AI/LLMs Navier-Stokes "Solved" by AI

[trappedslider]

Two things:

1) As already noted, this is still humans stealing from other humans - the AI didn't do this of its own accord - it was a tool for human ambition.

2) Justice DID NOT claim this was "worse" than other stealing. Just that it was an attempt to steal credit. The "worse" thing is all you.

Why are you asking people to justify statements that they didn't make?

So, if/when OpenAI gives credit, then this would be okay?

EDIT: looking back, that should have been my actual question and I apologize for that.

 

[Justice and Rule]

But that has nothing to do with AI. Scientists have been snooping on other scientists research trying to scoop them since basically the dawn of times.

I mean, during my first postdoc, my supervisor told me to immediately go and pick up every paper I printed, because people taking somebody else bibliography research to try and scoop novel ideas was apparently a thing.

Yup The Scandalous History of the Cubic Formula | Quanta Magazine and that's not even getting into the whole Calculus thing with Newton or Edison/Tesla and others. I guess when it's ONLY humans doing it to other humans it's okay.



I'll note that Open AI is not asking for the prize money.

Way to miss the point.

Y'all have been promoting AI as a big gamechanger in academia, since it can help do things like this. But this whole situation pours a ton of cold water on that: you are less likely to use that big game-changing technology when it is going to be used to scope all your work.

Like, you try to write it off as "It's happened before", but it's never been an inherent part of the new technology being used. Like, when I use a calculator to do work, I don't have a worry that Texas Instruments is going to somehow see that and steal my work. That is the problem: OpenAI, in its pursuit of getting good press and publicity, basically stole someone else's work and attempted to bully them out of the credit or force themselves into it at the cost of someone else who had been assisting. Li

so human theft is bad but adding AI makes it even worse?

No. Again, as an AI booster, you completely miss the actual problem with this. It is going to make AI less likely to be used even in situations where it'd be beneficial because we've now seen that an AI company is willing to completely undercut someone for a headline. People are going to be less likely to use AI, or at the very least cross-pollinate with different models as tools. It's an utterly boneheaded thing to do that poisons the well pretty thoroughly, especially OpenAI's actions after it was done.

So, AI doesn't make it worse, it just make way easier and way more likely. The hunger for press to offset the lack of profit margin means that AI companies are practically incentivized to do this sort of thing, and it is going to deter people whose work could actually benefit from these things.

See Justice using the word "stealing"

Now my using the Newton/Leibnitz and Edison were incorrect. But Justice did make the assertion that it was stealing and given that the topic is AI I think asking "why is this worse than humans stealing from other humans" is worth asking.

It is stealing. Absolutely, positively, totally and completely. They took someone else's notes and work without permission and used them to solve the problem before they could. Ask any academic and they are totally going to say that's intellectual theft. You can justify it in a dozen different ways if you like, it doesn't change what it actually is.

Edit: As to why it is "worse", I didn't say that but I suppose it's "worse" in the sense that it truly poisons the desire for a lot of academics to share and input their stuff into larger AI systems, particularly ones meant for academia. When you know someone might find out you are on to something and will steal your work for the press, you are less likely to use it.

It's not that the act of stealing here is worse, but what it reveals about AI companies and what they will do with your "private" work.

So, if/when OpenAI gives credit, then this would be okay?

EDIT: looking back, that should have been my actual question and I apologize for that.

No. It's an insane breach of trust and privacy to do what they did, all for a headline. "Giving credit" doesn't take away from them basically acting to steal credit from someone who did not desire or ask for their help. They should have waited for them to finish on their own and let them give credit to what was there (which would have included OpenAI's Codex anyways). They've seriously harmed AI usage in academia in a place where people like @The Firebird have (justifiably, to an extent) argued that it has the most beneficial use.

 

[FrogReaver]

I feel it worth noting: there is no proven theft here. There is speculation. There is theoretical opportunity, but no smoking gun - and likely never will be one.

 

[briggart]

Way to miss the point.

