This is not just PR and is very interesting. However, in my view, (and from a quick read of the paper) this is actually a very classical method in applied math work:
- Build a complex intractable mathematical model (here, Navier-Stokes)
- Approximate it with a function approximator (here, a Physics Informed Neural Network)
- Use the some property of function approximator to search for more solutions to the original model (here, using Gauss-Newton)
In a sense, this is actually just the process of model-based science anyway: use a model for the physical world and exploit the mathematics of the model for real-world effects.
This is very very good work, but this heritage goes back to polynomial approximation even from Taylor series, and has been the foundation of engineering for literal centuries. Throughout history, the approximator keeps getting better and better and hungrier and hungrier for data (Taylor series, Chebyshev + other orthogonal bases for polynomials, neural networks, RNNs, LSTMs, PINNs, <the future>).
You didn't say anything to the contrary, and neither did the original video, but it's very different than what some other people are talking about in this thread ("run an LLM in a loop to do science the way a person does it"). Maybe I'm just ranting at the overloading of the term AI to mean "anything on a GPU".
This is absolutely true, but it still makes use of the advantages and biases of neural networks in a clever way. It has to, because computationally-assisted proofs for PDEs with singularities is incredibly difficult. To me, this is not too similar from using them as heuristics to find counterexamples, or other approaches where the implicit biases pay off. I think we do ourselves a disservice to say that "LLMs replacing people" = "applications of AI in science".
I also wouldn't say this is entirely "classical". Old, yes, but still unfamiliar and controversial to a surprising number of people. But I get your point :-).
The R1-Zero paper shows how many training steps the RL took, and it's not many. The cost of the RL is likely a small fraction of the cost of the foundational model.
Yes, I don’t question the usefulness of the project by any means. To be frank, I’m personally very interested in it—I studied celestial mechanics at university many years ago and am still curious about simulations.
The graph on the chart I shared suggests that the peak of contributions was a couple of years ago, with occasional changes since then. This doesn’t make much sense to me, as the rendering quality looks great (at least in the videos—I’ll try the software a bit later), and it’s head and shoulders above what the scientific community is currently using.
I don't think that it's fair to compare the rendering to what is currently in use in the scientific community, for two main reasons:
The first is that different types of rendering have different uses; typically in scientific visualization this is broken down into essentially "viz for self, viz for peers, viz for others" and oftentimes the most well-used rendering engines are targeted squarely at the first and second categories. The visual language in those categories is qualitatively different than that used for more "outward facing" renderings.
The second reason is that I disagree with your assertion about the quality of the visualization techniques in use within science. There are some truly spectacular visualization engines for cosmology and galaxy formation -- just to pick two examples off the top of my head, the work done by Ralf Kaehler or that by Dylan Nelson. (There are many really good examples, however, and I feel guilty not mentioning more.)
As I said in another, rather terse and unelaborated comment, though, this is really, really impressive work. I think it's important that in praising it, however, we don't discount the work that's been done elsewhere. This need not be zero-sum.
I don’t mean to discount any other work. I have already disclaimed that I don’t work in academia and rely on second-hand feedback from my classmates (in Europe)—for example, the Fortran implementation of Yoshida’s method from N years ago that nobody could modify, or the pressure for publication. Building (or learning) a new rendering engine would be a losing strategy in an academic career, as it is a much more difficult path to getting published. There are far fewer postdoc positions than PhD positions, and rendering skills won’t help in this competition.
Regarding the work of Ralf Kaehler: I have seen his renderings and looked through his articles, but to the best of my knowledge, no source code is publicly available. I don’t consider it fair to count it as something actively used in the field, beyond his lab and affiliated projects.
Disclaimer: that doesn't mean that there are no others, but their availability to researchers is limited to be widely spread.
You can't imagine that someone working on something like this would slow down as the work neared completion? Why must a piece of software / code constantly be changing? What's your specific concern? You're making a very strong claim that the "project has stalled" without any real evidence. Furthermore, the project "stalling" makes it less... what, exactly?
Yes, I can imagine multiple reasons why an author might decide to change their pace for whatever reason. my observation was that it changed.
Based on my experience (both personal and from colleagues), when a project is not in active development, the team starts losing knowledge of the codebase along with its context. For example, something that was at your fingertips while actively working on the project would be much more difficult to recall after a year. The difficulty of maintaining or extending the project grows over time if it is not actively worked on.
‘Stalled’ = contributions become less and less frequent.
If a project has stalled, there isn’t much new happening. For a simulation like this, the sky is the limit—you can make it as accurate as possible (e.g., accounting for light pressure - esp. significant around blackhole acceleration disk, the Yarkovsky effect, etc.)
- Build a complex intractable mathematical model (here, Navier-Stokes)
- Approximate it with a function approximator (here, a Physics Informed Neural Network)
- Use the some property of function approximator to search for more solutions to the original model (here, using Gauss-Newton)
In a sense, this is actually just the process of model-based science anyway: use a model for the physical world and exploit the mathematics of the model for real-world effects.
This is very very good work, but this heritage goes back to polynomial approximation even from Taylor series, and has been the foundation of engineering for literal centuries. Throughout history, the approximator keeps getting better and better and hungrier and hungrier for data (Taylor series, Chebyshev + other orthogonal bases for polynomials, neural networks, RNNs, LSTMs, PINNs, <the future>).
You didn't say anything to the contrary, and neither did the original video, but it's very different than what some other people are talking about in this thread ("run an LLM in a loop to do science the way a person does it"). Maybe I'm just ranting at the overloading of the term AI to mean "anything on a GPU".