If a prompt can solve a useful math proof, then someone else will find that prompt. In which case, openAI won't get the street cred(wall street) it needs. Nobody is hiding secrets to gain advantage in this environment.
OpenAI cracked these frontier math proofs using models and amount of compute not available to general public. It's not the case that anyone can put "solve Navier–Stokes" in ChatGPT and get solution. Although less well-known problems can be solved with publicly available models.
Not only are the models and compute unavailable to the public, they are unavailable to anyone besides the labs, including the government. We better hope the furthermost tip of the jagged frontier stays in purely theoretical mathematics, because if it strays into the physical world then we may find ourselves in a situation where the power of the labs rivals the governments supposedly overseeing them.
Indeed. Their model they showed off that was ~5x better than 6 Astra in certain domains including mathematics is not public and I doubt they'd release it. They distill their internal large into their public models.
Do you think companies do not hide secrets? I think you are wrong but maybe I do not understand your point of view. However let's see my example.
Imagine: you are a boss of a company and own a powerful tool which can give you a lot of advantages. In short: it allows you to earn more money than competing companies. Would you share this tool to everybody, so that means competing companies could destroy your business or you keep the tool in secret and slowly build empire?
I do not know if OpenAI has powerful LLMs/AI things but I can be sure they do not share everything what they have. I think the only way they could share hidden "powerful tool" is situation where men in black visit CEO and talk with him/her about co-op. Of course CEO is "happy" because he/she can not refuse if you know what I mean. But still hidden powerful tool is hidden but this time another "more important guys" have access to it.
Its more about finding artifically created disabilities. My guess is public key cryptography is solved and the government at least knows. Which is why we had the "export" forbidden noises. Apply a old solution (export controll)to new problem knowledge intrferable.
What's the $-value of some street cred/publicity vs. $-value of finding some optimization trick that reduces training/inference cost by millions compared to their competitors? Of course they're going to keep that IP to themselves.
I think it depends on the situation. If there are profitable applications, they might keep the proof to themselves, but I agree with you in most cases.
Stupid me thinking they only spent time on brute force-able problems with pure compute, not sure there's much intelligence here. What could a mathematician do with $20M worth of compute to try to brute force one single problem?
My assumption on cryptography was that any published result could potentially help had actors, was unlikely to help good actors, and better to stay away completely. Can definitely see why they'd research but not publish in the other areas though.
Probably they have either tried using LLMs to crack most cryptographic systems, or they have strict rules against doing so. Either way, the US govt. will probably make sure it gets a chance to do so before this internal model is released. There's a good chance they have started already.
> In September 2nd, something like that, the biggest conjecture in our field was proved. Did your life change? My life? No. No. No. I mean, the only people that can actually tell you that it does change something are the mathematician who worked on it. The human understanding of mathematics is the only thing that at the end benefits society.
The Strip Mining of Mathematical Proofs (with Hugo Duminil-Copin) - Numberphile Podcast https://youtu.be/omuJpoV-Ko4
Remember, Trump has said if you don't use SI instead of AI you are an enemy. So, this title should be changed to "The math that OpenSI keeps for itself."
> get something that helps us understand what is going on, and live to see a better future.
Nah, I don't really buy the sales pitch that ultra advanced math will give us a better future. For decades we've had ultra advanced math already before AI like geometric langlands, higher topos theory, and Waldhausen categories and it's done nothing to give us a better future.
Maybe it helps to train people to think abstractly and inspire technical people through the beauty of math but that's only when it was a human endeavor anyway.
The idea of going to even more layers of abstraction and complexity with AI may be possible but it's just a mental game at that point and isn't going to help anyone. It will just waste energy.
As far as I'm concerned (as a former mathematician), they can take their new discoveries and shove it. Yes, science does help us live easier lives in some way but it doesn't follow that endless new discoveries always been better lives past a certain point. And I think this is a prime example of that.
well I for one thank the mathematicians engaging in number theory when there weren't any practical applications. it took a few hundred years for the technology to catch up to where its the reason we can even have a semblance of privacy in the modern age via cryptography.
the lesson there is that we have no way to know what the the abstract useless math today will enable us in the future.
If a prompt can solve a useful math proof, then someone else will find that prompt. In which case, openAI won't get the street cred(wall street) it needs. Nobody is hiding secrets to gain advantage in this environment.
