On the second day, there are two gaps: “between nothing and zero” and “between zero and nothing”. Two numbers spawn in those two gaps. Call them –1 and 1.
Got lost here. I think I'm officially too dumb for math.
Plot twist: they were both sold on early investment in companies that survived the .com bust. Now they’re VCs that everybody worships as business geniuses even though they’re just lucky idiots, and the sycophantic chatbots finally let them feel as smart as everyone says they are, and a whole bunch of hype-drunk fans are feeding into it.
I think it's just we don't have a way to say/write these numbers. So you make up a way to write them (-1 and 1) and continue.
The numbers don't matter and you could replace -1 and 1 with anything. It's just easier to begin your new fake number at - 1 and 1. Because position does matter.
There has to be some procedure how to come up with "new" numbers though, if you want to have more in the end than just a fancy binary tree - in particular if you want to map your "fake numbers" to the reals, infinity, etc.
This procedure is enough. If you define addition and other operations in a certain way (as Conway did), it turns out that on the omega-th day (i.e. after initial infinite steps), all reals will be born.
Basically what I take away is that we're inventing a new number system from scratch. So we're not "proving" that 1 is a number between 0 and the empty set. We're defining it as such, and it just so happens that a number system defined this way works out in convergent ways with other mathematics.
(Apparently I was extremely unclear with this text. For clarity: if you want to actually understand surreal numbers, go and read On Numbers and Games, by Conway, which is a delightful book; or get an LLM to talk you through Wikipedia. Original text follows.)
It’s a terrible explanation. A surreal number is defined as a pair of sets of surreal numbers (where you fiddle around the recursion in that definition by defining them in waves, so strictly speaking you’re defining “the surreal numbers born at time T” for each individual T given access to the surreal numbers born at all earlier times, and then you “take the union across all times”, scare quotes because there are too many times for this to result in a set). Zero is a surreal number but the LLM is using the word “zero” to mean “the set containing just the surreal number 0”; “nothing” here is the LLM’s obtuse word for the empty set. Wikipedia may actually be easier to follow.
I mean, I was intending to supply the words that would link the LLM’s explanation to a more normal one, not to explain it; apparently that was extremely unclear. An actual explanation is much longer, as indeed I attempted to indicate by pointing to Wikipedia and saying that it might be more clear.
“Doing better than a totally useless explanation in fewer characters” is in general impossible, of course, eg if the first explanation has only one character.
LLM didn't write anything in my post; these are all my words and my choices. Conway himself described surreal generation in short like this in ONAG:
We may say that Cantor was only interested in moving ever rightwards, whereas Dedekind stopped to fill in the gaps, so that R was always empty for Cantor, never empty for Dedekind. It is remarkable that by dropping these restrictions we obtain a theory that is both more general and more easy to work with.
This is precisely the intuition I present to the reader of the article. I am relying on visual aid (concretely, the ordered number line) to imply the machinery explicit in the actual recursive definition. The intended reader of this article is not a mathematician, and I think intuition is vastly more important here.
And I don't think I'm conflating 0 with {0} as you claim. When I say zero is "between nothing and nothing", I mean 0 := {|}. When I say one is "between zero and nothing", I mean 1 := {0|}. When I say 1/2 is "between 0 and 1", I mean 1/2 := {0|1}. And so on. I elide "the simplest number" because I am already going in the order of simplicity. I do not need to explain that alternative spellings like 1/2 = {0.2 | 1} are valid because it is not relevant to establishing the mental model of birthdays.
For the finite cases in my explanation, I do not need to explain that the left and the right parts form sets because I only ever need at most one surreal on either side to define the next generation. I also do not need to state the left/right order condition because it is already visually implied by the picture. For the same reason, I do not need to explicitly quantify over the set of earlier-born surreals, since in these finite cases, if we go birthday by birthday, each next day's surreals are definable via the numbers already constructed by the previous day.
I agree that these finite examples don't spell out how to handle infinitely many bounds at the omega-th day, which is where I believe the illustration embedded below is more helpful. I still think "a gap beyond 0, 1, 2, 3, ... with nothing on the right" is a useful intuition when we get there.
I'm hoping someone develops an interactive tutor that can teach any subject to any depth.
The tutor should optimize its pedagogy. It should use online RL to adapt to a learner's ideal learning style, model what the student understands and to what degree, and understand what the gaps and next steps are.
Broadly, universities are too expensive, inequitable, suboptimal, not portable, and slow. There's a huge amount of room for improvement.
Universities are great for networking, starting projects with other students (not the ones professors mandate), and learning lab sciences. In research, they're great for institutional knowledge, having a community of peers, getting guidance from research advisors, having real equipment and funding, etc. But there's a great need for AI tools to accelerate learning outside of that setting.
Anecdotally, I'm a working adult. I'm not going to waste time in college again. I need this for me.
I think he just wrote it in a confusing way. The quote before says:
(crucially, “to the left of all” and “to the right of all” also count as “gaps”)
So there are two "nothings" here, left of "all" - i.e. the zero - and right of it.
Though I'm not quite sure how you'd get infinite or irrational numbers by this procedure. Wouldn't you simply get the rational numbers by this?
(unless the "put a number" step is doing more work here than it seems. He doesn't really say which number to put there. In the examples, he mostly did "new number = (left number + right number) / 2", with special cases if any number is "nothing" - but he never actually wrote what the rules are here.)
Yeah I'm curious about the difference between "nothing" and "0", but I just decided to roll with it. Until the Greek letters made my head start to spin as they usually do.
I don't think this is your fault; the description isn't very explicit. Let me try to do a bit better. (I'll also try to go somewhat further, and you should not be discouraged if at some point it stops making sense.)
You can think of the "surreal numbers" as being built up step by step. We start out with no numbers at all, and then we repeatedly do a construction that makes some new numbers.
A surreal number is made from two sets of (pre-existing) surreal numbers. We typically call them L and R, for "left" and "right", and sometimes write it as L|R or {L|R} or something like that. The "left" numbers have to be smaller than the "right" numbers. The resulting number will turn out to be, in a certain sense, the "simplest" number in between all the left numbers and all the right numbers.
Now, as I said, we start out with no numbers at all. It might seem like that gives us no way to proceed, but it does: even given no numbers at all, we can still make a set of numbers, namely the empty set! So we can use that for both L and R, getting ∅|∅. Empty sets on both sides. We call this 0, and it will turn out to behave in the way you'd expect the number 0 to behave.
Now we suddenly have another set available, namely {0}, the set containing only zero. Which means that instead of being able to make one number, maybe we can make four: ∅|∅, ∅|{0}, {0}|∅, {0}|{0}. The first of these we already knew about. The last isn't actually admissible -- remember that the "left" numbers have to be smaller than the "right" numbers, which is "vacuously" true when one of those sets is empty (it means "if you have a number x in the left set, and a number y in the right set, then x<y", and if there are no numbers in the left set or no numbers in the right set then that's trivially true) but isn't true when both sets contain 0 because 0<0 is false.
So actually we get two new numbers: ∅|{0} and {0}|∅. The first fits into what OP calls the gap "between nothing and zero". The second first into what OP calls "the gap between zero and nothing". In both cases, "zero" means a number and "nothing" means a space where we don't yet have any numbers.
The number ∅|{0} is called -1 (it has to lie to the left of 0, and there's no constraint on its left, and -1 is "the simplest number less than 0") and the number {0}|∅ is called +1 (it has to lie to the right of 0, and there's no constraint on its right, and +1 is "the simplest number greater than 0").
I should explicitly acknowledge that I haven't defined what "less than" and "greater than" actually mean for these numbers, nor anything else about how they relate to one another that could possibly justify giving these things the specific names 0, -1, and +1. But there are definitions for "less than" and "greater than" and "plus" and "minus" and so forth, and the whole thing does turn out to work very nicely.
Anyway, once we've got these numbers we have eight possible sets that can go on the left or on the right. The requirement for left-things to be smaller than right-things reduces the possibilities somewhat, and the actual new numbers we get next time around are: ∅|{-1}, which turns out to be -2; {-1}|{0} which turns out to be -1/2; {0}|{+1} which turns out to be +1/2; {+1}|∅ which turns out to be +2. We also get some already-existing numbers in new ways; for instance, {-1}|{+1} is actually equal to 0 ("0 is the simplest number between -1 and +1"). Again, I should explicitly acknowlege that I haven't said anything about how you determine when two of these things are actually equal; again, it does all turn out to work properly.
If you keep going with this construction, you produce all the integers, two at a time, and also all the "dyadic rationals", meaning fractions where the denominator is a power of 2. And then, once you've got all those, at the next stage of construction you abruptly get all the real numbers -- e.g., the square root of 2 is L|R where L = {dyadic rational numbers that are negative or have a square smaller than 2} and R = {dyadic rational numbers that are positive and have a square larger than 2} -- and you also get {0,1,2,3,4,...}|∅, conventionally written as a lower-case Greek letter omega, which is an infinite number, larger than all the integers. (And its negation.) And {0}|{1,1/2,1/3,1/4,...} which is an infinitesimal number, positive but smaller than any ratio of positive integers. And you can then proceed further and construct a vast infinitude of numbers, including all the real numbers (which we've already made) and all of the so-called infinite ordinals (which you can kinda think of as being a sort of "infinite positive integer", though there's more to them than that) and much more, all in a system that lets you do arithmetic and suchlike. It's very elegant, if your brain has been twisted into the mathematician-y shape that finds such things elegant.
What an interesting construction. Thank you from a curious layman for your write-up. I thought it was pretty easy to follow. I'd heard of the surreal numbers before and never knew about the construction mind-game behind them.
Thank you for this explanation! The construction is so elegant, and in a way, the basic idea is simple (?) -- I wonder why it wasn't thought up of much earlier than it was. Maybe it's a little bit like https://en.wikipedia.org/wiki/Egg_of_Columbus
Not only is the basic idea simple, it's a sort of generalization of two other things that were already well known but before Conway were thought of as completely independent.
First: the construction of the real numbers from (traditionally) the rational numbers by means of "Dedekind cuts" (sometimes called "Dedekind sections"). The idea is that if you're trying to build up the machinery of mathematics from scratch, it's not too hard to go step by step from (say) sets to nonnegative integers to integers to rational numbers, but it's harder to get from there to the real numbers, and Dedekind's idea is to say that e.g. the square root of 2 is the way of chopping the rational numbers into "things less than the square root of 2" and "things greater than the square root of 2".
Second: the construction of the ordinal numbers (a sort of generalization of the notion of "nonnegative integer" that allows the numbers to get very infinite) due to von Neumann: you start off saying that zero "is" the empty set, and then you repeatedly say: the next ordinal "is" the set of all the ordinals you've constructed so far. So, e.g., 1 = {0}, and then 2 = {0,1}, etc. -- but once you've constructed all the nonnegative integers you can then look at {0,1,2,...} and that's a new ordinal typically called ω, and then you can take {0,1,2,...,ω} and call it ω+1, and so on and so forth.
Both of these are special cases of what Conway does: Dedekind's is the case where all the numbers are rational numbers and you don't allow either set to be empty, and von Neumann's is where you _require_ the right-hand set to be empty.