Y'all have been promoting AI as a big gamechanger in academia, since it can help do things like this. But this whole situation pours a ton of cold water on that: you are less likely to use that big game-changing technology when it is going to be used to scope all your work.

Like, you try to write it off as "It's happened before", but it's never been an inherent part of the new technology being used. Like, when I use a calculator to do work, I don't have a worry that Texas Instruments is going to somehow see that and steal my work. That is the problem: OpenAI, in its pursuit of getting good press and publicity, basically stole someone else's work and attempted to bully them out of the credit or force themselves into it at the cost of someone else who had been assisting. Li

I don't believe I have been promoting AI, even though I am amazed at what it can achieve, and especially at the speed of progress in the field.

I am not trying to write off what was done, I pointed out that is a human issue precisely because I do not believe that what happened is inherent to the technology. There are open AI models, which anybody can download and run locally with total data privacy if they desire.

On my laptop, which is a good one but not top of the line of what is currently available, I can run LLM which rival what was state-of-the-art a 1-2 years ago, at least for the technical tasks I am interested into.

With ~ 100k$ you can run locally models that are just a few months behind the current offer from OpenAI, Google, etc. This is not a regular Joe budget, but is within the reach of an academic institution.

EDIT: just to clarify, with 100$k hardware investment.

 

[The Firebird]

I don't believe I have been promoting AI, even though I am amazed at what it can achieve, and especially at the speed of progress in the field.

I am not trying to write off what was done, I pointed out that is a human issue precisely because I do not believe that what happened is inherent to the technology. There are open AI models, which anybody can download and run locally with total data privacy if they desire.

On my laptop, which is a good one but not top of the line of what is currently available, I can run LLM which rival what was state-of-the-art a 1-2 years ago, at least for the technical tasks I am interested into.

With ~ 100k$ you can run locally models that are just a few months behind the current offer from OpenAI, Google, etc. This is not a regular Joe budget, but is within the reach of an academic institution.

EDIT: just to clarify, with 100$k hardware investment.

Right. The game-changer for colleagues in academia was the Opus models plus Claude Code, which materialized around November 2025. Those are sufficient to have a substantial impact once that level of intelligence can be routed into our workflows, by which I mean: instrument computers, data analysis routines, and figure generation. (A clear majority of folks under the age of, say 35, exclusively use these products for figures already, but not the other things).

Fable and Astra are clearly interesting and useful but not really necessary. They will either find a niche in different tasks, or become so cheap they can be deployed everywhere.

Open models are sufficiently powerful but require institutional investment in compute, and most institutions seem to be partnering with Anthropic, or to a lesser extent OpenAI, or getting Gemini access from previous agreements with Google.

Qwen-27B is iirc close to the necessary performance (i.e., ~to Opus 4.5) and can be run locally but requires some setup. Currently token costs via subscription are cheap enough--and there are heavy subsidies, especially from Anthropic, at the moment--that no one is really looking at this. But the capability seems there.

ETA: For an idea of where things may go in terms of cost check out "Jimmy Chat". Its a demo from a company. Rather than encoding their model in software and using GPUs to run it, they encode it directly into the chip. That has a downside--you can only use the chip to run that model--but it is enormously fast. Currently they use it for a rather poor open source model. If you can manufacture the equivalent for sufficiently intelligent models, you may be able to run them at extreme speeds for a very low price.

 

[Stalker0]

Right. The game-changer for colleagues in academia was the Opus models plus Claude Code, which materialized around November 2025. Those are sufficient to have a substantial impact once that level of intelligence can be routed into our workflows, by which I mean: instrument computers, data analysis routines, and figure generation. (A clear majority of folks under the age of, say 35, exclusively use these products for figures already, but not the other things).

Fable and Astra are clearly interesting and useful but not really necessary. They will either find a niche in different tasks, or become so cheap they can be deployed everywhere.

This is my take as well. Around when Opus came online (and later as some of the less models were improved), was when I really noticed an uptick in the "intuitiveness" of the models. I went from having to explain everything to the models just figuring out what I wanted and needed with much less babysitting.

