Anyone who understands type theory and how theorem provers work? It's sort of akin to saying "how do you know that a massive C++ program that compiles to machine code compiled to the correct machine code that will actually run and it's just not a random string of bits!?!?!", you know because the compilation would have failed otherwise(this is different than saying the program behaves correctly, but that's precisely the difference between formal proofs and compiled programs).
The entire argument both you and Bessis are implying is that mathematics must be human intelligible. But there's absolutely no reason to assume that every mathematical statement must have a human intelligible representation. There is also not reason to assume that if we restrict ourselves to the subset of mathematical statements that are human intelligible that this is of any use.
Just because people who don't want this to be true, and I can understand the motivation, doesn't mean that it isn't still the case.
Would we still be able to use computers? Of course! They don't need the source code to run.
Would we nevertheless be in big trouble? Oh definitely. We'd need to write all software again, from scratch. Some critical parts we could reverse engineer. Maybe even derive some structures that translate back into source code, but only because a human wrote that source code in the first place.
Hopefully the point is clear: A proof, even if it is correct, that is totally obscure and unintelligable by humans is not very useful for mathematics. It's a black box that doesn't further understanding of the structures and approaches to think about them, and that's what math is all about.Just having a big binary blob of a program doesn't help much if you want to add a feature.
That's also why biology is so hard. There is no source code. It's just millions and millions of years of evolution and things have evolved in weird ways that don't really make it easy to understand them, even though they clearly work.
It remains all of those things, sitting alone in its own repository of 200kLOC; what benefit comes from merging it into mathlib?
> There is also not reason to assume that if we restrict ourselves to the subset of mathematical statements that are human intelligible that this is of any use.
This is obviously silly:
Things that aren’t human intelligible aren’t human usable, so the restriction is necessary to have a collection of things humans can utilize.
This is objectively false, people use things every single day they don't understand. We still have plenty of things about the world we don't understand but still find useful.
You are saying anything we know to be the case, but cannot understand why cannot be used? Can we just stop sleeping because we haven't reasoned why sleep is necessary even though we know it is necessary? I mean we still don't really understand gravity (we know how but not why)
In the 40k universe, modern humanity is below not just their peak of technological prowess, but even their peak of pre-AI prowess. Because they landed in a valley of superstition involving machines they do not understand — and have no way to rebuild towards their peak, without yet further losses by moving away from the magical machines.
That has nothing to do with the universe at large — and everything to do with human society and choices. Which we do have control over.
The comment I replied to equivocated between absolute limits to comprehensibility to humans and sociologically constructed limits, implying that thinking there may be the former makes someone ‘fine with’ the latter. That’s nonsense.
They approximately use them with varying degrees of success, but also mistakes, broken inferences, etc.
My exact point is that your view reduces our ability to do mathematics to that broken, flawed usage and thereby undermines its utility for logical precision: mathematics is only useful because we cleanly understand it.
When you try to use mathematics without understanding, you cause disasters: stock market crashes from mispricing options, Amazon’s 2018 hiring freeze from misallocating $1B, etc.
Note: neither of your examples (sleep, gravity) are things that people intentionally use. They just happen to people.
I think it’s very telling you couldn’t think of an example.
I think it might be fair to say that a proof cannot be without value if it proves something meaningful to a human, that a human can use somehow? But such proof probably doesn’t belong in a library seemingly explicitly dedicated to human-graspable proofs. Just because it violates the intent.
It’s not like such proofs mustn’t exist at all.
The telco that manages loads and allocates networks knows how voice compression works.
The farmers and supermarkets know how the supply chain works.
None of your questions show why mathematics should include blobs of incomprehensible gloop, where no mathematician, no logician, no philosopher, no man on the street can make sense of said gloop, or use it in any way to further human knowledge.
When it's been decomposed down we can discuss this further, but now it's like saying red is red, just because.
But I'll take your expanded statement, to include riding a horse, something even older than the engine. We don't understand fully how a horse works -- biology is still a matter of seeing fragments of the whole -- but people had no problem riding and breeding them before the invention of the car, and before the discovery of genetics.
Meanwhile, understanding the math of a thing -- like stock markets, or nuclear bombs -- does not prevent its use from going badly.
Math is useful and beautiful, and a helpful tool for expanding our understanding of the world, but it is not the whole of understanding, or the sole factor in successful application of science to the world .
Signed, a mathematician.
So what will you do if the doctor prescribes you an LLM-vibecoded drug that nobody understands how it works, yet it cures some deadly affliction with close to 100% efficacy?