There's a further connection, which I believe is how Conway found these things in the first place: if in the definition of surreal numbers you delete the requirement that everything in L has to be less than everything in R, then what you've got is (more or less) the definition of a position in a two-player game. L is the set of positions one player can move to, R is the set of positions the other player can move to. (I say "more or less" because e.g. in many games you're allowed to repeat positions, and games may have complicated winning conditions or involve chance or whatever.) And there's a whole rather nice thing called "combinatorial game theory" that's all about these, and from that perspective numbers are just one particular kind of (position in a) game. (Specifically, a number is a game in which at no point in the subsequent gameplay can it ever make your position better for you to make a move: you'd always rather pass if you could.)
This would make a lot more sense to me if "nothing" and "nothing" were instead "-inf" and "+inf"
Some other comments clarify that "nothing" is more accurately "the empty set". This is helpful because at first I wrongly synonomized "nothing" with zero. But now I get tripped up on the "between" language. Maybe it's a lack of background in sets, but I don't know what "between" implies for an integer (zero) and a set (the empty set).
I didn't want to introduce the notion of infinity because there are actual "infinite numbers" on the surreal number line. I've kind of tried to have both the simplicity of set-theoretic definition and the intuition of the number line, and slightly bungled the exposition. I hope the newly added diagram helps.
I'm not a mathematician and "nothing" doesn't really make sense for me either. But I guess the problem with inf might be that it'd be strange to get +2 as the next thing between +1 and +inf, while getting +1.5 between +1 and +2.
It's more that surreal numbers already include infinities like ω, ω + 1, and so on, so "inf" felt like a concept I want to avoid. The actual description is set-theoretic and uses empty sets there ("nothing to the left", "nothing to the right") so I took that a bit too literally.
Wow thanks. I'm a mathematician and also got lost at the step. This illustration makes the construction much more clear. The text isn't really describing this process well
In either case I believe people who can put AI to the most value are the mathematicians themselves
The net output of math will increase, and mathematicians have more work now to unravel all this, and make it useful. AI plays the role of a monkey in the infinite monkey theorem [1]. We now need an LLM corollary - Something like: A finite number of LLM agents will almost surely find all theorems given an infinite token budget.
"Given infinite thinking time a finite number of humans will solve all theorems"
I also love the angle that this was not intelligence just brute force. As if the mathematicians didn't reeaaally want to solve this they were just too lazy to give it a good try.
What does AI have to actually do before you realize these things are actually smart?
Machines have a much higher capacity for work than human beings. Saying that these proofs did not require equivelant intelligence, but benefitted from sheer volume, does not strike me as unreasonable.
It feels goalpost-movey to downplay exploring a large search space efficiently in regards to "intelligence". If we dug into a human genius's brain and found it was somehow trying out a million ways to solve a a problem at once, no one would seriously suggest the person isn't actually intelligent.
And our brains must something like that at some physical level. You can't have a "turtles all the way down" of reasoning - the building blocks must be simpler. It must reduce to something like pathfinding and brute force at some point, weighted by factors in the system and maybe some randomness.
We have a romantic view of intelligence, perhaps stemming from intuition within the context of scientific discovery. Given enough intelligence, and enough context, a brilliant person can have a stroke of inspiration that allows them to make a major leap (a-la General Relativity or Fermats last theorem). We haven't seen THAT same capacity from a machine, but we see the more ordinary, unsexy grinding type of progress that represents 99.9% of scientific reality.
Is there truly anything new under the sun? Hasn't all of existence alway been here? All math, all physics? We could have merely discovered it. Intuition might be nothing more than combinations of what already exists rather than some sort of divine insight that unlocks previously unknowable mysteries.
Agreed. I don't believe intuition and creativity would be more than pattern recognition, remixing ideas, and trial and error combined with a kind of "genetic algorithm" approach if you deconstructed them into what the brain is actually doing.
I would find it very interesting to train a model on information only available prior to the discovery of e.g. relativity or calculus and see if it can invent it. My intuition is that modern frontiers absolutely could. Not to take away from their brilliance, but Newton and Einstein were brilliant people who also happened to be in the perfect place at the perfect time - there's not so much "low hanging (i.e. approachable by one brilliant individual) but immensely valuable fruit" anymore.
Yes my only point here is that you can argue about the semantics of how smart they really are but they are undeniably smart.
Today it cost massive effort but it's possible 10-20yrs from now an AI could solve a problem like this in under an hour with a single thread on a free subscription paid for by serving an ad.
These arguments are so weak because you'll then have to make the same one a few years from now when it does something else impossible. The argument only stands if we assume no progress will occur.
Funny that you imagine a future where AI can solve complex math quickly, but humans are still watching ads, for some reason.
I just think you and the other guy have different definitions of "smart". There's no denying that LLMs are useful, but I don't know if I'd classify them as "smart". There were probably people in the 80s saying computers were "smart" because they could compute 78971 * 12341 faster than a human.
In what sense is a LLM "undeniably smart" but a CPU from the 80s isn't? Or would you define such a CPU as "smart"?
There certainly are such proofs. Even for simple decidable theories we have very large lower bounds on decision complexity (like double exponential), which implies large lower bounds on the function from "length of theorem statement" to "length of shortest proof".
For undecidable theories, there is no computable function bounding this blowup from theorem length to proof length (otherwise, the theory would be decidable.)
It's impossible for finite number of LLMs to solve all theorems. This would imply that the busy beaver sequence is computable which implies the halting problem is decidable.
For any finite program (eg some LLMs), there is a true math theorem which they cannot prove or disprove (given fixed input of the statement with no other information sources). If that weren’t true, BB would be computable.
Math is beyond computation. Since AI is just bits in bits out, it has this fundamental limitation.
Any magic of AI systems comes from the transformed meaning of its input data. With fixed weights any LLM is just an artifact. For example a human prompting an LLM constitutes an extra information source, which removes the above limitations. In theory any input from the natural world would remove the limitations too. The natural world is a black box and we don't know what kind of meaning or intelligence could underly it.
Even if the busy beaver sequence were computable and the halting problem were decidable, Gödel's incompleteness theorems would still prevent all theorems from being solved, regardless of if one used LLMs or not.
I think there's a really important sense in which Godel's argument is not the full story.
IIUC, Godel's incompleteness is less about theorems and more about axiomatic systems. Given an axiomatic system, there are statements within it which cannot be proven or disproven. It's relatively unrelated to the platonic ideal of the theorem itself. The statements it considers are axiomatic-system-specific.
Another way to view it is, who cares if we can't prove or disprove "This statement is false". Ok, the axiomatic system is incomplete; fine. What's important is can the system prove a real theorem that I care about.
The busy beaver computability argument addresses these issues. The problem format is always "For Turing machine T with no input, does T halt?". This format can encode many math problems. And we know already that BB(432) is independent of ZF, aka, there is a 432-state TMs which ZF can't prove or disprove the halting behaviour of.
So BB looks at real theorems, ranks them, and we can ask what axiomatic systems can solve them or not. Godel looks at 1 axiomatic system and produces a toy theorem which the system can't solve. That's an extremely important difference!
The core issue is that any fixed LLM can only encode so many axiomatic systems in its states, and the fixed systems implies an upper bound in terms of the BB number which it can solve. Godel is only looking at one system at a time, while BB is a way to use a common problem format to rank every axiomatic system on an infinite number line.
But a 432-state TM is a problem that we would like to "prove" is it not? It's not even a particularly complex one to begin with, my smartwatch has orders of magnitude more state then that and yet here we see that all of our math "fails" at it.
I'm no mathematician, but this is also the crux of Gödel's theorem, he just showed it in a more "hacky" and clever way - but BB(432)'s relation to ZF is also a consequence of Gödel's more general idea, is it not?
Even more concretely, the halting problem for turing machines with halting problem oracle would be undecidable for them. And if you could solve that you won't believe what problem would be undecidable. It's turtles all the way up.
The problem with what you're saying is that any old random true proposition about the integers is not necessarily interesting enough to be called a theorem. GIT (or the uncomputability of the Busy Beaver problem) does not establish a limitation on proving theorems, but rather on determining whether a proposition is true or not. Most propositions are ugly and irrelevant. So GIT/Busy Beaver is irrelevant.
-----
Oh, and: All proofs are conditional on axioms. If those axioms are computably enumerable, then all of their consequences are computably enumerable too.
Most propositions may be ugly and irrelevant, but how do you know how many are not so and we just can't prove it? Also, what about stuff like Continuum Hypothesis, would you add it or not?
We are talking about the same thing, but I would actually put this the other way around.
Computation and computability is "the final frontier". Math is a "subset" of that. Doesn't matter if we choose ZFC or in the future discover some "better" subset of core axioms, we will always hit limits where BB will trivially skip over whatever we could prove (let alone Gödel's theorems).
given fixed input of the statement with no other information sources
Also, this is just trivially avoidable, so not sure if we really should be concerned about this limitation. An LLM in a loop where it can write on a tape can be Turing complete, ergo it can compute anything computable and is "bigger" than math at that point.
Computation and computability is "the final frontier". Math is a "subset" of that.
Maybe I'm misunderstanding you point, but I don't know how widely this would be held as true. Are you defining "math" as _only_ what can be proven under some particular formal system?
Well, I only know how to define computability in terms of Turing machines.
For math I don't have a fix definition, but it's surely a bit more specific than that (e.g. I wouldn't consider the computation that prints a 0 at the same place for infinity math) - but of course I do see the circularity in my argument: a Turing machine is a mathematical object in and of itself. Though being able to talk about something doesn't necessarily change which is "bigger".
As for the other direction, this gets a bit more into the philosophy behind math itself. Constructive math's territory is "easy" - but I am on the opinion that if humans (or any intelligent physical entity) are at most Turing-complete [1], then any non-constructive math "steps" or thoughts must also be at most computable. Well, unfortunately I can't prove whether math done by transcendent entities are also computable, though.
In any case, I am no mathematician, so whatever I think regarding this topic may not have much relevance to anyone, only done CS course with quite a bit of math, but that's obviously not the same.
[1] I believe religion is an escape hatch here from an argument perspective
if humans (or any intelligent physical entity) are at most Turing-complete
This is a bit of a strange assumption to make. I do agree that a human, if it had infinite memory, would be an universal machine, i.e. capable of computing any given Turing machine [0]. But would that be the limits of its capabilities? It's far from certain.
You'll get into the philosophy of free will (funnily enough, a sort of inverted Turing test), i.e. for a given human with infinite memory, is there a Turing machine that exactly replicates the behavior of that human? Is our behavior governed entirely by rules? Would that imply that a human themselves is a kind of Chinese room [1]?
any non-constructive math "steps" or thoughts must also be at most computable.
What does it mean for a "thought" to be computable? Compare to Gödel's incompleteness theorem. Clearly the act of stating the thought, or writing down the theorem, is computable. But proving it to be true or false may very well be impossible.
any physical system can be evaluated to any degree of accuracy by a computer
That's an interesting hypothesis, but I don't know why you'd assume it to be true at face value. It's a bit unclear how you would even define "evaluated", given that we don't yet have a mathematical model of all of physics as we know it. [0] And then consider unknown unknowns.
Do you agree that humans are physical systems?
Do you consider humans _with infinite memory_ as physical systems? Do you consider computers _with infinite memory_ as physical systems?
What does it mean for a "thought" to be computable?