OpenAI cracked these frontier math proofs using models and amount of compute not available to general public. It's not the case that anyone can put "solve Navier–Stokes" in ChatGPT and get solution. Although less well-known problems can be solved with publicly available models.
Not only are the models and compute unavailable to the public, they are unavailable to anyone besides the labs, including the government. We better hope the furthermost tip of the jagged frontier stays in purely theoretical mathematics, because if it strays into the physical world then we may find ourselves in a situation where the power of the labs rivals the governments supposedly overseeing them.
Indeed. Their model they showed off that was ~5x better than 6 Astra in certain domains including mathematics is not public and I doubt they'd release it. They distill their internal large into their public models.
Do you think companies do not hide secrets? I think you are wrong but maybe I do not understand your point of view. However let's see my example.
Imagine: you are a boss of a company and own a powerful tool which can give you a lot of advantages. In short: it allows you to earn more money than competing companies. Would you share this tool to everybody, so that means competing companies could destroy your business or you keep the tool in secret and slowly build empire?
I do not know if OpenAI has powerful LLMs/AI things but I can be sure they do not share everything what they have. I think the only way they could share hidden "powerful tool" is situation where men in black visit CEO and talk with him/her about co-op. Of course CEO is "happy" because he/she can not refuse if you know what I mean. But still hidden powerful tool is hidden but this time another "more important guys" have access to it.
Its more about finding artifically created disabilities. My guess is public key cryptography is solved and the government at least knows. Which is why we had the "export" forbidden noises. Apply a old solution (export controll)to new problem knowledge intrferable.
What's the $-value of some street cred/publicity vs. $-value of finding some optimization trick that reduces training/inference cost by millions compared to their competitors? Of course they're going to keep that IP to themselves.
I think it depends on the situation. If there are profitable applications, they might keep the proof to themselves, but I agree with you in most cases.
The NSA won’t give you that choice.
[dead]
Stupid me thinking they only spent time on brute force-able problems with pure compute, not sure there's much intelligence here. What could a mathematician do with $20M worth of compute to try to brute force one single problem?
My assumption on cryptography was that any published result could potentially help had actors, was unlikely to help good actors, and better to stay away completely. Can definitely see why they'd research but not publish in the other areas though.
[dead]
Probably they have either tried using LLMs to crack most cryptographic systems, or they have strict rules against doing so. Either way, the US govt. will probably make sure it gets a chance to do so before this internal model is released. There's a good chance they have started already.
They would use parallel construction to avoid having to explain the fact?
> In September 2nd, something like that, the biggest conjecture in our field was proved. Did your life change? My life? No. No. No. I mean, the only people that can actually tell you that it does change something are the mathematician who worked on it. The human understanding of mathematics is the only thing that at the end benefits society.
The Strip Mining of Mathematical Proofs (with Hugo Duminil-Copin) - Numberphile Podcast https://youtu.be/omuJpoV-Ko4
Maybe it is just me, but what annoys me even more than the AI-slop texts is the uniform Claude, ripoff visual designs everywhere.
Remember, Trump has said if you don't use SI instead of AI you are an enemy. So, this title should be changed to "The math that OpenSI keeps for itself."
> get something that helps us understand what is going on, and live to see a better future.
Nah, I don't really buy the sales pitch that ultra advanced math will give us a better future. For decades we've had ultra advanced math already before AI like geometric langlands, higher topos theory, and Waldhausen categories and it's done nothing to give us a better future.
Maybe it helps to train people to think abstractly and inspire technical people through the beauty of math but that's only when it was a human endeavor anyway.
The idea of going to even more layers of abstraction and complexity with AI may be possible but it's just a mental game at that point and isn't going to help anyone. It will just waste energy.
As far as I'm concerned (as a former mathematician), they can take their new discoveries and shove it. Yes, science does help us live easier lives in some way but it doesn't follow that endless new discoveries always been better lives past a certain point. And I think this is a prime example of that.
Thinking about giving “us” a better future makes me wonder how many people alive don’t even have access to clean water or electricity.
well I for one thank the mathematicians engaging in number theory when there weren't any practical applications. it took a few hundred years for the technology to catch up to where its the reason we can even have a semblance of privacy in the modern age via cryptography.
the lesson there is that we have no way to know what the the abstract useless math today will enable us in the future.