Humans are incredibly good at adapting. A few days ago AI solved Navier-Stokes and I was blown away. Now I'm already thinking: "Well, it was only a counterexample and it brute-forced its way to it." lol
Sure. More precisely: they resolved the Navier-Stokes Millennium problem as posed by the Clay Institute. Not sure what else "solving Navier-Stokes" could reasonably mean. A general closed-form solution probably doesn't exist. And numerical solutions have existed for decades. But of course there are still open questions like unforced solutions etc.
I suspect EdwardDiego is referring to the brouhaha about whether OpenAI's training for the model that produced the alleged solution to the Millennium Problem about the Navier-Stokes equations was trained on material that included conversations Tristan Buckmaster and Levent Alpöge had had with earlier OpenAI systems.
I think there's a bit less to that than meets the eye. Yes, OpenAI's result builds on human work. It's possible that it builds on more human work than OpenAI admitted. But even if we suppose that everything Buckmaster and Alpöge did (which, btw, was itself very heavily LLM-assisted/generated work) was a necessary precursor to what OpenAI released, it's still the case that OpenAI's clankers completed the solution and Buckmaster and Alpöge didn't.
My understanding from what Buckmaster has written about this is that the deep mathematical ideas behind their work (and presumably OpenAI's) are due to Córdoba and Martínez-Zoroa. Those ideas are in the published literature, and human mathematicians and AI systems alike are allowed to use them, and doing so doesn't mean they didn't actually do something impressive. Mathematicians build on one another's work; that's how mathematics progresses and always has been.
It may very well be that OpenAI's announcement has a serious problem of professional ethics, especially as their first version of it didn't even list Córdoba and Martínez-Zoroa in its references. (On the specific question of what if anything they learned from B&A's work before that was published: OpenAI are now claiming that after investigating carefully they are confident that the model was not trained on anything Buckmaster and Alpöge did after early July. B&A had been working on this thing for much longer than that. However, on Buckmaster's account of things it wasn't until mid-August that they got beyond what he calls "preliminary results".)
But! The results of B&A were themselves largely AI-generated. (From Buckmaster's statement: "on August 15th, we obtained the blow up results, with smooth forcing, for both Boussinesq and Euler. I can say the first LLM generated proof Levent sent me was the most horrendous I have ever read; we verified it on Lean on August 22nd. Since this point, we have been working around the clock to understand this proof and turn it into something readable." That is: the LLMs found the proof, and B&A had to work to understand what the LLMs had done. It's not that humans did the thinking and AIs just did the gruntwork. (Except in so far as one might want to give all the credit for Real Deep Cleverness to C&MZ.)
And! What OpenAI say their model has proved goes well beyond what B&A did.
I don't see any way of slicing this that makes it unreasonable to say (unless it turns out that there's an error in the proof -- unlikely, given that it comes with Lean verification, but there have been misformalizations and Lean bugs in the past and there surely will be in the future) that AIs solved the N-S problem. No, they couldn't have done it without the work of C&MZ, but again: important mathematical work almost always builds on earlier important mathematical work, that's just how it is. Yes, if OpenAI are lying through their teeth their model might have had early access to B&A's ideas -- but it seems like most of the B&A work was actually done by AI systems anyway.
It is (I think -- I am not an expert and in particular I have not so much as looked at OpenAI's publication) reasonable to say that the deepest ideas here came from humans, and that it was already widely expected that the N-S problem would be solved in the not-impossibly-distant future in something like the way it has been. So, sure, what the AIs have done here is much less impressive than if they'd settled the Riemann Hypothesis or (probably even harder) PvNP. But it's still a resolution of a famous mathematical problem that any human mathematician would have been very proud to have achieved.
Can someone explain if i understand this correctly: Are they saying that they started training this new model on August 28th and then started using it on September 1st? Does training a new model only take 3 days?
OpenAI finished another pre-train in late August, and they are now building models off that base. He's saying the specific model OpenAI used to solve this problem is currently in post-training, which started on August 28.
Not entirely, it's just a late stage of the overall training process. It's an early checkpoint in post training (you can use the model at different stages of training), so it will probably become even stronger with more post training.
> I've heard it said that true understanding is demonstrated when someone can explain difficult concepts well.
I think this not generally. I worked together with this amazing engineer, but he really struggled to sometimes explain what he was trying to do. He came up with great solutions, but often took us some time to figure out what he was trying to get at.
Being good at a thing and explaining it well are different skills. Part of the problem is that you have to empathize with your audience when explaining. It's not sufficient to break a problem down into its essence; its "essence" varies depending on who you're talking to.
An example from something I've had to iterate on: When explaining an event loop multiprocessing runtime sort of thing, I eventually found I had to hand-wave "and your CPU hates that" to establish an appropriate premise to the problem and solution (referring to item-by-item dynamic dispatch with a large number of task types as the specific demon which needed to be slain while discussing that subset of the design). People in the know didn't need more understanding. People not in the know were happy to brush their lack of microarchitectural understanding under the rug. With that premise, both crowds were able to understand what followed.