Well, given our scientific knowledge it's a molecule-level (only important to disregard quantum physics to make the case easier) physical/chemical process, that we should in principle be able to simulate on any other medium, including a Turing machine.
Nonetheless, I can accept the definition of math where it's about "truths" and truths can obviously exist without being computable.
Does that mean that humans could produce mathematical proofs that are entirely logical and verifiable by other humans, but that cannot be formalised in any automatically verifiable language such as lean?
Computation and computability is "the final frontier". Math is a "subset" of that.
In what sense? BB(n) is a prime example of an object that can be mathematically defined, yet is not computable. Or see BBB(n) for an "even more" uncomputable function. [0]
An LLM in a loop where it can write on a tape can be Turing complete
What does this mean? A given LLM, like a given C program, can't really be Turing complete or not in a meaningful sense. The C programming language, or the concept of LLMs in general can be said to be Turning complete or not. Do you mean to state that LLMs in general are not Turing complete, but being "in a loop" somehow makes a difference?
it can compute anything computable and is "bigger" than math at that point
Again, in what sense is it "bigger" than math? Lots of things are Turing complete, I wouldn't classify lambda calculus as "bigger" than math.
Instead, I have a more complicated view, which I actually expressed in my essay The Two Cultures of Mathematics a quarter of a century ago, and which can be summarized by saying that there is a spectrum of attitudes in mathematics to the relationship between problem-solving and conceptual understanding. At one end of the spectrum you have mathematicians who are primarily motivated by the wish to solve problems, who see conceptual understanding as a very important means to that end. At the other you have mathematicians who are primarily motivated by the wish to attain conceptual understanding, who see problem-solving as a very important means to that end.
Before, understanding and problem-solving-ability were so interdependent that distinguishing between the two was practically very difficult and probably wouldn’t have changed anyone’s research agenda. Now, they’re not connected, and this guy just did the ultimate meta-experiment of seriously undertaking a project that is intentionally 100% problem-solving and 0% understanding to prove it (maybe 99% and 1% but pretty close. In his transcripts, he never asks ChatGPT about the math, only about its opinions of the math).
As we (as a society) sit around asking ourselves what mathematicians (and software engineers, and anyone in deep technical fields) should be doing all day, we now have this case study to show us how wide our range of options has become.
So, assuming my proof doesn’t rely on a Lean kernel bug, it’s likely to be legit too.
He lacks the understanding to verify his solution properly, and has to lean on those who do have the understanding to verify it, only being able to say himself that it's "likely" to be correct. (And what do those mathematicians get for laboriously checking the generated proof? 40 grand?)
Seems to me problem solving is as dependent on understanding as ever.
Moverover, the version I linked above is intentionally paranoid so it doesn't use any third-party code except Mathlib. If you allow usage of CombinatorialGames and trust its definitions, the part that needs to be checked narrows down to exactly 20 lines of code: https://github.com/gaearon/conway-refinement/blob/264445c93b...
What is your point, exactly? Increasing number of people working in and around mathematics are relying on Lean kernel's correctness. That's kind of the point of tools like Lean. Why is it a problem for me to publish a result that relies on it? How do you think other Lean proofs work?
My point is what I said. Without understanding you are only able to say your proof is "likely" to be correct. It's clear from your writing that you understand that your proof will only be accepted once thoroughly reviewed by human mathematicians, who will certainly not be just verifying the definition. Bugs in Lean exist (you're a programmer and it's a program, why would you assume they don't?) and reward hacking and finding bugs are both well established LLM behaviours.
Why is it a problem for me to publish a result that relies on it?
Bit over-sensitive here. I never said it was a problem for you to publish a result. You can do what you like on your blog and spend your tokens however you choose, just as I'm free to have my own opinions on the value of such an effort. I was responding to, and disputing, a commenter's assertion that understanding and problem-solving ability are "now ... not connected".
While Lean is tightening things up after the recent LLM-driven hacks, I agree that bugs are possible. Although usually code that exploits them is obviously aggressive and is deliberately using the more obscure features related to metaprogramming. Also note that my solution has passed the nanoda kernel as well (https://palomar-registry.org/entry?id=PALOMAR-2026-09-03-000...).
That said, again, I never implied that I'm asking mathematicians to "laboriously [check] the generated proof" which is what your parent comment says. The value to mathematicians is knowing that the conjecture is probably right, and knowing the rough path the LLM has taken to it. Instead of checking the Lean proof line by line, what mathematicians are interested in doing (at least, the ones I've been in contact with) is finding a shorter and more direct proof now that they're aware of the outline and main intermediate claims. As for how much value they find in that, I presume they would be able to speak to that when/if they would like to make their research public.
The value to mathematicians is knowing that the conjecture is probably right, and knowing the rough path the LLM has taken to it.
Ah, or is the value to mathematicians that their LLMs can build results on top of this? (In which case, did this do more than save them some tokens?) Or is the value the deep mathematical insight that this result incidentally gives a few mathematicians the confidence to develop for their own personal satisfaction (e.g. if they decide to go and prove it for themselves, and come to the same result after a lot of work)?
(IMO, the deep, scary question: what if it’s soon impossible to make anything at all that anyone who doesn’t know you personally would bother to look at or use? https://www.smbc-comics.com/comic/crack)
Finally, you might be wondering about the token cost. I wasn’t running this project in a particularly token-efficient way and have repeatedly maxed out my 20x Pro subscriptions for both Claude and ChatGPT every week. I also briefly had access to a prerelease model in the last few days, which did not have a usage cap. I was not tracking my actual token usage consistently. Some AI analysis from the recovered logs roughly estimates that we’re totaling around 40 billion tokens, of which around 210 million were output tokens. Over 95% were cache reads.
Who can pay for a 20x Pro subscriptions and also get prerelease models? something weird is going on here.
Several of my coworkers — it’s not that unusual that if you can max out a couple accounts, the companies will obviously notice you (as a high cost customer), and sometimes offer more
I don't know why the author could claim this is "their" proof, and they kept saying "they" did this, "they" built that. but in reality everything is done by the LLM and the author is merely asking it to do things. i guess they did contribute money at least...
Author here! My impression is that it's customary in the mathematical community to take responsibility for the result with your name, regardless of whether it came from LLM etc (as long as you disclose LLM usage). I am perfectly fine calling it "LLM's proof" or somehow else, but it's "my" in the sense that "if there is a mistake in it, it is my mistake".
When you use a drill to put a hole in the wall, do you take credit for it, or do you credit the drill? Without intent, a tool, whether it be a drill or an LLM, is just an inert object.
Someone had to choose the problem, steer the model, and check the output - it's clearly taken a lot of time. That's authorship with a powerful tool, same as it's always been.
They still needed to invest time and other resources into this. It's listed clearly in the 2nd paragraph. Mathematicians, or anyone for that matter, don't figure out everything from scratch. They lean on the work that others have done before them to save time. How is this any different?
I don't understand your comment. At first it seems like you think the LLM should get credit for the work. But then you mock the author for asking about the LLM's mood, which makes me think you believe the LLM is just a tool and not capable of receiving credit (FWIW I would agree.)
Just saying (as an author) I don't believe that LLMs have conscious experiences, but the word "mood" was a good languagespace anchor for the kind of information I wanted to get out of the LLM at the time.
LLMs need a lot of help to get to any kind of complicated proof, really. And yes, they get in moods. I spent 3 weeks trying to prove something and frequently had to try and convince Claude that it wasn't impossible and that it could really do it.
The Claude output in the first one-shot counterexample attempt is hilarious. I hate its writing most of the time but this stuff is next level deep-fried slop.
And the control column confirms the resonance-necessity conjecture empirically: break the skeleton alignment and the joint kernel dies at the constrained window, exactly as the transversality heuristic predicted.
That makes perfect sense. Using language in unorthodox ways to convey very specific concepts that only make sense to you and sound like bullshit to others.
I never expected a machine to be better at articulating complex thoughts with a spare number of semantic coordinates, but then I never expected to find out that my poetry was golfing in the latent space
I had a manic friend in college who later became a Emmy-winning television writer. One night over a quarter century ago, he stayed up about 24 hours straight doing nothing but writing, completely free form, some of it prose, some structured rhyme, some of it dialogue with stage directions. He pinned it to his dorm walls like it was wallpaper for a week, then we took it out to the field and burned it. The surviving paper that didn't burn made up bizarre strings of words that sounded much like this, which he retyped and called it poetry. It totally worked.
This sort of thing is why STEM students need more humanities classes. The idea that poetry is semantically denser than prose is, like, obvious, to anybody who cares about poetry.
I never got the point of poetry, even after giving it an honest shot many times. It felt far too compressed to express any thing. I could not separate it from how I would appreciate music. Maybe its because I don't start from the same base as others.
You are right, maybe some humanities classes would've helped or I'm just wired this way.
It's pretty similar to music, yeah. Or perhaps music can be similar to poetry. Just this morning I was re-reading the lyrics for "Pale Green Things" & "Woke Up New" by The Mountain Goats. There's a sort of specificity in intention that the form tends to end up dictating, that can really carry a rich emotional context or metaphor more directly into the listener's brain.
(Of course, prose is sort of like a superset of poetry - it's not a hard dividing line. You may find especially poetic lines in your favorite longform works, especially in good fiction books.)
I do think that many English teachers do a poor job of encouraging a love, or even understanding, of poetry. One of my highschool teachers was in love almost with categorizing the forms of poetry - sonnets and sestets and villanelles, and seemed to think that if we learned the rules about what lines had to rhyme or which syllables to stress, we'd enjoy poetry. I hated it! But then the next year, even against this backdrop of learned avoidance, my teacher was so in love with the ability of poetry to communicate - the unfolding of a few words into meaning like a flower in bloom - something clicked. He focused almost not at all on the structure of poetry, but rather on finding works that spoke to you. I'm still not a lover of poetry, even now, but I have poems I appreciate.
Here's my conjecture. Large Language Models are the great filter. They represent a local maximum in the technological advancement of a species from which we will not escape.
People will just limit publishing valid works to avoid becoming a hapless plagiarism victim class. Same thing happened to tech bloggers ripped off by low-effort you-tube content makers.
Isomorphic plagiarism makes people feel 23% smarter, but it also provably degrades core skills by 17%.
LLM are great at context search, but are also trivially proven degenerative under recursive self improvement scenarios. We look forwards to stripping their assets at a heavy discount.
Also, we shouldn't kink shame peoples cognitive dildo choices. =3
A wonderfully made introduction to the surreal numbers and their surrounding game theoretic concepts is this video on Hackenbush[0], a winner in 3Blue1Brown's Summer of Math competition.
I’ve emailed some of the mathematicians with a few proposed typo fixes, and I got confirmation that at least a few of those fixes seemed real. However, some of the problems that weren’t backed by Lean also turned out to be misunderstandings.
I think this project is really neat, but is it appropriate to cold email specialists before you've put in enough hours of effort to describe yourself as more than an "amateur"? OP's emails may have been helpful, but billions of people use these LLMs to wade into new areas and email is already low signal-to-noise.
Yeah it's a pretty tough question! I've resisted doing that until I had a relatively high certainty that their published results contained minor mistakes, which I assumed they would want to know about. I've also been explicitly apologetic and tried to keep it super brief.