Fable I've used peridoically for some analsyis work but I haven't found it that much better than Opus and its a absolute token hog. Haven't gotten to try Astra yet.

 

[Cadence]

Y'all have been promoting AI as a big gamechanger in academia, since it can help do things like this.

RE: AI in Academia

258 papers so far this year by one author...
https://statmodeling.stat.columbia.edu/2026/08/27/258/
https://www.linkedin.com/posts/abel...or-has-posted-share-7499495028915048448-F3Lp/

Papers coming out so fast the "author" can't have even legitimately had a chance to read them. I wonder if he'll use agents to pick the best ones to submit and do the submitting too...

 

[briggart]

RE: AI in Academia

258 papers so far this year by one author...
https://statmodeling.stat.columbia.edu/2026/08/27/258/
https://www.linkedin.com/posts/abel...or-has-posted-share-7499495028915048448-F3Lp/

Papers coming out so fast the "author" can't have even legitimately had a chance to read them. I wonder if he'll use agents to pick the best ones to submit and do the submitting too...

Yes, this is IMO a more relevant concern for AI use in academia. Fortunately my field has been relatively immune to this until now, but the risk is there.

 

[FrogReaver]

RE: AI in Academia

258 papers so far this year by one author...
https://statmodeling.stat.columbia.edu/2026/08/27/258/
https://www.linkedin.com/posts/abel...or-has-posted-share-7499495028915048448-F3Lp/

Papers coming out so fast the "author" can't have even legitimately had a chance to read them. I wonder if he'll use agents to pick the best ones to submit and do the submitting too...

If he has a process where he can generate 258 papers that pass peer review in a year then what's the issue?

 

[Cadence]

If he has a process where he can generate 258 papers that pass peer review in a year then what's the issue?

It will be interesting to see how many of these he submits for peer review.

One issue is the number of peer reviewers needed and strain on the editorial staff. In my field each of those would be gone over in depth by an associate editor (AE) and 2-4 reviewers, with the AE sometimes have needed to ask a half dozen or more people to find those who agree to review, and the editor would have had to do at least a pre-screening, AE assignment, and then check the AE/reviewer recommendations. Do we just offload the review process in large part to LLMs (which are not actually expert systems, are they) too? Or do we start requiring everyone to non-AI assisted review at least one paper for each journal they co-author a submission/revision to when asked? (Maybe some sort of global review credit system). AEs and reviewers in my field are currently uncompensated and even if used an AI on expert mode to check things isn't free.

Another issue is authorship. Can you actually be called an author of something you never even read? (Does that happen in the lab sciences sometimes or an area where someone only focused on the data analysis?) In this case, can you have a paper where literally none of the authors ever read it? Should we require them to list the AI model as the author in an extreme case like this? Should journals start distinguishing between author and contributor? (Some in other fields have checklists each author uses to indicate what part of the process they contributed to that come out with the article).

Can journals copyright any articles made this way at all or does it defacto make them all open? (What is the latest court case related to this?)

How is the output judged for tenure and promotion? Right now in my field it is common to divvy up percent effort when describing the paper in the tenure file (what percent of the work goes to each person). If it was literally just a very broad idea and a prompt, what percent is that? Will the motivation be somewhat self correcting? Somewhat self-correcting only for junior faculty? Will fields where AI is able to do more of these be down-weighted greatly in terms of perceived intellectual merit?

Are the LLMs better at avoiding plagiarism now? (I wonder what percent of papers it still happens in, and if you have it spit out 200+ what the odds are of them all being clean?). If we ran it like we do academic integrity for students cheating, I wonder how long until someone doing this is blackballed for having one really egregious case or a few iffy ones. Is it fraud to cite a non-extant paper? I assume it is easy to make an LLM-callable tool that find fake cities and stop them. Can journals make the online terms of use be that you can't feed their articles into an LLM? What legal enforcement do they have then? Will LLMs do a lot worse if they can only second-hand cite things?

As an aside, similar issues are occurring in grant proposals.