That wasn't my first attempt. I have a bad habit of trying to explain those missing details as well, especially when it's clear the listener doesn't know them yet -- trying to get them into a position where they could've built the thing themselves -- but that only lands well with like 1-5% of people I've met.
Critically, agreeing with you, that's a communication failure, not an engineering failure. I understood the problem just as well in both cases; I just didn't understand the full extent of the people problem.
I've experienced something similar, I still think the quote works -
If you can explain difficult concepts well, it's a demonstration of true understanding. Not that poor communication demonstrates lack of it.
Oh but I think it can demonstrate the lack of it. There's a distinct difference between learning by memory and learning by understanding, and the only indicator is being able to explain it in a novel way. After all you can also memorize someone else's explanation. In practice, being able to explain it is the only actual difference, if you can't then it's the same as not understanding it and the burden of proof is on the one claiming otherwise. That's the approach upon which the world's education is generally built.
You can do things extremely well without having the foggiest about the actual underlying principles, just from observations and intuition. Doubly so if the process can be machine automated, which by this point encompasses nearly everything to some extent. Sufficiently advanced overfitting is indistinguishable from generalization.
For some tasks, this is absolutely true. We can identify e.g. a dog in picture in 100ms or so, and nobody is capable of explaining how. We can understand and speak languages, without the slightest idea of how it works. Maths is of course not comparable to these cognitive functions, but people with high levels of expertise do have a lot of their knowledge "automated", and not open to introspection.
You can also understand the very simple basic essence of something, but get lost in the complexity when scaling up.
Binary is very simple, but scaled up: look what we've created with software.
When it comes to explanation: pulling from rote memory, requires someone to attempt to hold all the short-term details in mind.
There are biological limitations to how well we can do this, but we can also exercise our brains to improve this ability.
But when something is deeply learned, in long-term memory, the effort of recall is much less than rote memory of short-term details. Our context window is limited, fills up, and we must recover. When you're remembering long-term details, context seems easier to swap in and out (sorry to sound like an LLM, but they do simulate thinking).
Whether or not someone is a master of any given domain of knowledge comes from demonstration. Maybe that is teaching the essence of a subject in a way that demonstrates you can visualize and move around the subject with ease. Or maybe you can create something very useful, or tasteful.
We accept that you have spent time in this area and probably can revral truth to us. You are credible.
If you can't demonstrate mastery through teaching, exchanging ideas to bring me closer to your level: them other forms of credentials are sought: like how well they code, or how useful their products become.
But life isn't about usefulness and will just lead to unhappiness. Just be the best version of yourself you can be. Life is too much to understand all at once.
Expert mathematicians may not be able to explain how they find a proof of a difficult mathematical statement, but once a correct proof is found, it can be, with sufficient work, be formalized (most mathematicians don't do this part). This formal proof can be mechanicaly checked, without any creativity, step-by-step according to the axioms and inference rules of a mathematical logic.
Up to the limits of the Goedels incompletness theorem.
Dependenting on the used notation the formal proof can be very long. For example, Principia Mathematica took about 300 pages to prove that 1 + 1 = 2.
It doesn't matter if you can explain it in a novel way if your novel way doesn't correspond to your listener's ability. So it is possible they can explain it in a different way but still don't know how to explain it in a way that their listener can gain understanding from it.
My sense (I could be wrong) is that once that amazing engineer really understood the class of problems (i.e. he handled it multiple times) and was really comfortable that they would in fact be able to communicate plainly about the decisions and architecture.
Being a good communicator is a skill in itself and requires understanding not only of the problem, not only of yourself, not only of the language but also of other people. And not everyone is capable in all of those at the same time.
Yeah, there's this thing called the curse of knowledge. If an engineer has a deep understanding of something, it's not a given that they can explain it well. For them, the topic feels so simple, and they've done it so many times that they may have forgotten other people aren't as knowledgeable. They will throw terms around without explaining them, etc.
Why are people so negative about this? It feels like a fun project and at 17 the stakes are not really high. Something one could easily do on summer break in a couple of weeks.
Probably some of those issues are not a big problem when you have a good salary (like child/healthcare). And a Silicon Valley salary in Idaho sounds like a pretty sweet deal.
I think you didn't understand what I was trying to say.
Fraunhofer didn't do licencing before joining up with Thomson. That's why - back in the 90s - there were no litigation anywhere even though everyone ignored their patents.
Then Frauenhofer joined up with Thomson and that's where your understanding joins up with reality: now the Frauenhofer does licencing and sues anyone that's not paying them.
My whole comment was a technical tidbit - maybe meaningless from today's perspective... But the previous view interpreted a forward thinking strategy with the monetization of mp3 that just wasn't there
I feel your argument is shifting. Also, it was Fraunhofer that was threatening developers at the end of the 90s not Thomson. Fraunhofer will very aggressively defend their patents/licenses.
I think what GP is saying is that, upon Fraunhofer's licensing of their technologies to Thomson, Thomson demanded (perhaps contractually) that Fraunhofer defend their patents so that Thomson could actually "get something" for having licensed those patents.