A counterpoint - I was told a story by one of my professors in the late 1990s, about one of his professors -- he'd written a thesis, gotten hired somewhere like Princeton, and taught there for a few years as Dr. <Somebody>. One day he received a letter pointing out a construction flaw in his thesis. He brought it to the department head who read the letter, and said "Well, Mr. Somebody, ..." Ultimately he fixed the proof.
Upshot, if there are real errors in published work, I think most mathematicians want to know about them.
Yea but that was a person who actually put in work and had to think about and understand the problem. They didn't just generate something with a magic box.
As a software engineer, I could not care less about how a bug was found or by whom, as long as I can quickly verify it is correct, I always appreciate being able to improve my work. I don't see why it should be different for mathematicians (the ones I knew would think similarly, I would assume.)
If I may offer an analogy, what I tried to do is essentially reporting a bug after having a failing unit test that exercises the public API, without looking into the black box of internals.
(Background: trained, published, but still amateur mathematician.)
This is a cool blog post and I think you're going the right way, and beginning to get an understanding of the proof as you go.
I'd recommend continuing on the simplification and understanding route, until you yourself can follow the proof. Some suggestions, as I did something similar:
1. See if (or ask the AIs) if individual parts of the proof can be found elsewhere, i.e., is an argument just a copy of something else? If so, it's important to attribute this, but also this usually allows simplification ("by Theorem X", etc.)
2. Look for redundant patterns and try to combine them.
3. Ask the AI to be a critical reviewer from some journal, and try to fix its criticisms.
4. Continue simplifying! Assume that the final result may actually be relatively short.
OP has reportedly been in contact with Prof. Mantova, who actually worked (jointly with S. L'Innocente) on the key human-authored results behind this AI proof and is arguably in the best position to understand exactly what the AI added that wasn't known before. (See the OP's thread on the Lean Zulip.) So this is happening, and we might see an actual paper publication of this result down the line (possibly encompassing multiple roughly self-contained papers, building up to the final result). The current AI-written version is way too obscure for that, and the AI-written human-targeted "summaries" are not really helpful. Again the OP is quite aware of this.
For some reason, this approach makes me think of the difference between “wizardry” and “sorcery” in some fantasy magic systems. The magic of “wizards” is fundamentally based on a deep study and understanding of arcane things, perhaps assisted by some (necessary or helpful) tools of great power. “Sorcerers” summon supernatural beings and are able to control them, cajole them, and protect themselves and others against them (with more or less success)... but the actual desired magical effect is performed by those beings.
Computing has historically been a field of wizardry. It's... interesting (?) to see so many people pushing so hard in the direction of sorcery, and in fact applying that sorcery to other fields, in which they themselves aren't quite able to validate whether the spell worked or not.
Heh, I like this. but it should be pointed out that from the other point of view, software developers were the supernatural beings (dare I say demons), which the sorcery of a good project manager could tame (with more or less success) to perform the desired magical effect
And now everyone is a sorcerer: they can trap small demons inside metal boxes and force them to do their bidding. As before, any supernatural effects are performed by those beings (which were willed into existence by siphoning the wisdom from the wizard's own grimoires...)
Differently than wizards that lock their knowledge in towers and in sects, software development has a tradition of being open for the most part, so the sorcery and wizardry analogy works more like a spectrum. It just depends on how close to the metal the apprentice would like their consciousness.
That’s fair! I think what struck me, though, is that what we’re seeing is large numbers of “wizards” jump ship to and actively promote “sorcery” instead, in this sense. That is, it’s interesting to me to see Dan Abramov, whose blog primarily consists of painstaking explanations of React internals based on the deep knowledge he developed over years as the most visible member of the core team, switch over to “do a breakthrough.”
I try to address this in the post explicitly in a few places.
Primarily I thought of this as a sort of "epistemic performance art project", maybe similar to playing Elden Ring blindfolded having never played it before, or speedrunning a game by opening a box a thousand times and overflowing some counter. It's funny and absurd to do knowledge work without the knowledge.
I think it's also a stress test of meta skills. Like, how much can we do without knowing? What kind of processes can we set up around these demons that would constrain them into our requirements? How can we know when things are going wrong? In some sense, this isn't too different from engineering management.
Naturally, I'm also interested in how much of my role in this could've been automated away. Can there be a skill for that? Then "do a breakthrough" is an irrelevant implementation detail of that skill.
Note that "do a breakthrough" actually produced the worst results over the runs. The best results were from more directed runs like searching for first obstacle towards the next milestone.
The wizards don’t become sourcerers themselves - they become enthusiastic users of someone else’s sourcery. Their years of learning don’t protect them from mistaking access to power for mastery of it.
I’ve felt like a warlock for about half a year now - talking to demons which summon code from the abyss. Exhilarating and terrifying, especially when you can tell the demons get better faster.
It's a truly wonderful read, made better by the fact the author doesn't quite understand what's going on.
The usual format that fun mathematics is presented (being talked at by someone who is very well versed in the subject) comes with a heavy cognitive burden - and often I just can't really make it through.
When the author is not an expert the writing is just so much more accessible - it's easier to understand and making it through feels more of an adventure and less of a lecture.
I've never thought previously how much I would enjoy this format though. I'm here to see more amateurs stumbling through mathematics.
Also, wasn't expecting this sort of side-quest from the guy who got me into React.
I find the way the author communicates with the LLM fascinating. For example:
However, I didn’t just want any result; I wanted something that pulls me.
Initially, I asked Claude:
Me: which unsolved problems in the Surreal Numbers research program pull you the most and why?
Note the switch from "pulls me" to "pull[s] you". What is the author's perception of the relationship/boundary between them and the LLM here?
1. Are they using it to find things it flags as interesting in hopes they might also find it interesting?
2. Do they consider "interesting" to be a universal (observer-independent) trait and are using the LLM to find things that are interesting?
3. Have they delegated their desire to find something interesting to the LLM so that it can instead find something that it flags as interesting, regardless of how the author feels?
4. Do they see it as a part of their thought process, and so do not distinguish "you" from "me"?
5. Do they see it as part of them, and are referring to the combined entity in the second person?
Let me first clarify my relationship with mathematics. I think of myself as "an awestruck observer from a distance". I find some parts that I understand beautiful, and I have also tried to understand some of the basics rigorously. However, I generally just can't make my way through any serious paper, as I both lack the prerequisites and struggle with the amount of inference mathematics tends to place on the reader. That's the "from a distance" part.
Now, about picking the problem. I am genuinely "pulled by" surreal numbers themselves. I find them irresistibly beautiful. There is also a bit of bitterness around how they haven't fulfilled their promise (yet?) as Conway hoped they would be able to become a better foundation for some mathematics. But they are a bit too difficult to prove things about so far, and we know too little about them. So what "pulls me" also is a possibility of making enough dents in this that we would be able to use them more broadly, and learn even more things about them.
However, I do not know the details of the latest research. I don't know which problems have actually been solved, which pursue Conway's original vision vs narrower approaches, and which are elegant enough to feel "awestruck" enough about. So this is an invitation from me to LLM to share what it "feels pulled by" (for whatever definition; I think of it as just navigating the languagespace) , and then sifting through that list to see if something it lists makes me feel something. I would assume that with the field currently being so small (serious mathematicians mostly don't care about surreals), it's easy to get the LLM "excited" (again, just a vector in the languagespace) enough that it would give me genuinely interesting candidates. Then it's up to me to sift through them and see if they "speak" to me.
It's like asking a mathrock nerd to share their favorite mathrock albums. Niche enough that you'd likely get good results. Then you can listen and form an opinion.
In this particular example, the "ONAG birthday" and "maybe last Conway's unsolved conjecture about surreals" part spoke to me emotionally, the statement itself amazed me with its simplicity, and I felt "blood in the water" related to the recent results bringing the conjecture closer. So I felt the pull myself and went with it.
This year, LLMs have been involved in a number of interesting proofs of conjectures. But that is not even half of mathematics. Has anyone tried to use an LLM to generate a mathematically interesting conjecture, on the level of Conway's refinement conjecture? If so, what happened?
With all the talk of mathematicians possibly being obsolete, I'm wondering where the future conjectures that future LLMs would prove might come from.
My experience has been similar; I find ChatGPT to be much stronger and more precise at math and in communication. I also can not do better with a multiple agent flow than I can with a single agent.
My main thought is that he was performing something general here that is actually valuable and hard for a large proportion of humanity. It's like when Google search was a difficult thing, or troubleshooting a PC. what he is able to do here is actually a rare skill, called intelligence and he may think it's nothing, but it's actually very rare and hard for most people. The kind of judgment and interpretation of output without deep expertise is actually a very rare ability.
Regardless of whether the target result(s) are ultimately correct, isn't it almost guaranteed that supporting infrastructure for surreals-in-lean is a real contribution? Is it a goal to make those polished/reusable, or more like throw-away harness, and just a stepping stone to the proof?
I'm a little tired from the project so not eager to jump back into it right away. But yes, I'd love for useful pieces to make their way into https://github.com/vihdzp/combinatorial-games. Violeta, who maintains CG, expressed interest in ultimately integrating the proof in some shape into the repo, but I think more work needs to be done to understand what makes it work.
this is a great blog post, documenting a very real process of what it's like to create large results with fallible models
though I am an expert at coding, the author's process sounds very similar. constantly double checking, asking for explanations, having AI adversarially check its own work, trying to detect bullshit
Take all the numbers you have so far. Then, “spawn” a new number in every gap between the numbers you already have (crucially, “to the left of all” and “to the right of all” also count as “gaps”). Apply this step forevermore, and you’ll get surreal numbers.
I'm not a mathematician. Can someone explain to me how this approach gets you beyond the rational numbers?
Also, this was formatted as a blockquote, but as far as I can see, this blog post is the only instance of this formulation online.
Got lost here. I think I'm officially too dumb for math.
I feel the same way. Want to become a dumb and dumber duo? I believe you could be my friend.
Pitch: a mini-series where a Dumb and Dumber duo get access to unlimited tokens via a roommates's account (who is an intern at a frontier lab).
In each episode they make a major science-fiction style breakthrough and grapple with the consequences without revealing themselves.
Plot twist: they were both sold on early investment in companies that survived the .com bust. Now they’re VCs that everybody worships as business geniuses even though they’re just lucky idiots, and the sycophantic chatbots finally let them feel as smart as everyone says they are, and a whole bunch of hype-drunk fans are feeding into it.
I think it's just we don't have a way to say/write these numbers. So you make up a way to write them (-1 and 1) and continue.
The numbers don't matter and you could replace -1 and 1 with anything. It's just easier to begin your new fake number at - 1 and 1. Because position does matter.
There has to be some procedure how to come up with "new" numbers though, if you want to have more in the end than just a fancy binary tree - in particular if you want to map your "fake numbers" to the reals, infinity, etc.
This procedure is enough. If you define addition and other operations in a certain way (as Conway did), it turns out that on the omega-th day (i.e. after initial infinite steps), all reals will be born.
Only as a mental abstraction that's based on our experience/concept of space+time.
This is very helpful!
Basically what I take away is that we're inventing a new number system from scratch. So we're not "proving" that 1 is a number between 0 and the empty set. We're defining it as such, and it just so happens that a number system defined this way works out in convergent ways with other mathematics.
Is that roughly right?
Exactly.