What if, say, these incomprehensible math results lead to a revolution in quantum physics which unlocks chip topologies that are orders of magnitude faster than human comprehensible designs?
Would the high priestess of human reason pass her divining rod over such chips or life-saving drugs and reject it as the work of the AI devil?
Again, my exact point is that mathematics loses its utility when you reduce it to that inaccurate usage. You no longer can have any faith in the conclusions — just like sometimes psychiatrists kill their patients with an SSRI prescription because they don’t understand the drugs.
> Would the high priestess of human reason pass her divining rod over such chips or life-saving drugs and reject it as the work of the AI devil?
My point is you can’t know if you’re turning it over for life saving drugs or poison, if you don’t understand what you’re getting.
If you give me an inference chip that runs 200x faster, yes, it could be backdoored to take control of my dishwasher and kill me in my sleep - but I can't deny it runs 200x faster an account of nobody being able to explain why. The same for the mistery cancer drug that cured everyone who took it up to now, but could, without doubt, kill the next patient.
Reasoning by analogy...
If you go to the fiction side of things, you’ll get a lot of things that humans would like to do, but can’t right now as we’re blocked by physics laws and our own biology. We have harnessed electricity, why can’t we harness gravity?
Yet they don't merge this stuff, for many different reasons, including that those 200K lines are free as in puppies.
You can have all the type theory of the world, but the library still needs maintenance.
Change a simp-lemma in a file close to the root. Oooh noes, now there's 987 errors in the 200K loc that we merged last week. And there's nobody there who understands how to fix them.
Just use AI to fix. Or maybe just don't merge the code and let it sit in a downstream library?
And wait until there's evidence that the code is stable, high-quality, with well-designed APIs. And then decide that it might be worthy including in a more foundational library.
I'll remind you of the underhanded C contest and the fact that LLMs are widely documented to take such routes.
So why don;t they use AI to write Lean programs? That should make the AI-proofs more readily human undersrndable.
> In particular, the process of converting a medium-sized Lean document (containing a few thousand lines of code, with some proofs AI-generated) into a nicely golfed and structured Mathlib submission has been an interesting experience. AI agents can be used to perform local golfs that can shave the size of the code somewhat, but global refactoring decisions, such as noticing that a certain argument appears multiple times across the document and can be abstracted into a standalone lemma that can may have additional utility beyond the file, is still largely beyond the reach of current AI tools. (I find that I can explain such a refactor to an AI agent, who can then execute it, but they struggle to spontaneously discover such refactors on their own.)
Here is a other one: hello_world.c versus hello_world.exe (apologies for windows extensions, just for illustration).
One is made by a human for human consumption and extension (though legible by a machine). The source code.
One is made by a machine for a machine. Unreadable by a human. The "binary", though that's a terrible misnomer. (Sure you can disassemble but any nontrivial program is too much to cope with as a whole).
Source vs binary. Both are useful but only one is useful for human consumption.
If there are other lemmas, etc buried inside that 200k blob that can be factored out and proved and used themselves, so much the better. But denying a machine-valid proof just because it's incomprehensible with what a human being considers a reasonable effort made to unpack it just seems odd to me. I see no reason not to accept the vibe coded blob if Lean says it's kosher except anthropocentrism.
So yeah, generated machine-valid proof can be denied if it's incomprehensible, same as human machine-valid proof can be denied for same reasons.
just as we don't re-prove Fermat's little theorem every time I use it in a proof, because well, it's a theorem.
Replacing thought and curation with repeated automation is tech debt, pushed down to fundamental knowledge and understanding.
Exactly! There's a shared foundation, and everyone builds upon it. A mathematical paper is a whole bunch of Lego blocks being added to that foundation, and combining them in a hopefully-useful new interface.
But if the entire paper is just one giant black box, you only get to use the final interface: you lose the ability to meaningfully repurpose the individual Lego blocks to build a different interface. You end up having to reinvent the same Lego blocks over and over again, just with a slightly different color each time.
Don't want to re-prove each and every Lego brick? Then you shouldn't accept giant black box proofs.
Why not just fork the original master branch of human science to an ai-enhanced one and see where that brings us?
As far as I can tell from colleagues in other domains, it’s the same there. One paper will mention something off-hand and that’ll cause someone else to have a spark of insight, which turns into it’s own valuable research
I couldn't disagree more.
A lot of mathematical "problems" are almost entirely pointless. Nobody genuinely cares about the moving sofa problem, or about square packing, or about the minimum number of colors needed to draw a map - it is the math that is developed during the solving process that is valuable!
