:03
  Speaker 3: behind it in the dimensions we kept failing.
• 17:06
  Speaker 2: Delicate is the right word. Critical systems are famous for
• 17:10
  Speaker 2: being sensitive. A forest at the edge of a spreading fire.
• 17:14
  Speaker 2: A rock barely permeable to oil. The math says the
• 17:18
  Speaker 2: change is not a switch. It is a steepening slope.
• 17:21
  Speaker 2: Knowing that slope is continuous lets engineers and physicists trust
• 17:26
  Speaker 2: a whole family of approximations.
• 17:29
  Speaker 3: So the impact is quiet. No new device ships tomorrow.
• 17:33
  Speaker 3: A class of models gets a foundation it was borrowing
• 17:36
  Speaker 3: on credit.
• 17:37
  Speaker 2: Quiet impact is still impact, and the cultural impact is louder.
• 17:42
  Speaker 2: A Fields medalist named the bulldozer, and the bulldozer arrived
• 17:45
  Speaker 2: on something like his schedule. That is a data point
• 17:49
  Speaker 2: about the speed of formal AI, not a prophecy about
• 17:52
  Speaker 2: every open problem.
• 17:54
  Speaker 3: I want that caveat said out loud. One conjecture is
• 17:58
  Speaker 3: not the end of human proof. It is evidence that
• 18:01
  Speaker 3: the queue of formalizable problems just got shorter for groups
• 18:05
  Speaker 3: that can run these systems.
• 18:07
  Speaker 2: And evidence that prediction inside the field was good. Dumanil
• 18:11
  Speaker 2: Kopin was not surprised by the direction. He was worried
• 18:15
  Speaker 2: about the texture of the work that would remain.
• 18:18
  Speaker 3: Texture might be the treasure. The side theorems he found
• 18:22
  Speaker 3: while failing. The essays. The choice of which lighthouse to
• 18:25
  Speaker 3: steer toward. A model can hit a stated target. Choosing
• 18:29
  Speaker 3: targets that matter is still a taste.
• 18:32
  Speaker 2: Taste trained on the world. Broadbent and Hammersley were not
• 18:36
  Speaker 2: optimizing a benchmark. They were watching liquid find a way.
• 18:40
  Speaker 2: The conjecture inherited that attention.
• 18:43
  Speaker 3: If I had to tell a friend the story in
• 18:45
  Speaker 3: one breath, I would say this. For decades, mathematicians knew
• 18:49
  Speaker 3: how random openings become infinite connection in easy dimensions, and
• 18:54
  Speaker 3: not in the middle ones. An AI system just supplied
• 18:57
  Speaker 3: a checked proof that the middle behaves smoothly, days after
• 19:01
  Speaker 3: a medal-winning mathematician said the machines would get there first.
• 19:04
  Speaker 2: And the people who spent their lives on it are glad,
• 19:07
  Speaker 2: and a little bereft. Both reactions fit in the same cup.
• 19:12
  Speaker 3: Bereft is strong. I might say disoriented. The mountain is
• 19:15
  Speaker 3: still there, The flag was planted by a different kind
• 19:18
  Speaker 3: of expedition.
• 19:20
  Speaker 2: Disoriented, then. Johnnel's question about future medals is really a
• 19:24
  Speaker 2: question about recognition. If we only applaud the final inference,
• 19:29
  Speaker 2: we will applaud labs. If we applaud the question, the audit,
• 19:33
  Speaker 2: and the explanation, the room stays mixed.
• 19:36
  Speaker 3: I will applaud the audit. Someone has to say the
• 19:39
  Speaker 3: formal statement is the sponge we meant. That sentence is
• 19:43
  Speaker 3: small and load-bearing.
• 19:45
  Speaker 2: It is. and it is a good place to leave
• 19:47
  Speaker 2: the doubt. Not doubt that a continuous transition is the
• 19:50
  Speaker 2: right picture. Doubt, healthy doubt, that every green check is
• 19:55
  Speaker 2: the conjecture in the headline. Read the formalization. Read the
• 19:59
  Speaker 2: objections if they come. That is the scientific posture, even
• 20:03
  Speaker 2: when the author is a model.
• 20:05
  Speaker 3: I also want the older papers kept in the story.
• 20:08
  Speaker 3: Keston in 1980. the high-dimension work, Domeno Copan's essay on
• 20:13
  Speaker 3: August 30th. The proof is a chapter, not a replacement
• 20:17
  Speaker 3: for the book.
• 20:18
  Speaker 1: Yes.
• 20:19
  Speaker 2: Emergence did not get less mysterious because we pinned one transition.
• 20:23
  Speaker 2: We learned that in dimensions 3 through 10, infinite connection
• 20:27
  Speaker 2: does not burst into being. It gathers. That is a
• 20:31
  Speaker 2: beautiful constraint on how order can appear.
• 20:34
  Speaker 3: Gathers. I can live with that verb. Random pipes, and
• 20:39
  Speaker 3: then a path that has no author, except the probabilities.
• 20:42
  Speaker 2: And now a second author, of a sort, for the
• 20:45
  Speaker 2: argument that the path must exist in that smooth way.
• 20:49
  Speaker 2: We can hold both without forcing a moral.
• 20:52
  Speaker 3: The moral I will take home is modest. Ask better
• 20:55
  Speaker 3: questions than you can finish. Some of them will wait
• 20:58
  Speaker 3: for tools you do not have yet. Some of the
• 21:01
  Speaker 3: waiting is the understanding.
• 21:03
  Speaker 2: That sounds like Dumanil Kopan, translated into a kitchen. I
• 21:07
  Speaker 2: will take the sponge. The next time water finds its
• 21:10
  Speaker 2: way through, I will remember there is a threshold, and
• 21:13
  Speaker 2: that the crossing can be gentle, even when the proof
• 21:16
  Speaker 2: was not.
• 21:17
  Speaker 3: Gentle for the water, abrupt for the mathematicians. Both can
• 21:21
  Speaker 3: be true.
• 21:22
  Speaker 2: Both are true on the evidence we have today. A
• 21:25
  Speaker 2: holy grail of probability theory has a formal proof, posted
• 21:29
  Speaker 2: by Anthropic and reported by Scientific American on September 30, 2026.
• 21:33
  Speaker 2: The transition is continuous. The community is still deciding what
• 21:39
  Speaker 2: a proof is worth when the hand that wrote it
• 21:41
  Speaker 2: is not a hand.
• 21:43
  Speaker 3: And we get to sit with that, which is still
• 21:45
  Speaker 3: a human job. Thanks for walking the middle dimensions with me.
• 21:50
  Speaker 2: Thanks for making the coffee that proved the metaphor. Until
• 21:53
  Speaker 2: the next strange result, keep an eye on the small
• 21:56
  Speaker 2: random things. Sometimes they connect.