(Apparently I was extremely unclear with this text. For clarity: if you want to actually understand surreal numbers, go and read On Numbers and Games, by Conway, which is a delightful book; or get an LLM to talk you through Wikipedia. Original text follows.)
It’s a terrible explanation. A surreal number is defined as a pair of sets of surreal numbers (where you fiddle around the recursion in that definition by defining them in waves, so strictly speaking you’re defining “the surreal numbers born at time T” for each individual T given access to the surreal numbers born at all earlier times, and then you “take the union across all times”, scare quotes because there are too many times for this to result in a set). Zero is a surreal number but the LLM is using the word “zero” to mean “the set containing just the surreal number 0”; “nothing” here is the LLM’s obtuse word for the empty set. Wikipedia may actually be easier to follow.
If you can’t explain it better in the same amount of characters (or fewer), then I don’t think you’re qualified to “nuh-uh!!!” anyone. Sorry buddy.
I mean, I was intending to supply the words that would link the LLM’s explanation to a more normal one, not to explain it; apparently that was extremely unclear. An actual explanation is much longer, as indeed I attempted to indicate by pointing to Wikipedia and saying that it might be more clear.
“Doing better than a totally useless explanation in fewer characters” is in general impossible, of course, eg if the first explanation has only one character.
... wut? :/
LLM didn't write anything in my post; these are all my words and my choices. Conway himself described surreal generation in short like this in ONAG:
This is precisely the intuition I present to the reader of the article. I am relying on visual aid (concretely, the ordered number line) to imply the machinery explicit in the actual recursive definition. The intended reader of this article is not a mathematician, and I think intuition is vastly more important here.
And I don't think I'm conflating 0 with {0} as you claim. When I say zero is "between nothing and nothing", I mean 0 := {|}. When I say one is "between zero and nothing", I mean 1 := {0|}. When I say 1/2 is "between 0 and 1", I mean 1/2 := {0|1}. And so on. I elide "the simplest number" because I am already going in the order of simplicity. I do not need to explain that alternative spellings like 1/2 = {0.2 | 1} are valid because it is not relevant to establishing the mental model of birthdays.
For the finite cases in my explanation, I do not need to explain that the left and the right parts form sets because I only ever need at most one surreal on either side to define the next generation. I also do not need to state the left/right order condition because it is already visually implied by the picture. For the same reason, I do not need to explicitly quantify over the set of earlier-born surreals, since in these finite cases, if we go birthday by birthday, each next day's surreals are definable via the numbers already constructed by the previous day.
I agree that these finite examples don't spell out how to handle infinitely many bounds at the omega-th day, which is where I believe the illustration embedded below is more helpful. I still think "a gap beyond 0, 1, 2, 3, ... with nothing on the right" is a useful intuition when we get there.
For a more precise but accessible treatment, I think https://www.infinitelymore.xyz/p/surreal-numbers is much clearer than Wikipedia.
I'm hoping someone develops an interactive tutor that can teach any subject to any depth.
The tutor should optimize its pedagogy. It should use online RL to adapt to a learner's ideal learning style, model what the student understands and to what degree, and understand what the gaps and next steps are.
I'd subscribe in a heartbeat.
There's a skill for that (haven't tried myself but intend to; other of author's skills I've used have been a game changer).
- https://github.com/mattpocock/skills/blob/main/skills/produc...
- https://www.youtube.com/watch?v=s5T5oQJcJ6U
Ask for analogy in terms of a thing you are an expert in.
ie. i am an expert at zig, explain this c++ in terms of zig
This is called college.
Broadly, universities are too expensive, inequitable, suboptimal, not portable, and slow. There's a huge amount of room for improvement.
Universities are great for networking, starting projects with other students (not the ones professors mandate), and learning lab sciences. In research, they're great for institutional knowledge, having a community of peers, getting guidance from research advisors, having real equipment and funding, etc. But there's a great need for AI tools to accelerate learning outside of that setting.
Anecdotally, I'm a working adult. I'm not going to waste time in college again. I need this for me.
The image helped for me at least.
I think he just wrote it in a confusing way. The quote before says:
So there are two "nothings" here, left of "all" - i.e. the zero - and right of it.
Though I'm not quite sure how you'd get infinite or irrational numbers by this procedure. Wouldn't you simply get the rational numbers by this?
(unless the "put a number" step is doing more work here than it seems. He doesn't really say which number to put there. In the examples, he mostly did "new number = (left number + right number) / 2", with special cases if any number is "nothing" - but he never actually wrote what the rules are here.)
You have infinite steps. Pi is just taking the correct turn an infinite number of times.
Yeah I'm curious about the difference between "nothing" and "0", but I just decided to roll with it. Until the Greek letters made my head start to spin as they usually do.
Hope the newly added picture helps see each step.
I don't think this is your fault; the description isn't very explicit. Let me try to do a bit better. (I'll also try to go somewhat further, and you should not be discouraged if at some point it stops making sense.)
You can think of the "surreal numbers" as being built up step by step. We start out with no numbers at all, and then we repeatedly do a construction that makes some new numbers.
A surreal number is made from two sets of (pre-existing) surreal numbers. We typically call them L and R, for "left" and "right", and sometimes write it as L|R or {L|R} or something like that. The "left" numbers have to be smaller than the "right" numbers. The resulting number will turn out to be, in a certain sense, the "simplest" number in between all the left numbers and all the right numbers.
Now, as I said, we start out with no numbers at all. It might seem like that gives us no way to proceed, but it does: even given no numbers at all, we can still make a set of numbers, namely the empty set! So we can use that for both L and R, getting ∅|∅. Empty sets on both sides. We call this 0, and it will turn out to behave in the way you'd expect the number 0 to behave.
Now we suddenly have another set available, namely {0}, the set containing only zero. Which means that instead of being able to make one number, maybe we can make four: ∅|∅, ∅|{0}, {0}|∅, {0}|{0}. The first of these we already knew about. The last isn't actually admissible -- remember that the "left" numbers have to be smaller than the "right" numbers, which is "vacuously" true when one of those sets is empty (it means "if you have a number x in the left set, and a number y in the right set, then x<y", and if there are no numbers in the left set or no numbers in the right set then that's trivially true) but isn't true when both sets contain 0 because 0<0 is false.
So actually we get two new numbers: ∅|{0} and {0}|∅. The first fits into what OP calls the gap "between nothing and zero". The second first into what OP calls "the gap between zero and nothing". In both cases, "zero" means a number and "nothing" means a space where we don't yet have any numbers.
The number ∅|{0} is called -1 (it has to lie to the left of 0, and there's no constraint on its left, and -1 is "the simplest number less than 0") and the number {0}|∅ is called +1 (it has to lie to the right of 0, and there's no constraint on its right, and +1 is "the simplest number greater than 0").
I should explicitly acknowledge that I haven't defined what "less than" and "greater than" actually mean for these numbers, nor anything else about how they relate to one another that could possibly justify giving these things the specific names 0, -1, and +1. But there are definitions for "less than" and "greater than" and "plus" and "minus" and so forth, and the whole thing does turn out to work very nicely.
Anyway, once we've got these numbers we have eight possible sets that can go on the left or on the right. The requirement for left-things to be smaller than right-things reduces the possibilities somewhat, and the actual new numbers we get next time around are: ∅|{-1}, which turns out to be -2; {-1}|{0} which turns out to be -1/2; {0}|{+1} which turns out to be +1/2; {+1}|∅ which turns out to be +2. We also get some already-existing numbers in new ways; for instance, {-1}|{+1} is actually equal to 0 ("0 is the simplest number between -1 and +1"). Again, I should explicitly acknowlege that I haven't said anything about how you determine when two of these things are actually equal; again, it does all turn out to work properly.
If you keep going with this construction, you produce all the integers, two at a time, and also all the "dyadic rationals", meaning fractions where the denominator is a power of 2. And then, once you've got all those, at the next stage of construction you abruptly get all the real numbers -- e.g., the square root of 2 is L|R where L = {dyadic rational numbers that are negative or have a square smaller than 2} and R = {dyadic rational numbers that are positive and have a square larger than 2} -- and you also get {0,1,2,3,4,...}|∅, conventionally written as a lower-case Greek letter omega, which is an infinite number, larger than all the integers. (And its negation.) And {0}|{1,1/2,1/3,1/4,...} which is an infinitesimal number, positive but smaller than any ratio of positive integers. And you can then proceed further and construct a vast infinitude of numbers, including all the real numbers (which we've already made) and all of the so-called infinite ordinals (which you can kinda think of as being a sort of "infinite positive integer", though there's more to them than that) and much more, all in a system that lets you do arithmetic and suchlike. It's very elegant, if your brain has been twisted into the mathematician-y shape that finds such things elegant.
What an interesting construction. Thank you from a curious layman for your write-up. I thought it was pretty easy to follow. I'd heard of the surreal numbers before and never knew about the construction mind-game behind them.
Thank you this was very well explained
Thank you for this explanation! The construction is so elegant, and in a way, the basic idea is simple (?) -- I wonder why it wasn't thought up of much earlier than it was. Maybe it's a little bit like https://en.wikipedia.org/wiki/Egg_of_Columbus
Not only is the basic idea simple, it's a sort of generalization of two other things that were already well known but before Conway were thought of as completely independent.
First: the construction of the real numbers from (traditionally) the rational numbers by means of "Dedekind cuts" (sometimes called "Dedekind sections"). The idea is that if you're trying to build up the machinery of mathematics from scratch, it's not too hard to go step by step from (say) sets to nonnegative integers to integers to rational numbers, but it's harder to get from there to the real numbers, and Dedekind's idea is to say that e.g. the square root of 2 is the way of chopping the rational numbers into "things less than the square root of 2" and "things greater than the square root of 2".
Second: the construction of the ordinal numbers (a sort of generalization of the notion of "nonnegative integer" that allows the numbers to get very infinite) due to von Neumann: you start off saying that zero "is" the empty set, and then you repeatedly say: the next ordinal "is" the set of all the ordinals you've constructed so far. So, e.g., 1 = {0}, and then 2 = {0,1}, etc. -- but once you've constructed all the nonnegative integers you can then look at {0,1,2,...} and that's a new ordinal typically called ω, and then you can take {0,1,2,...,ω} and call it ω+1, and so on and so forth.
Both of these are special cases of what Conway does: Dedekind's is the case where all the numbers are rational numbers and you don't allow either set to be empty, and von Neumann's is where you _require_ the right-hand set to be empty.
There's a further connection, which I believe is how Conway found these things in the first place: if in the definition of surreal numbers you delete the requirement that everything in L has to be less than everything in R, then what you've got is (more or less) the definition of a position in a two-player game. L is the set of positions one player can move to, R is the set of positions the other player can move to. (I say "more or less" because e.g. in many games you're allowed to repeat positions, and games may have complicated winning conditions or involve chance or whatever.) And there's a whole rather nice thing called "combinatorial game theory" that's all about these, and from that perspective numbers are just one particular kind of (position in a) game. (Specifically, a number is a game in which at no point in the subsequent gameplay can it ever make your position better for you to make a move: you'd always rather pass if you could.)
For more, see these books.
https://en.wikipedia.org/wiki/Winning_Ways_for_Your_Mathemat...
https://web.archive.org/web/20230307045844/https://www.cs.st...