An answer to a question like "what is the exact area of a unit circle" is a mere curiosity. Calculating a good-enough approximation is trivial, after all. But wanting an exact answer leads to developing calculus, which leads to most modern physics. Science was able to make a giant leap forwards due to the techniques developed, while the actual answer itself is mostly useless.
But let’s consider a hypothetical: what if an intuitive understanding of the true “boundaries” of mathematics (if such things exist) is beyond the capabilities of a human mind? If there truly is no way to simplify some proofs down from 200,000 line incomprehensible gibberish to something you could teach to a high schooler or undergraduate or even a PhD. Is the proof still worthless? Sure, at the moment, it might be. Finding such a proof and understanding the implications of it are different skills, the latter of which AI almost certainly does not possess at the moment. But there may come a time where the AI can view the bigger picture and make the leaps you described (say, an eka-Calculus from an eka-unit-circle). These leaps may be as unintelligible to us as the proof in OP is.
I guess the question is: assuming that we can’t make the proof beautiful enough to spark deeper human understanding, do we still want it if it sparks deeper AI understanding?
Personally I would hate to live in a universe where the boundaries of science are beyond intuitive human understanding, but I think it’s almost certainly the case. The idea that the rules are all within our grasp reeks of anthropocentrism to me. I would love for the universe to prove me wrong though. It’d be a pleasant, hilarious coincidence if they do fit within the boundaries of our understanding.
By extension: why should we assume that a human would still understand the problems - or the answers? If all of it is complete gibberish to a human and can never be applied in any way, shape, or form, then what's the point?
The way I view it there are two options here: either you completely ignore it and end up burning a massive amount of electricity on what is essentially a bunch of LLMs jerking each other off, or you blindly follow it and end up with a Machine God who can justify a genocide with a "This is the correct thing to do. Trust me bro, I have irrefutable proof - you won't understand it". There's just no sensible way to do post-human math in an inherently human world.
Mathematics and science are both about building structure on large classes of facts about the (logical or physical, respectively) universe that allow us to generalize our knowledge and make predictions about facts we cannot yet observe. This also lets us make claims about things that are, individually, too complex for the human mind to grasp, by abstracting away the complex details into a simpler structure and then making the claim about anything that satisfies that structure.
To put it another way, mathematics is about finding beauty — specifically those things that exhibit structure that humans can grasp. Modern mathematics, for the most part, makes no claims about what lies outside that space: if it turns out that the universe consists only of things that humans can describe through mathematics that would be neat, but in the much more likely case that it doesn't mathematics continues as it is today.
It's an interesting and maybe even useful trivium if (according to some set of axioms, such as those implemented by your favourite proof assistant) a fact is true, but it's not (human) mathematics, and if there's no useful way for humans to generalize from it there's no point in including it in a library of mathematics. It wouldn't be that surprising (but would be very interesting!) if there were an entire parallel class (or, more likely, family of classes, one for each AI architecture) of AI mathematics comprising structures that AIs can usefully generalize from. Such a thing has no reason to bear any resemblance to human mathematics, which is based on mapping structures to innate human linguistic and spatial intuitions, and may not yield any insights to human mathematicians.
"Okay, so what determines what is and is not allowed in the collection?"
"Whether the given substatement is true or not, of course."
Like this is obviously silly, right. In your view you could have two guys both trying to prove or disprove that the area of the unit circle is 3, and yet only the guy doing it with some vision of nobility where he's building up to this grand theory of approximations is the one actually doing "real" mathematics. The guy who's doing it just because he thinks it's neat and would like an answer to the problem itself doesn't count, and you suspect he couldn't even exist.
Now if you ask "does the area of unit circle equal 4?", I don't really know, but we can go back to the oracle and ask again (we haven't learned the general pattern).
Also, I'm not sure that assuming this 'area of circle' question was cutting edge math, that the oracle wouldn't say 'yes, to a certain level of tolerance'. Can't count how many times I've seen agent decide a test needs to be loosened or deleted because its an "edge case" or "blocking". If you don't understand the proof you might get back 'yes' for some versions of 'is the area of unit circle 3' (depending on complexity of that ask).
What did the alchemist truly learn from this interaction? Is the answer in any way helpful for him? Will it lead to him making gold, and the implied endless riches it entails? Why should it bring him him to ask follow-up questions with useful answers like "The technology to do this doesn't exist yet and can't be developed within several lifetimes" and "It'll never be economically profitable to turn lead into gold", instead of blindly being fed piecemeal steps in the impossible task of one medieval guy building a LHC?
