Finite-time blow-up for the Navier–Stokes equations
github.comI don't, at all. In fact, and this may be controversial, I kinda like the OpenAI approach of crediting the model, and not an author. The model did the vast majority of the work.
I feel bad for folks like Kevin Buzzard, doing the FLT Lean formalization by hand on a grant and then watching AI blow past it.
I also feel bad for the young people entering the field in such "interesting" times.
This is mostly OpenAI and Anthropic racing each other with a nearly unlimited token budget, hard to feel sympathy for that.
Yea? What does that mean exactly?
I'm working on a quite expansive project and most of its time actually seems to be spent avoiding accidental complexity blowup in the tactics it writes, rather than any difficulty writing the actual proofs. Obviously it's not Navier-Stokes, but it's still super complex stuff. Easily PhD level and way beyond my skill.
Take some form of gradient descent local optimizer and wrap it with a bunch of global genetic algorithms / particle swarm optimizers, and you can get something like this.
It is still applying massive amounts of compute power to grind on a problem, incessantly, but it is a less naive search than just randomly throwing shit at the wall until finding something that sticks.
The route to the Clay problem through a smooth force, options c and d in Fefferman’s statement of the problem, is the route Luis and Diego opened and the one Levent and I had quietly chosen to attack. Almost nobody else I know of was working on it. It is not the direction one arrives at in a few days by giving a model the problem statement. When I heard “forced,” it was a bright red flag.
I'm not sure how anyone walks away from this story and remembers it as "The model did the vast majority of the work."
And who does the model assign credit to? It'd be interesting to see when all these proofs are cleaned up and understood by humans, how much actual innovation in information terms was added by the model and how much of it was its sheer bandwidth to correlate wildly disparate fields of mathematics to come up with proofs of conjectures (i.e., working within an entirely closed system). Don't get me wrong - forming those connections across those fields is impressive and useful in its own right.
But if there is no "innovation", then we are still in Ford's "faster horse" regime, not true intelligence.
Unless it just regurgitated the ideas that were fed into the training data by the human. We might be at or near the point where the models won't depend on that anymore, but in this particular case I think we shouldn't discount this possibility. Even OpenAI hasn't ruled it out afaik.
Really, all the technologies that went into disseminating your comment did it: telecommunications, the PC revolution, the internet, etc. etc.
'Ghiblify' was the epitome, where a marketing demo was deliberately framed to be rude to talented human artists, for no business-sensible reason. The style-transfer was played around with for a few months and then died out as a generally cringe thing to do. Of course, OpenAI's future isn't affected by a bunch of people with pencils in Japan who are liked for practicing a deliberately low-tech craft. On the other hand, OpenAI has not yet made an animated movie.
Now, applying 'ghiblify' dynamics to interactions with scientists, which might be kind of important... the rumour is that OpenAI people talk to scientists, hear something they don't like, and start saying "we should just ask ChatGPT".
The natural counterpoint will be scientists getting contacted by OpenAI for help with something and replying (in more polite terms) "why are you bothering to talk to us if you think we're obsolete? Just ask your AI :)"
If you cleaned a room with vacuum cleaner, do you say the machine cleaned it or you?
The race hasn’t changed, people are now retroactively pretending that the race was always about something other than it was commonly understood to be, because they don’t like that machines are beating them at it.
Obviously there are more aspects here:
- is a conjecture true or false - this has utility on its own
- finding a proof - this is an ability game/race - this used to be a part of math, but I'm arguing at this point this got muddy unfortunately
- finally, is there utility (other than personal desire to understand) in having people do math
Surely, just like we don't mix F1 cars with Usain Bolt, we can still have IMO.
I think this is uncharitable.
It is true that a math paper is useful to the world regardless of whether it was authored by human or machine.
It is also true that people have a deep-seated need to feel esteemed and valued by their community. Many people derive that from their profession and many mathematicians derive it from being able to manipulate symbols in their head in a way few others can to discover truths about the world.
It is cruel to watch that meaning get ripped away and shrug simply because the same math papers are coming out in the end anyway.
The proof of the blowup has zero effect on our ability to predict weather.
Most mathematicians have long expected the Riemann hypothesis to be true — it matters whether techniques from the proof may be used for something, whether there is an insight , not just whether there is a proof.
It would be like feeling bad for an incredible chess player, using a chess engine, losing to a completely incompetent chess player, using a better chess engine.
I think he'd be mildly disappointed that he couldn't publish before GTA6 is released and nobody has time to look at his work because they are busy playing.
I know the comment is just jokes and memes but this is so absurd that I ask people to give me a break.
edit: They didn't and I totally deserved it.
A real world approximation, where the fluid started to be "weird", for some definition of "weird", would be the ultimate Exploratorium exhibit.
People forget that just because classical things are not true, doesn't mean that are not coherent, or that some other classical universe (which might be dull without chemistry etc. but whatever) isn't still "possible".
If I understand the result correctly, it means that the Navier–Stokes equations fail on classical terms, because they don't conserve energy in all cases. This is a theoretical failure, not just an empirical one.
If you are familiar with the Maxwell equations, they perfectly describe all of the electromagnetic field of physics. Navier-Stokes do not have that property for the fluids, unfortunately.