I made a picture, hope this helps: https://excalidraw.com/#json=zfKWWn1h7GzdFca6RDdXl,plr_WeaCt...
Sorry it was confusing.
Edit: the picture is now edited into the article.
This would make a lot more sense to me if "nothing" and "nothing" were instead "-inf" and "+inf"
Some other comments clarify that "nothing" is more accurately "the empty set". This is helpful because at first I wrongly synonomized "nothing" with zero. But now I get tripped up on the "between" language. Maybe it's a lack of background in sets, but I don't know what "between" implies for an integer (zero) and a set (the empty set).
I didn't want to introduce the notion of infinity because there are actual "infinite numbers" on the surreal number line. I've kind of tried to have both the simplicity of set-theoretic definition and the intuition of the number line, and slightly bungled the exposition. I hope the newly added diagram helps.
My favorite intro to surreal numbers is https://www.infinitelymore.xyz/p/surreal-numbers, but it is behind a registration wall.
I'm not a mathematician and "nothing" doesn't really make sense for me either. But I guess the problem with inf might be that it'd be strange to get +2 as the next thing between +1 and +inf, while getting +1.5 between +1 and +2.
It's more that surreal numbers already include infinities like ω, ω + 1, and so on, so "inf" felt like a concept I want to avoid. The actual description is set-theoretic and uses empty sets there ("nothing to the left", "nothing to the right") so I took that a bit too literally.
Wow thanks. I'm a mathematician and also got lost at the step. This illustration makes the construction much more clear. The text isn't really describing this process well
No problem! I've added it to the article, appreciate the feedback.
Someone please vibe-prove that ZFC is inconsistent.
That's awesome! Congratulations!
I'd imagine that in three months when we all have access to communicating agent swarms this should be easier
The net output of math will increase, and mathematicians have more work now to unravel all this, and make it useful. AI plays the role of a monkey in the infinite monkey theorem [1]. We now need an LLM corollary - Something like: A finite number of LLM agents will almost surely find all theorems given an infinite token budget.
[1] https://en.wikipedia.org/wiki/Infinite_monkey_theorem
lol same with people.
"Given infinite thinking time a finite number of humans will solve all theorems"
I also love the angle that this was not intelligence just brute force. As if the mathematicians didn't reeaaally want to solve this they were just too lazy to give it a good try.
What does AI have to actually do before you realize these things are actually smart?
Machines have a much higher capacity for work than human beings. Saying that these proofs did not require equivelant intelligence, but benefitted from sheer volume, does not strike me as unreasonable.
It feels goalpost-movey to downplay exploring a large search space efficiently in regards to "intelligence". If we dug into a human genius's brain and found it was somehow trying out a million ways to solve a a problem at once, no one would seriously suggest the person isn't actually intelligent.
And our brains must something like that at some physical level. You can't have a "turtles all the way down" of reasoning - the building blocks must be simpler. It must reduce to something like pathfinding and brute force at some point, weighted by factors in the system and maybe some randomness.
We have a romantic view of intelligence, perhaps stemming from intuition within the context of scientific discovery. Given enough intelligence, and enough context, a brilliant person can have a stroke of inspiration that allows them to make a major leap (a-la General Relativity or Fermats last theorem). We haven't seen THAT same capacity from a machine, but we see the more ordinary, unsexy grinding type of progress that represents 99.9% of scientific reality.
Is there truly anything new under the sun? Hasn't all of existence alway been here? All math, all physics? We could have merely discovered it. Intuition might be nothing more than combinations of what already exists rather than some sort of divine insight that unlocks previously unknowable mysteries.
Agreed. I don't believe intuition and creativity would be more than pattern recognition, remixing ideas, and trial and error combined with a kind of "genetic algorithm" approach if you deconstructed them into what the brain is actually doing.
I would find it very interesting to train a model on information only available prior to the discovery of e.g. relativity or calculus and see if it can invent it. My intuition is that modern frontiers absolutely could. Not to take away from their brilliance, but Newton and Einstein were brilliant people who also happened to be in the perfect place at the perfect time - there's not so much "low hanging (i.e. approachable by one brilliant individual) but immensely valuable fruit" anymore.
Yes my only point here is that you can argue about the semantics of how smart they really are but they are undeniably smart.
Today it cost massive effort but it's possible 10-20yrs from now an AI could solve a problem like this in under an hour with a single thread on a free subscription paid for by serving an ad.
These arguments are so weak because you'll then have to make the same one a few years from now when it does something else impossible. The argument only stands if we assume no progress will occur.
Funny that you imagine a future where AI can solve complex math quickly, but humans are still watching ads, for some reason.
I just think you and the other guy have different definitions of "smart". There's no denying that LLMs are useful, but I don't know if I'd classify them as "smart". There were probably people in the 80s saying computers were "smart" because they could compute 78971 * 12341 faster than a human.
In what sense is a LLM "undeniably smart" but a CPU from the 80s isn't? Or would you define such a CPU as "smart"?
This isn't necessarily true. There may be proofs so complex, they could exceed the limit of human cognition.
There certainly are such proofs. Even for simple decidable theories we have very large lower bounds on decision complexity (like double exponential), which implies large lower bounds on the function from "length of theorem statement" to "length of shortest proof".
For undecidable theories, there is no computable function bounding this blowup from theorem length to proof length (otherwise, the theory would be decidable.)
It's impossible for finite number of LLMs to solve all theorems. This would imply that the busy beaver sequence is computable which implies the halting problem is decidable.
For any finite program (eg some LLMs), there is a true math theorem which they cannot prove or disprove (given fixed input of the statement with no other information sources). If that weren’t true, BB would be computable.
Math is beyond computation. Since AI is just bits in bits out, it has this fundamental limitation.
Any magic of AI systems comes from the transformed meaning of its input data. With fixed weights any LLM is just an artifact. For example a human prompting an LLM constitutes an extra information source, which removes the above limitations. In theory any input from the natural world would remove the limitations too. The natural world is a black box and we don't know what kind of meaning or intelligence could underly it.
Even if the busy beaver sequence were computable and the halting problem were decidable, Gödel's incompleteness theorems would still prevent all theorems from being solved, regardless of if one used LLMs or not.
Pretty sure Gödel’s theorems imply the halting problem if you squint hard enough.
I think there's a really important sense in which Godel's argument is not the full story.
IIUC, Godel's incompleteness is less about theorems and more about axiomatic systems. Given an axiomatic system, there are statements within it which cannot be proven or disproven. It's relatively unrelated to the platonic ideal of the theorem itself. The statements it considers are axiomatic-system-specific.
Another way to view it is, who cares if we can't prove or disprove "This statement is false". Ok, the axiomatic system is incomplete; fine. What's important is can the system prove a real theorem that I care about.
The busy beaver computability argument addresses these issues. The problem format is always "For Turing machine T with no input, does T halt?". This format can encode many math problems. And we know already that BB(432) is independent of ZF, aka, there is a 432-state TMs which ZF can't prove or disprove the halting behaviour of.
So BB looks at real theorems, ranks them, and we can ask what axiomatic systems can solve them or not. Godel looks at 1 axiomatic system and produces a toy theorem which the system can't solve. That's an extremely important difference!
The core issue is that any fixed LLM can only encode so many axiomatic systems in its states, and the fixed systems implies an upper bound in terms of the BB number which it can solve. Godel is only looking at one system at a time, while BB is a way to use a common problem format to rank every axiomatic system on an infinite number line.
But a 432-state TM is a problem that we would like to "prove" is it not? It's not even a particularly complex one to begin with, my smartwatch has orders of magnitude more state then that and yet here we see that all of our math "fails" at it.
I'm no mathematician, but this is also the crux of Gödel's theorem, he just showed it in a more "hacky" and clever way - but BB(432)'s relation to ZF is also a consequence of Gödel's more general idea, is it not?
Even more concretely, the halting problem for turing machines with halting problem oracle would be undecidable for them. And if you could solve that you won't believe what problem would be undecidable. It's turtles all the way up.
The problem with what you're saying is that any old random true proposition about the integers is not necessarily interesting enough to be called a theorem. GIT (or the uncomputability of the Busy Beaver problem) does not establish a limitation on proving theorems, but rather on determining whether a proposition is true or not. Most propositions are ugly and irrelevant. So GIT/Busy Beaver is irrelevant.
-----
Oh, and: All proofs are conditional on axioms. If those axioms are computably enumerable, then all of their consequences are computably enumerable too.
Most propositions may be ugly and irrelevant, but how do you know how many are not so and we just can't prove it? Also, what about stuff like Continuum Hypothesis, would you add it or not?
We are talking about the same thing, but I would actually put this the other way around.
Computation and computability is "the final frontier". Math is a "subset" of that. Doesn't matter if we choose ZFC or in the future discover some "better" subset of core axioms, we will always hit limits where BB will trivially skip over whatever we could prove (let alone Gödel's theorems).
Also, this is just trivially avoidable, so not sure if we really should be concerned about this limitation. An LLM in a loop where it can write on a tape can be Turing complete, ergo it can compute anything computable and is "bigger" than math at that point.
Maybe I'm misunderstanding you point, but I don't know how widely this would be held as true. Are you defining "math" as _only_ what can be proven under some particular formal system?
Well, I only know how to define computability in terms of Turing machines.
For math I don't have a fix definition, but it's surely a bit more specific than that (e.g. I wouldn't consider the computation that prints a 0 at the same place for infinity math) - but of course I do see the circularity in my argument: a Turing machine is a mathematical object in and of itself. Though being able to talk about something doesn't necessarily change which is "bigger".
As for the other direction, this gets a bit more into the philosophy behind math itself. Constructive math's territory is "easy" - but I am on the opinion that if humans (or any intelligent physical entity) are at most Turing-complete [1], then any non-constructive math "steps" or thoughts must also be at most computable. Well, unfortunately I can't prove whether math done by transcendent entities are also computable, though.
In any case, I am no mathematician, so whatever I think regarding this topic may not have much relevance to anyone, only done CS course with quite a bit of math, but that's obviously not the same.
[1] I believe religion is an escape hatch here from an argument perspective
This is a bit of a strange assumption to make. I do agree that a human, if it had infinite memory, would be an universal machine, i.e. capable of computing any given Turing machine [0]. But would that be the limits of its capabilities? It's far from certain.
You'll get into the philosophy of free will (funnily enough, a sort of inverted Turing test), i.e. for a given human with infinite memory, is there a Turing machine that exactly replicates the behavior of that human? Is our behavior governed entirely by rules? Would that imply that a human themselves is a kind of Chinese room [1]?
What does it mean for a "thought" to be computable? Compare to Gödel's incompleteness theorem. Clearly the act of stating the thought, or writing down the theorem, is computable. But proving it to be true or false may very well be impossible.
[0] https://en.wikipedia.org/wiki/Universal_Turing_machine [1] https://en.wikipedia.org/wiki/Chinese_room
Do you agree that humans are physical systems?
My understanding is that any physical system can be evaluated to any degree of accuracy by a computer, no?
That's an interesting hypothesis, but I don't know why you'd assume it to be true at face value. It's a bit unclear how you would even define "evaluated", given that we don't yet have a mathematical model of all of physics as we know it. [0] And then consider unknown unknowns.