Writing true statements is absolutely trivial, anyone can do that. There's an endless amount of true statement which have absolutely no value to humanity whatsoever. Their proof is meaningless without something to give it context. "Substatement 1304 is true" has significantly less value than "here's something I call 'group theory', and it might also be helpful solving open issues in fields X, Y, and Z".
Mathematics is about widening our collective understanding. There will of course be some people out there who'll hear "The answer to life, the universe, and everything is forty-two", think "Neato!", and go on with their day without giving it any further thought - but I'd have a hard time calling them mathematicians.
That doesn't mean it is completely useless, of course. A "prove that no efficient algorithm exists to break this new encryption scheme" would be very helpful indeed. Until you realize it neglected to take quantum computing into account, of course.
[0]: https://home.cern/alice-detects-conversion-lead-gold-lhc/
> Nobody genuinely cares about the moving sofa problem, or about square packing, or about the minimum number of colors needed to draw a map
Wild examples. These are all extremely useful.I agree that solving problems along the way is incredibly useful too, but there's no math problem that I'm aware of that's utterly useless. Thing is, math is about abstraction and structure. You're developing tools. Those tools can be used for more than one thing. That's why math is often called a language. It's also why mathematicians and physicists end up being so helpful in other fields, because it was (most of them) about understanding the language of complex problem solving. Even code is math (I find it quite odd programmers are against math and elegance as these are directly in line with their goals, but I guess there's always a Dijkstra EWD already calling this out)
Where I fully agree with you is that solving hard problems advances us. TBH, it doesn't matter what it is. Fund NASA and you not only get men to the moon but velcro, GPS, and global wireless communication. Fund particle accelerators and you get the internet. Fund math and, like you said, you get a tool (calculus) which is worth more money than we can even imagine. The thing is that solving hard problems forces the unknown unknowns to become, at least known unknowns.
Mathematics has to be also understood from the perspective of theory building, not just problem solving.
The beauty of the weapon IS the point. What it proves is derived from its use.
The laws of mathematics exist and their truths hold before they are proven by humans or our machines, so in a very real sense the entire point of proving anything in the first place is anthropocentrism.
That, plus cleaning it up may reveal it contains proofs of other things we also want to know. Imagine if this happened to also contain as a sub-part a proof of all the open Millennium Prize Problems? We don't know until we investigate. (If it was a specific list of things to check from rather than expanding humanity's library, we could just ask an AI to do it for us… but as The Wachowski sisters wrote in their most famous script: "I say your civilization because as soon as we started thinking for you, it really became our civilization").
This future of "the AI built the compiler and it's totally incomprehensible spaghetti, but don't worry, it verified the compiler works using this AI-generated proof assistant whose codebase is also pure spaghetti" terrifies me.
Please show me a proof of `False` in Lean.
LLM agents will a) discover dumb counterintuitive stuff like this and b) exploit it to satisfy the kernel with c) the questionable lines being buried behind millions of lines of math slop. Humans have to check this stuff.
The idea behind the challenge to provide a few lines of human-written/checked scaffolding that defines an "API surface". The API surface is resistant to junk/slop filling out the sorry's. This is supposed to reduce the number of lines that humans have to read to only a handful, while the full valid AI-generated proof is tens of thousands of lines. The challenge was just solved a couple weeks, here's the discussion about it: https://leanprover.zulipchat.com/#narrow/channel/583336-Auto...
(i) accepting that a piece of code is a valid Lean proof (ii) merging a valid Lean proof into Mathlib.
Valid Lean proofs need maintenance. Mathlib is a living blob of code. People care about how fast the proofs typecheck. Many other properties of code play a role.
Not everything that is true is worthy of immediately including in the Encyclopaedia Brittanica.
The novel approaches to solving challenging proofs is very useful.
Have I treated your argument unfairly?
Of course there's no reason AI couldn't do that too.
> A clear explanation can be found in Alex Kontorovich’s account of his own learning curve with formalized mathematics. In a nutshell: Mathlib, the dominant Lean library, is a human-curated formalization of an ever-growing fraction of existing human mathematics. It exposes clean APIs and abstractions, without which no autoformalization could take place. By contrast, Math Inc’s autoformalized proof of Viazovska’s results exposes no intelligible interface. Who in their right mind would merge a 200,000-line unaudited vibe-coded blob into the master branch of global human science?
https://davidbessis.substack.com/p/the-fall-of-the-theorem-e...