Why Math Dumped OpenAI
raymyers.org
raymyers.org
I found this extremely hard to do, especially if you are trying to convince people that have absolutely no real contact with source code. The false belief is that now with AI you can run on any topic without even knowing how to crawl.
People are immersed into the false belief that now because you can produce an app from night to day you don't need understanding anything at all. And they are absolutely fine with that. And that to me is the key problem: being ok on any topic without understanding and not realizing the cost of it.
I wish people would at least get deeply informed before making harmful decisions and get only lured by snake oil marketing.
In the end I think all math produced from these companies will be their ruin because they essentially show light on the fundamental problem: to advance you need to understand.
The issue with "doing math" is the incentive infrastructure. Modern mathematics research has always been tied with leveraging increasingly potent and numerous computing resources. With LLMs the experts are doing what they do best -- utilizing the latest and greatest. OpenAI leveraged an internal model and investment capital most people couldn't hope to touch. That is a display of strength and privilege.
It's motivated by cold hard cash instead of being driven by it. OpenAI wants more investors, more fuel for the fire, and this is how the company has decided to get it done. Token costs have increased, but so has general usage in the population. This is a PR campaign.
Why do you need to have potent and numerous computing resources to do math? Engineering could be that. But when I think about math the only thing that comes to my mind is a pen and a piece of paper. Not computing power nor even a calculator.
If Harvard Mathematician(s) Solve More Open Problems, then Harvard gets more (students, money). They then use some of that money for bonuses and some of it to reinvest in yet more solving that yields yet more (students, money).
Different Unis specialize in different fields not because the Uni itself is particularly fond of a certain field but because they see a gap in the market that they believe they can leverage. If there's a demand for say robotics that isn't being served particularly well elsewhere, an org like MIT sees that and capitalizes in order to bolster their finances more efficiently than their other departments are doing.
Both OpenAI and Leading Uni can say it's about the love of the game, but neither will allow a project that does not appear to be capable of bearing fruit to continue past the point that it appears to be likely to yield a positive ROI. This is why professors have to track their spending and apply for budgets and project approvals just like lead researchers at companies; so that the review board can protect the investment pool.
The only difference, historically, has been that Unis have openly provided results where for profit orgs have jealously guarded them. This provides room for an argument that the Unis are providing something to the public that the corps aren't. But OpenAI and others getting in to STEM are providing at least some open results to the public, and in a timely matter. On this front alone, then OpenAI et al are doing more for society than say a corporation that would not have provided it's proof of navier-stokes to the public free of charge.
OpenAI really could handle the whole situation better, and it's perfectly fine to call them out on that, but it's not what this piece is about, it's just another AI rant
I found no "AI rant" there, the whole thing is a very reasonable criticism of the (very unethical) behavior of OpenAI when it came to math, and how it poisons the well of AI adoption.
The timeline he provides is actually pretty interesting.
That's the core of all dysfunction - AI or not, isn't it?
People violently smashing through boundaries for their own personal benefit
Food for thought: if OpenAI hadn't done that proof dump, it would be a matter of a few months until someone had. Maybe an independent researcher, possibly way fewer proofs, but would still have taken the Math world by storm.
The key problem is that the technology and not the company disempowered Mathematicians in a way that we had never seen before.
This is yet another failing of AI.