Do you consider humans _with infinite memory_ as physical systems? Do you consider computers _with infinite memory_ as physical systems?
[0] https://en.wikipedia.org/wiki/Physics_beyond_the_Standard_Mo...
Well, given our scientific knowledge it's a molecule-level (only important to disregard quantum physics to make the case easier) physical/chemical process, that we should in principle be able to simulate on any other medium, including a Turing machine.
Nonetheless, I can accept the definition of math where it's about "truths" and truths can obviously exist without being computable.
Does that mean that humans could produce mathematical proofs that are entirely logical and verifiable by other humans, but that cannot be formalised in any automatically verifiable language such as lean?
In what sense? BB(n) is a prime example of an object that can be mathematically defined, yet is not computable. Or see BBB(n) for an "even more" uncomputable function. [0]
What does this mean? A given LLM, like a given C program, can't really be Turing complete or not in a meaningful sense. The C programming language, or the concept of LLMs in general can be said to be Turning complete or not. Do you mean to state that LLMs in general are not Turing complete, but being "in a loop" somehow makes a difference?
Again, in what sense is it "bigger" than math? Lots of things are Turing complete, I wouldn't classify lambda calculus as "bigger" than math.
[0] https://wiki.bbchallenge.org/wiki/Beeping_Busy_Beaver
I find this whole post fascinating in the context of https://news.ycombinator.com/item?id=49738091 and particularly this excerpt from Gowers:
Before, understanding and problem-solving-ability were so interdependent that distinguishing between the two was practically very difficult and probably wouldn’t have changed anyone’s research agenda. Now, they’re not connected, and this guy just did the ultimate meta-experiment of seriously undertaking a project that is intentionally 100% problem-solving and 0% understanding to prove it (maybe 99% and 1% but pretty close. In his transcripts, he never asks ChatGPT about the math, only about its opinions of the math).
As we (as a society) sit around asking ourselves what mathematicians (and software engineers, and anyone in deep technical fields) should be doing all day, we now have this case study to show us how wide our range of options has become.
Yes, the timing of this post just after Gowers' post is fascinating. It is almost as if the marketing machine is well oiled.
What marketing machine? You think someone's paying me to do this?
When you descend into conspiracy theorizing to defend your prejudices, it's time to stop and reconsider.
You can see by who entered the discussion (not you) and immediately sank certain comments that this is a protected submission. One wonders why.
Or perhaps the criticisms are objectively unhinged and are so down voted without having to invent a conspiracy.
He lacks the understanding to verify his solution properly, and has to lean on those who do have the understanding to verify it, only being able to say himself that it's "likely" to be correct. (And what do those mathematicians get for laboriously checking the generated proof? 40 grand?)
Seems to me problem solving is as dependent on understanding as ever.
Author here. No one's asking mathematicians to check the generated proof. I explain it in this part: https://overreacted.io/how-i-vibed-a-proof-of-conways-conjec...
The only thing that needs a check is this 500-line file: https://github.com/gaearon/conway-refinement/blob/264445c93b.... If this file is correct and Lean kernel is correct, the proof is correct.
Moverover, the version I linked above is intentionally paranoid so it doesn't use any third-party code except Mathlib. If you allow usage of CombinatorialGames and trust its definitions, the part that needs to be checked narrows down to exactly 20 lines of code: https://github.com/gaearon/conway-refinement/blob/264445c93b...
There are two ifs in this sentence.
What is your point, exactly? Increasing number of people working in and around mathematics are relying on Lean kernel's correctness. That's kind of the point of tools like Lean. Why is it a problem for me to publish a result that relies on it? How do you think other Lean proofs work?
My point is what I said. Without understanding you are only able to say your proof is "likely" to be correct. It's clear from your writing that you understand that your proof will only be accepted once thoroughly reviewed by human mathematicians, who will certainly not be just verifying the definition. Bugs in Lean exist (you're a programmer and it's a program, why would you assume they don't?) and reward hacking and finding bugs are both well established LLM behaviours.
Bit over-sensitive here. I never said it was a problem for you to publish a result. You can do what you like on your blog and spend your tokens however you choose, just as I'm free to have my own opinions on the value of such an effort. I was responding to, and disputing, a commenter's assertion that understanding and problem-solving ability are "now ... not connected".
I see, we don't seem to disagree much.
While Lean is tightening things up after the recent LLM-driven hacks, I agree that bugs are possible. Although usually code that exploits them is obviously aggressive and is deliberately using the more obscure features related to metaprogramming. Also note that my solution has passed the nanoda kernel as well (https://palomar-registry.org/entry?id=PALOMAR-2026-09-03-000...).
That said, again, I never implied that I'm asking mathematicians to "laboriously [check] the generated proof" which is what your parent comment says. The value to mathematicians is knowing that the conjecture is probably right, and knowing the rough path the LLM has taken to it. Instead of checking the Lean proof line by line, what mathematicians are interested in doing (at least, the ones I've been in contact with) is finding a shorter and more direct proof now that they're aware of the outline and main intermediate claims. As for how much value they find in that, I presume they would be able to speak to that when/if they would like to make their research public.
Ah, or is the value to mathematicians that their LLMs can build results on top of this? (In which case, did this do more than save them some tokens?) Or is the value the deep mathematical insight that this result incidentally gives a few mathematicians the confidence to develop for their own personal satisfaction (e.g. if they decide to go and prove it for themselves, and come to the same result after a lot of work)?
(IMO, the deep, scary question: what if it’s soon impossible to make anything at all that anyone who doesn’t know you personally would bother to look at or use? https://www.smbc-comics.com/comic/crack)
What is your point? Please don't be obtuse, it is more constructive to make your points clearly.
Perhaps mathematicians will undergo the same split as what happened to philosophy and natural sciences.
Finally, you might be wondering about the token cost. I wasn’t running this project in a particularly token-efficient way and have repeatedly maxed out my 20x Pro subscriptions for both Claude and ChatGPT every week. I also briefly had access to a prerelease model in the last few days, which did not have a usage cap. I was not tracking my actual token usage consistently. Some AI analysis from the recovered logs roughly estimates that we’re totaling around 40 billion tokens, of which around 210 million were output tokens. Over 95% were cache reads.
Who can pay for a 20x Pro subscriptions and also get prerelease models? something weird is going on here.
Several of my coworkers — it’s not that unusual that if you can max out a couple accounts, the companies will obviously notice you (as a high cost customer), and sometimes offer more
ah, if it's anyone it'd be dan abramov
I don't see why that's weird
$200 a month is what people pay for morning coffee in the Bay Area.
I don't know why the author could claim this is "their" proof, and they kept saying "they" did this, "they" built that. but in reality everything is done by the LLM and the author is merely asking it to do things. i guess they did contribute money at least...
LOL. mood??
Author here! My impression is that it's customary in the mathematical community to take responsibility for the result with your name, regardless of whether it came from LLM etc (as long as you disclose LLM usage). I am perfectly fine calling it "LLM's proof" or somehow else, but it's "my" in the sense that "if there is a mistake in it, it is my mistake".
When you use a drill to put a hole in the wall, do you take credit for it, or do you credit the drill? Without intent, a tool, whether it be a drill or an LLM, is just an inert object.
Someone had to choose the problem, steer the model, and check the output - it's clearly taken a lot of time. That's authorship with a powerful tool, same as it's always been.
They still needed to invest time and other resources into this. It's listed clearly in the 2nd paragraph. Mathematicians, or anyone for that matter, don't figure out everything from scratch. They lean on the work that others have done before them to save time. How is this any different?
I don't understand your comment. At first it seems like you think the LLM should get credit for the work. But then you mock the author for asking about the LLM's mood, which makes me think you believe the LLM is just a tool and not capable of receiving credit (FWIW I would agree.)
Just saying (as an author) I don't believe that LLMs have conscious experiences, but the word "mood" was a good languagespace anchor for the kind of information I wanted to get out of the LLM at the time.
LLMs need a lot of help to get to any kind of complicated proof, really. And yes, they get in moods. I spent 3 weeks trying to prove something and frequently had to try and convince Claude that it wasn't impossible and that it could really do it.
free time spent talking to llm, what an achievement!
The Claude output in the first one-shot counterexample attempt is hilarious. I hate its writing most of the time but this stuff is next level deep-fried slop.
Discovering that poetry was cognitive compression is alone one of the latent findings LLMs unlocked.
You can so easily imagine this shit being read in a 90s slam poetry coffeehouse. Trust me I was there
That makes perfect sense. Using language in unorthodox ways to convey very specific concepts that only make sense to you and sound like bullshit to others.
I never expected a machine to be better at articulating complex thoughts with a spare number of semantic coordinates, but then I never expected to find out that my poetry was golfing in the latent space
I had a manic friend in college who later became a Emmy-winning television writer. One night over a quarter century ago, he stayed up about 24 hours straight doing nothing but writing, completely free form, some of it prose, some structured rhyme, some of it dialogue with stage directions. He pinned it to his dorm walls like it was wallpaper for a week, then we took it out to the field and burned it. The surviving paper that didn't burn made up bizarre strings of words that sounded much like this, which he retyped and called it poetry. It totally worked.
This sort of thing is why STEM students need more humanities classes. The idea that poetry is semantically denser than prose is, like, obvious, to anybody who cares about poetry.
I never got the point of poetry, even after giving it an honest shot many times. It felt far too compressed to express any thing. I could not separate it from how I would appreciate music. Maybe its because I don't start from the same base as others.
You are right, maybe some humanities classes would've helped or I'm just wired this way.
It's pretty similar to music, yeah. Or perhaps music can be similar to poetry. Just this morning I was re-reading the lyrics for "Pale Green Things" & "Woke Up New" by The Mountain Goats. There's a sort of specificity in intention that the form tends to end up dictating, that can really carry a rich emotional context or metaphor more directly into the listener's brain.
(Of course, prose is sort of like a superset of poetry - it's not a hard dividing line. You may find especially poetic lines in your favorite longform works, especially in good fiction books.)
I do think that many English teachers do a poor job of encouraging a love, or even understanding, of poetry. One of my highschool teachers was in love almost with categorizing the forms of poetry - sonnets and sestets and villanelles, and seemed to think that if we learned the rules about what lines had to rhyme or which syllables to stress, we'd enjoy poetry. I hated it! But then the next year, even against this backdrop of learned avoidance, my teacher was so in love with the ability of poetry to communicate - the unfolding of a few words into meaning like a flower in bloom - something clicked. He focused almost not at all on the structure of poetry, but rather on finding works that spoke to you. I'm still not a lover of poetry, even now, but I have poems I appreciate.
Either I really misunderstood high school English classes or this is what poems have always been known to do.
Absolute bad-metaphor-laden slop, in a dramatic writerly voice. These LLMs are trained on too much pretentious writing.
Here's my conjecture. Large Language Models are the great filter. They represent a local maximum in the technological advancement of a species from which we will not escape.
People will just limit publishing valid works to avoid becoming a hapless plagiarism victim class. Same thing happened to tech bloggers ripped off by low-effort you-tube content makers.
Isomorphic plagiarism makes people feel 23% smarter, but it also provably degrades core skills by 17%.
LLM are great at context search, but are also trivially proven degenerative under recursive self improvement scenarios. We look forwards to stripping their assets at a heavy discount.