(Unless the solution requires non-physical parameters, like negative viscosity or something, which would be much less interesting).
I wonder if we have enough compute, and/or modelling accuracy, to do a molecule level fluid simulation to see what would REALLY happen as this blow-up scenario is approached? It's not necessarily going to be weird.
I'm not expert though, so someone correct me.
- Cavitation or other phase transitions
- Ionization and even nuclear reactions (which is kind of just an extreme example of the above)
- Viscosity strongly deviating from the linear approximation in N-S.
- Radiative heat transfer away from the hotter regions.
On the flip side, black holes appear to exist and can be formed from fluids. So one might argue that, if you take a fluid and subject it to finite, smooth initial conditions that will turn it into a black hole, then the system really does blow up. But the blowup is described by Einstein’s equation plus some relativistic fluid model, not N-S.
Sadly (IMO), reports of bubble fusion in otherwise mostly ordinary fluids seem to be largely discredited, but I have no real doubts that sufficiently extreme conditions involving real liquids would induce actual fusion or fission. (Fission bombs are generally just imploding blobs of stuff. They might not be liquid, but they’re certainly in a regime where it’s reasonably to attribute some fluid-like properties to them. Fission bonds are conceptually similar except that so much implosion is needed that a fusion bomb is used to supply the energy.)
Anyway, the point is that compressible Newtonian fluids with no friction plus gravity really does develop structure and areas of substantial overdensity in finite time, and if you make a universe and fill it with gas at a density that is just slightly dependent on position, you can end up with black holes!
This is why I wrote "for some definition of "weird'".
Higher level abstractions such as Newtonian physics, or Navier Stokes are not meant to be correct in all situations - they are meant to be useful tools.
If you are using the wrong tool for the job, then don't blame the tool.
We had forced Navier-Stokes and the OpenAI October release already, but I think the generalized Riemann Hypothesis and unforced Navier-Stokes in the same day comfortably sets a new standard for most significant day in the history of mathematics.
``` The formalization contains no data. The certificate whose hash the paper gives is not in this repository; the paper says that it will be released. That the certificate satisfies the conditions of Appendix B is verified by a rigorous computation, involving interval arithmetic (128 bits of precision, with the Arb library through FLINT), which is not part of the formalization and is not in this repository either.
The source of the paper is not in this repository. ```
If finite precision math is acceptable in a "rigorous" verification for some math reason I don't understand, I don't see why we would care that the number of bits is 128 (or less). I suspect that this was AI written and not revised carefully, and I'm really annoyed that this is such common practice now.
Like wow, but whats the application for this neat puzzle you just disentangled.
Number theory was just an esoteric curiosity for mathematicians until modern cryptographic applications started proliferating in the 1970s-1980s.
Lack of present application does not mean something is useless. It could underpin you or your descendant’s reality and livelihood in 20-200 years.
Yes, but in this specific community?
Griping about the lack of applications would be like choosing to visit the Tate Modern or Guggenheim, then walking around tutting that “modern art is all a bit shit isn’t it?”
I’m open minded enough to engage with those opinions, it’s just jarring to see them here of all places and, well, I fail to see the practical application ;)
I don't think the art criticism angle works because your own comment points out that there's not a present application. It's more like walking around the Tate in an empty gallery and a curator excitedly telling you "Two decades from now, we're going to have a really impressive show in here".
What’s narrow minded is trying to draw parallels between concise mathematical tools like linear algebra and the convoluted constructs of 100-page proofs covering the quirks of simulations that appear when the simulation is applied to contexts we already know it has no meaning in.
Last, I think when you look at this through the lens of value capture in addition to value creation, it's hard to think of a cleaner way to accrue value directly to society as a whole than fundamental discovery.
My crypto professor in college loved to read the incipit of a century old number theory book, where the author pride themselves for working on a field of math with no practical application and existing solely “for the glory of the human mind”.
Hadn't they be long dead when modern cryptography arose, they would have been very annoyed I guess.
If you have thought of a stupid idea, some of them have thought of that stupid idea too. You'd better believe Musk and co "watch too many movies"!
When you or I have a big dumb idea like, maybe we should gut government spending without doing any kind of impact analysis, we lack the ability to execute it. Or we have friends. Friends that would say, don’t be an idiot. Powerful people don’t have friends. They have resources.
from my 5,000 foot perspective they seem interesting just because they weren’t solved, while the ramifications, potential, consequence seems missing
What are my blind spots here, can we do anything for society with this knowledge? Does it change an underlying assumption that elevates a stalled branch of technological evolution? is that important at all, just checking because sometimes academics consider it insulting to assume something needs utility, I just have no context here and want some
Will we be able to get an actual simulation of a system, which follows the Navier-Stokes equations, that will actually show the blow-up happen in real time, so we can potentially gain an intuitive understanding of where the equations don’t match reality?
The OpenAI release a couple months ago had a rendering of some arrows ascending in a tornado shape, but I don’t know if that’s just an “artist’s rendition” kind of thing or not… it would be interesting to see if a software simulation that simply followed the NS equations could be given some start conditions and then blow up to infinity, in a way I could see for myself. But I don’t know if the proof actually works that way.