A surfboard is also an amazing tool, but there's more to operating one than telling it which way to go.
Many people want self-driving cars so they can drink in the back seat watching movies. They'll find their jobs replaced by AI, with a poor quality of life because we're a selfish species. In contrast Niki Lauda trusted fellow Formula 1 race car driver James Hunt to race centimeters apart. Some people want AI to help them drive that well. They'll have great jobs as AI evolves.
Gary Kasparov pioneered "freestyle" chess tournaments after his defeat by Big Blue, where the best human players were paired with computers, coining the "centaur" model of human-machine cooperation. This is frequently cited in the finance literature, where it is recognized that AI-guided human judgement can out-perform either humans or machines.
Any math professor knows how to help graduate students confidently complete a PhD thesis, or how to humiliate students in an oral exam. It’s a choice. To accomplish more work than one can complete alone, choose the former. This is the arc of human evolution: we develop tools to enhance our abilities. We meld with an abacus or a slide rule, and it makes us smarter. We learn to anticipate computations, like we’re playing a musical instrument in our heads. Or we pull out a calculator that makes us dumber. The role we see for our tools matters.
Programmers who actually write better code using AI know this. These HN threads are filled with despair over the poor quality of vibe coding. At the same time, Anthropic is successfully coding Claude using Claude.
Can you share more about your architecture & process? Also a researcher involved in math research (though not strictly speaking a mathematician, but I digress). I've often thought about using AI on my notes, but they are messy and even then I can't quite figure out what to ask: prioritization, connecting ideas, lit search, etc.
I'd love to hear what you do.
There is definitely a gap in academic tooling, where an "association engine" would be very useful for a variety of fields (and for encouraging cross-pollination of ideas between fields), but I don't think LLMs are anywhere near the frontier of what can be accomplished with a given amount of computing power. I would expect simpler algorithms operating over more explicit ontologies to be much more useful. (The main issue is that people haven't made those yet, whereas people have made LLMs.) That said, there's still a lot of credit due to the unreasonable effectiveness of literature searches: it only usually takes me 10 minutes a day for a couple of days to find the appropriate jargon, at which point I gain access to more papers than I know what to do with. LLM sessions that substitute for literature review tend to take more than 20 minutes: the main advantage is that people actually engage with (addictive, gambling-like) LLMs in a way that they don't with (boring, database-like) literature searches.
I think developing the habit of "I'm at a loose end, so I'll idly type queries into my literature search engine" would produce much better outcomes than developing the habit of "I'm at a loose end, so I'll idly type queries into ChatGPT", and that's despite the state-of-the-art of literature search engines being extremely naïve, compared to what we can accomplish with modern technology.
I also agree that neural net LLMs are not the inevitable way to implement AI. I'm most intrigued by the theoretical underpinnings of mathematical proof assistants such as Lean 4. Computer scientists understand the word problem for strings as undecidable. The word problem for typed trees with an intrinsic notion of induction is harder, but constructing proofs is finding paths in this tree space. Just as mechanical computers failed in base ten while at the same time Boole had already developed base two logic, I see these efforts merging. Neural nets struggle to simulate recursion; for proof assistants recursion is baked in. Stare at these tree paths and one sees thought at the atomic level, begging to be incorporated into AI. For now the river runs the other way, using AI to find proofs. That river will reverse flow.
While CoIC has recursion "baked in", HOL does not. It turns out that we can treat structural recursion as a derived property, even over coinductively-defined types. We don't even need a notion of ordinals for this! (See https://www.tcs.ifi.lmu.de/staff/jasmin-blanchette/card.pdf and https://matryoshka-project.github.io/pubs/bindings.pdf.)
I still think you're overgeneralising. What actual thing does your poetic tree / thought / river analogy correspond to?
"Throw a massive neural network at it" is an extremely inefficient way to get results, and doesn't generalise well – for instance, there's no easy way to get online learning for a transformer model, whereas that capability just falls out of most search engine database systems. (The underlying relational database engines had a lot of work put in to make online CRUD work reliably, but that work has been done now, and we can all build on top of it without a second thought.)
It seems likely that PhD students in the subfields of the authors are capable of solving these problems. What makes them interesting is that they seem to require fairly high research level context to really make progress.
It’s a test of whether the LLMs can really synthesize results from knowledge that require a human several years of postgraduate preparation in a specific research area.
> the answers are known to the authors of the questions but will remain encrypted for a short time.
Ok. But humans may be able to solve the problems too. What prevents Anthropic or OpenAI from hiring mathematicians, have them write the proof and pass it off as LLM written? I'm not saying that's what they'll do. But shouldn't the paper say something about how they're going to validate that this doesn't happen?
Honest question here. Not trying to start a flame here. Honestly confused how this is going to test what it wants to test. Or maybe I'm just plain confused. Someone help me understand this?
Yep. "possible but unlikely" was my take too. As another person commented, this isn't really a benchmark, and as long as that's clear, it seems fair. My only fear is that some submissions may be AI-assisted rather than fully AI-generated, with crucial insights coming from experienced mathematicians. That's still a real achievement even if it's human + AI collaboration. But I fear that the nuance would be lost on news media and they'll publish news about the dawn of fully autonomous math reasoning.
This is what is special about them:
> a set of ten math questions which have arisen naturally in the research process of the authors. The questions had not been shared publicly until now;
I.e. these are problems of some practical interest, not just performative/competitive maths.
And this is what is know about the solutions:
> the answers are known to the authors of the questions but will remain encrypted for a short time.
I.e. a solution is known, but is guaranteed to not be in the training set for any AI.
Not a mathematician and obviously you guys understand this better than I do. One thing I can't understand is how they're going to judge if a solution was AI written or human written. I mean, a human could also potentially solve the problem and pass it off as AI? You might say why would a human want to do that? Normal mathematicians might not want to do that. But mathematicians hired by Anthropic or OpenAI might want to do that to pass it off as AI achievements?
Of course a math expert could solve the problems themselves and lie by saying that an AI model did it. In the same way, somebody with enough money could secretly film a movie and then claim that it was made by AI. That's outside the scope of what this paper is trying to address.
The point is not to score models based on how many of the problems they can solve. The point is to look at the models' responses and see how good they are at tackling the problem. And that's why the authors say that ideally, people solving these problems with AI would post complete chat transcripts (or the equivalent) so that readers can assess how much of the intellectual contribution actually came from AI.