Also, we shouldn't kink shame peoples cognitive dildo choices. =3
On the other hand here we have an amateur that could accelerate their learning and experimentation faster than ever possible before.
I don't know if this necessarily qualifies as "accelerating their learning". The user appears to know what the proof is doing, but not how or why.
Really nice "proof guide": https://gaearon.github.io/conway-refinement/#/highlights
and "proof map": https://gaearon.github.io/conway-refinement/#/map/conway-ref...
My god, that proof map has so many parts O.O
A wonderfully made introduction to the surreal numbers and their surrounding game theoretic concepts is this video on Hackenbush[0], a winner in 3Blue1Brown's Summer of Math competition.
[0]https://www.google.com/search?q=video+introduction+to+surrea...
https://www.youtube.com/watch?v=ZYj4NkeGPdM
I think this project is really neat, but is it appropriate to cold email specialists before you've put in enough hours of effort to describe yourself as more than an "amateur"? OP's emails may have been helpful, but billions of people use these LLMs to wade into new areas and email is already low signal-to-noise.
Yeah it's a pretty tough question! I've resisted doing that until I had a relatively high certainty that their published results contained minor mistakes, which I assumed they would want to know about. I've also been explicitly apologetic and tried to keep it super brief.
A counterpoint - I was told a story by one of my professors in the late 1990s, about one of his professors -- he'd written a thesis, gotten hired somewhere like Princeton, and taught there for a few years as Dr. <Somebody>. One day he received a letter pointing out a construction flaw in his thesis. He brought it to the department head who read the letter, and said "Well, Mr. Somebody, ..." Ultimately he fixed the proof.
Upshot, if there are real errors in published work, I think most mathematicians want to know about them.
Yea but that was a person who actually put in work and had to think about and understand the problem. They didn't just generate something with a magic box.
As a software engineer, I could not care less about how a bug was found or by whom, as long as I can quickly verify it is correct, I always appreciate being able to improve my work. I don't see why it should be different for mathematicians (the ones I knew would think similarly, I would assume.)
If I may offer an analogy, what I tried to do is essentially reporting a bug after having a failing unit test that exercises the public API, without looking into the black box of internals.
(Background: trained, published, but still amateur mathematician.)
This is a cool blog post and I think you're going the right way, and beginning to get an understanding of the proof as you go.
I'd recommend continuing on the simplification and understanding route, until you yourself can follow the proof. Some suggestions, as I did something similar:
1. See if (or ask the AIs) if individual parts of the proof can be found elsewhere, i.e., is an argument just a copy of something else? If so, it's important to attribute this, but also this usually allows simplification ("by Theorem X", etc.)
2. Look for redundant patterns and try to combine them.
3. Ask the AI to be a critical reviewer from some journal, and try to fix its criticisms.
4. Continue simplifying! Assume that the final result may actually be relatively short.
Good luck!
OP has reportedly been in contact with Prof. Mantova, who actually worked (jointly with S. L'Innocente) on the key human-authored results behind this AI proof and is arguably in the best position to understand exactly what the AI added that wasn't known before. (See the OP's thread on the Lean Zulip.) So this is happening, and we might see an actual paper publication of this result down the line (possibly encompassing multiple roughly self-contained papers, building up to the final result). The current AI-written version is way too obscure for that, and the AI-written human-targeted "summaries" are not really helpful. Again the OP is quite aware of this.
For some reason, this approach makes me think of the difference between “wizardry” and “sorcery” in some fantasy magic systems. The magic of “wizards” is fundamentally based on a deep study and understanding of arcane things, perhaps assisted by some (necessary or helpful) tools of great power. “Sorcerers” summon supernatural beings and are able to control them, cajole them, and protect themselves and others against them (with more or less success)... but the actual desired magical effect is performed by those beings.
Computing has historically been a field of wizardry. It's... interesting (?) to see so many people pushing so hard in the direction of sorcery, and in fact applying that sorcery to other fields, in which they themselves aren't quite able to validate whether the spell worked or not.
Heh, I like this. but it should be pointed out that from the other point of view, software developers were the supernatural beings (dare I say demons), which the sorcery of a good project manager could tame (with more or less success) to perform the desired magical effect
All this time, I've been considering myself the warlock. When, in fact, I've simply been the Imp. Dastardly news.
The same corollary occurred to me, as well!
And now everyone is a sorcerer: they can trap small demons inside metal boxes and force them to do their bidding. As before, any supernatural effects are performed by those beings (which were willed into existence by siphoning the wisdom from the wizard's own grimoires...)
Differently than wizards that lock their knowledge in towers and in sects, software development has a tradition of being open for the most part, so the sorcery and wizardry analogy works more like a spectrum. It just depends on how close to the metal the apprentice would like their consciousness.
The sorcerers have always outnumbered the wizards. Before AI, we called them code monkeys.
And before those code monkeys were making money, we called them skiddies.
That’s fair! I think what struck me, though, is that what we’re seeing is large numbers of “wizards” jump ship to and actively promote “sorcery” instead, in this sense. That is, it’s interesting to me to see Dan Abramov, whose blog primarily consists of painstaking explanations of React internals based on the deep knowledge he developed over years as the most visible member of the core team, switch over to “do a breakthrough.”
I try to address this in the post explicitly in a few places.
Primarily I thought of this as a sort of "epistemic performance art project", maybe similar to playing Elden Ring blindfolded having never played it before, or speedrunning a game by opening a box a thousand times and overflowing some counter. It's funny and absurd to do knowledge work without the knowledge.
I think it's also a stress test of meta skills. Like, how much can we do without knowing? What kind of processes can we set up around these demons that would constrain them into our requirements? How can we know when things are going wrong? In some sense, this isn't too different from engineering management.
Naturally, I'm also interested in how much of my role in this could've been automated away. Can there be a skill for that? Then "do a breakthrough" is an irrelevant implementation detail of that skill.
Note that "do a breakthrough" actually produced the worst results over the runs. The best results were from more directed runs like searching for first obstacle towards the next milestone.
This reminds me of Terry Pratchett’s Sourcery.
The wizards don’t become sourcerers themselves - they become enthusiastic users of someone else’s sourcery. Their years of learning don’t protect them from mistaking access to power for mastery of it.
I’ve felt like a warlock for about half a year now - talking to demons which summon code from the abyss. Exhilarating and terrifying, especially when you can tell the demons get better faster.
It's a truly wonderful read, made better by the fact the author doesn't quite understand what's going on.
The usual format that fun mathematics is presented (being talked at by someone who is very well versed in the subject) comes with a heavy cognitive burden - and often I just can't really make it through.
When the author is not an expert the writing is just so much more accessible - it's easier to understand and making it through feels more of an adventure and less of a lecture.
I've never thought previously how much I would enjoy this format though. I'm here to see more amateurs stumbling through mathematics.
Also, wasn't expecting this sort of side-quest from the guy who got me into React.
I find the way the author communicates with the LLM fascinating. For example:
Note the switch from "pulls me" to "pull[s] you". What is the author's perception of the relationship/boundary between them and the LLM here?
1. Are they using it to find things it flags as interesting in hopes they might also find it interesting?
2. Do they consider "interesting" to be a universal (observer-independent) trait and are using the LLM to find things that are interesting?
3. Have they delegated their desire to find something interesting to the LLM so that it can instead find something that it flags as interesting, regardless of how the author feels?
4. Do they see it as a part of their thought process, and so do not distinguish "you" from "me"?
5. Do they see it as part of them, and are referring to the combined entity in the second person?
I would love clarification on this.
Hah, very interesting question!
Let me first clarify my relationship with mathematics. I think of myself as "an awestruck observer from a distance". I find some parts that I understand beautiful, and I have also tried to understand some of the basics rigorously. However, I generally just can't make my way through any serious paper, as I both lack the prerequisites and struggle with the amount of inference mathematics tends to place on the reader. That's the "from a distance" part.
Now, about picking the problem. I am genuinely "pulled by" surreal numbers themselves. I find them irresistibly beautiful. There is also a bit of bitterness around how they haven't fulfilled their promise (yet?) as Conway hoped they would be able to become a better foundation for some mathematics. But they are a bit too difficult to prove things about so far, and we know too little about them. So what "pulls me" also is a possibility of making enough dents in this that we would be able to use them more broadly, and learn even more things about them.
However, I do not know the details of the latest research. I don't know which problems have actually been solved, which pursue Conway's original vision vs narrower approaches, and which are elegant enough to feel "awestruck" enough about. So this is an invitation from me to LLM to share what it "feels pulled by" (for whatever definition; I think of it as just navigating the languagespace) , and then sifting through that list to see if something it lists makes me feel something. I would assume that with the field currently being so small (serious mathematicians mostly don't care about surreals), it's easy to get the LLM "excited" (again, just a vector in the languagespace) enough that it would give me genuinely interesting candidates. Then it's up to me to sift through them and see if they "speak" to me.
It's like asking a mathrock nerd to share their favorite mathrock albums. Niche enough that you'd likely get good results. Then you can listen and form an opinion.
In this particular example, the "ONAG birthday" and "maybe last Conway's unsolved conjecture about surreals" part spoke to me emotionally, the statement itself amazed me with its simplicity, and I felt "blood in the water" related to the recent results bringing the conjecture closer. So I felt the pull myself and went with it.
I use you, me, us interchangably because I don't think it really matters for anything and I don't need to reaffirm my individuality with such words.
This year, LLMs have been involved in a number of interesting proofs of conjectures. But that is not even half of mathematics. Has anyone tried to use an LLM to generate a mathematically interesting conjecture, on the level of Conway's refinement conjecture? If so, what happened?
With all the talk of mathematicians possibly being obsolete, I'm wondering where the future conjectures that future LLMs would prove might come from.
My experience has been similar; I find ChatGPT to be much stronger and more precise at math and in communication. I also can not do better with a multiple agent flow than I can with a single agent.
My main thought is that he was performing something general here that is actually valuable and hard for a large proportion of humanity. It's like when Google search was a difficult thing, or troubleshooting a PC. what he is able to do here is actually a rare skill, called intelligence and he may think it's nothing, but it's actually very rare and hard for most people. The kind of judgment and interpretation of output without deep expertise is actually a very rare ability.
Perhaps one thing you should devote effort to is ensuring this has not already been proved in the literature.
I've confirmed with the mathematicians working in that field that this is a new result.
Regardless of whether the target result(s) are ultimately correct, isn't it almost guaranteed that supporting infrastructure for surreals-in-lean is a real contribution? Is it a goal to make those polished/reusable, or more like throw-away harness, and just a stepping stone to the proof?
I'm a little tired from the project so not eager to jump back into it right away. But yes, I'd love for useful pieces to make their way into https://github.com/vihdzp/combinatorial-games. Violeta, who maintains CG, expressed interest in ultimately integrating the proof in some shape into the repo, but I think more work needs to be done to understand what makes it work.
Damn, I was not expecting Dan Abramov when I read the title.
this is a great blog post, documenting a very real process of what it's like to create large results with fallible models
though I am an expert at coding, the author's process sounds very similar. constantly double checking, asking for explanations, having AI adversarially check its own work, trying to detect bullshit
I'm not a mathematician. Can someone explain to me how this approach gets you beyond the rational numbers?
Also, this was formatted as a blockquote, but as far as I can see, this blog post is the only instance of this formulation online.