Join the discussion
Write your take first — we'll ask for email only when you're ready to publish.
- Hacker News
- In chats with models like deepseek the LLM makes many basic math wrong claims. Maybe more advanced models can in special circumstances, guided by the work of humans and Lean, prove theorems. But just like in RL, the brute force approach of failing a million times may not be enough when the reward is far.
In math the reward is finding the decomposition that reduces the problem to simpler cases, and in many cases that reward is very far. Furthermore, the LLM may process the many cases fast, but the decomposition often comes from a human, and the model takes the credit for the proof.
This foundation could provide the LLMs with the decompositions, generated by human ingenuity, that they need. But once supplied, the model can run the cases and get the credit. Hence more credits for LLMs and fewer jobs and worse prospects for humans, even if it isn't an intended consequence.
by Trusteando - Nearly everyone rejected the idea of a CERN-like project for mathematics when it was recently proposed. It should have been more seriously considered.
Now that things are settling a bit after the Navier-Stokes and theorem-dump shocks, something like what mathematics will look like post-AI is starting to emerge. Of course, things should be conditioned on the assumption that capabilities will continue to improve, but imagine something like this--
An institute with a tremendous amount of compute, either rented or owned, built on the Humboldtian model of the unity of teaching and research. The main draw for big names would be the access to compute, free of the fear of being scooped. New forms of attribution and collaboration would develop, and we'd gain some clarity for students and PhDs on what a future career looks like.
The current melange of math university departments all following their own paths through the crisis without consolidating resources makes the transition more difficult. Think of the path physics took beginning with scattered cyclotrons (the basement of the Niels Bohr institute was even dug out to put one in), then the larger labs that followed. That transition was gradual because physics led it from within, whereas with mathematics every university department is suddenly at a massive compute disadvantage with respect to labs that aren't following the same model, and are driven by different incentives.
It could be this simple--maybe we don't need to completely throw away the old model, attribution practices, and so on, and maybe we can preserve the transfer of knowledge to the next generation of mathematicians, if we simply give them lots of compute.
An analogy to physics to highlight how the divergent interests of the labs could be managed in a state of "peaceful coexistence" might be: physicists got well-funded accelerators and labs that followed standard academic practices. These were separate from "war work" (Los Alamos) that followed government priorities, even with the crossover of personnel. You could have the same thing with a "math accelerator" working on the academic model, while the AI labs follow their own priorities, and some subset of people cross between.
It might be objected that this is merely a conservative attempt to preserve the existing academic conventions and practices currently under threat, but tenure itself already does that, and is arguably the very thing that exacerbates the problem for young mathematicians.
by defgeneric - I really can't empathize with treating mathematicians as some sort of AI refugees who need coddling and support. AI is a tool, and competent mathematicians should learn to use it like everyone else. I find it hard to believe that AI doesn't represent a net potential benefit to human understanding of mathematics. Yes it can be misused, but any tool can be misused. Just don't misuse it...by gradus_ad
- My effort to help scientists navigate the AI era:
On ChatGPT: Settings -> Data controls -> Turn off “Improve model for everyone”
On Claude: Settings -> Privacy -> Turn off “Help improve our AI models”
They can scrape your work after you publish it.
by an0malous - So they want mathematicians to keep prompting and giving feedback to the models for clarifications on their proofs. Then write it all up and publish it in a journal and mention their AI usage. Then suddenly the models end up producing readable and understandable proofs through the "emergence" of the skills because of the scaling laws.
- > > Under a partnership, the Simons Foundation, Anthropic and OpenAI will commit $15 million in combined funding and resources.
OAI and Ant got blowback for essentially strip mining mathematical discoveries into PR wins for their models, accidentally laying waste to the research field in process. Throwing ~$5M each into PR whitewashing (partially in token credits), seems self-serving.
Kudos to the Simons Foundation for seizing the moment to offer the labs a fig leaf of respectability. Might have been the fastest "Yes" on a grant proposal ever.
by mrandish - > Under a partnership, the Simons Foundation, Anthropic and OpenAI will commit $15 million in combined funding and resources. The foundation will manage and award the funding, including through an open call for proposals, to support workshops and initiatives that help researchers evaluate recent AI-assisted results, learn and use new tools, share perspectives and shape collective standards for AI use.
The foundation was started in 1994, Wikipedia says "one of the largest charitable organizations in the United States with assets of over $5 billion in 2022" - Jim Simons founded the Renaissance Technologies hedge fund.
by simonw - I think a lot about how Terrence Tao and the usual 30 (already well-funded) math guys will receive this cash while us early-career math students are left holding the bag.
It's akin somewhat to "donating 100 dollars to the node JS foundation, not the packages that appear when you run npm fund".