This is pretty common when a company exclusively licenses a patent: they demand that the patent holder sue anyone using the tech who isn't them, so that they actually benefit in the market from the exclusivity. Without that enforcement, the money they were paying for exclusivity would be for nothing.
Thomson just made this kind of arrangement in the context of a non-exclusive license — getting Fraunhofer to start legally enforcing their patent rights at all, so that Thomson and whoever else licenses Fraunhofer's patents could benefit in the market from having licensed the patents, by ensuring that non-licensees aren't getting to use the patented tech for free.
In the US, such an arrangement is unnecessary (to the point that it might be causing a miscommunication in this conversation!), since patent holders tend to legally protect their patents all on their own, suing non-licensees for use of the patented tech even when there are no licensees demanding to "get the benefit they paid for." In Europe, patents aren't sued over by default, with enforcement instead only beginning when the patent is licensed and the licensee demands enforcement so that they can benefit.
I am from Germany and public transit is hit and miss here also. Big cities like cologne are great. My commute to work is also fine with public transit (although I mostly bike, actually) But very often there are places I have to go where public transit is pretty inconvenient (adding a lot of time to a trip). I am all for public transit and it should be cheap (maybe even free) and widely available, but it simply doesn’t work for me all the time.
Yeah, fake testimonials and sales numbers are problematic both in the US and in the EU and can actually get fairly expensive. I would just leave this stuff off or be very prepared to back the claims. I think a problem is that a competitor (like slack or jira or whoever) might file a complaint.
So, OpenAI and Anthropic say the Chinese models are only as good because they distill their models. How true is that. I am sure it adds something. But is it more like a marginal 1% improvement or something really significant?
OpenAI's Head of Strategic Futures just this week posted this about the latest Kimi release: "It's a very good model! I don't think its performance can be explained away by distillation or anything like that."
It was part of a longer post that kicked off quite a firestorm about open models and OpenAI's position on them, but it's also notable that labs are no longer contending that open models are essentially just distilled versions of frontier models: https://x.com/deanwball/status/2078133895766114412
> They can invent it. They can build it. And it is only a matter of them before they can scale that last barrier of American hegemony- market it.
And at some point we'll see very capable chips coming out of China: Huawei, Baidu and Alibaba already have some stuff. I think it's only a matter of time before they come up with some AI accelerator doing 80% of the job at 20% of the price.
I am strongly in favor of open models, open source ML more broadly, and am pretty critical of the cynical positions adopted by major US labs vis a vis open models.
But this is an insane characterization. Literally every single researcher and executive at OpenAI and Anthropic would say that "these Chinese labs have a lot of talented people." They hire from them (and vice versa). Tencent's chief AI scientist was poached directly from Deepmind, who poached him from Anthropic, etc etc etc. Do you think there are just zero people from China working at US frontier labs?
And even beyond that, the entire ML ecosystem (including people at OpenAI and Anthropic) get excited about research published by Chinese labs. Deepseek's GRPO paper set the ecosystem on fire for a little while.
The contention from OpenAI and Anthropic around distillation has basically been "Labs that distill from us get to bootstrap their model at a much lower price point". Or, in other words, "If we didn't invest in building the teacher model, it wouldn't be possible for these labs to distill their student model." Which I'm not very sympathetic to, but is a far cry from how you're characterizing it.
Very true, and once the models get even better and smaller and operate locally at a reasonable level there will be even more smart people particularly young people that will get access. The fun has only just started. Like the dawn of the personal computer era.
I think this also maps cleanly on the American blueprint of enshitification. Facebook took off by cleanly integrating and siphoning from Myspace so users could get the best of both on Facebook. Once Facebook took over the market they locked it up tight so no competitor could do the same.
if distilling was so easy and could give you frontier LLM on openai/anthropic output, then how come there are no hundreds of frontier labs in the US market, all distilling and competing for the TRILLION dollar market valuation ????
its all bs spread by oai/anthropic in order to ban open weight models and monopolize the market for two US companies and protect their trillion dollar valuations
Because no VC will give you $5-$10 billion in cash to attempt a catch-up run with Anthropic, OpenAI, and Gemini at this point. Untold billions have been pushed into Grok and it can't keep up. X has had the GPUs, the engineers (reasonably), the cash and the datacenters necessary - it's a very, very, very hard task. Microsoft could afford spend $100 billion on trying to catch up and they might fail at it.
It's a critical national imperative for China. If they were to lose the AI race, it would be economically devastating over the coming decades. Their demonstrated capabilities in the open-weight space are making it fairly clear they are not going to fall behind at this juncture.
As a nation, if you don't have your own GPT equivalent, you will be beholden to a master (right now it's mainly either the US or China, pick one). The EU for example is putting their group of nations at risk in a big way by not going all in on having at least two cutting edge independent competing models (Mistal is not enough). Economically the EU is plenty large enough to accomplish that, nobody is driving the bus the right way.
Distilling isn't necessarily easy, there is a huge cottage industry of services middle-manning ChatGPT and Claude to collect huge amounts of data. It is still vastly cheaper than training yourself, but it is certainly not easy or feasible for most organizations. And I'm sure a flock of lawyers would show up if someone in America was found doing it.
reply