This is either a gift to math, or the start of a very quiet mistake: letting a machine decide which questions deserve human attention.
On paper, the idea behind FAR sounds clean. You start with a broad direction. The system scans the literature for open problems. It tries to make progress on lots of them at once. Then it filters what it finds so only the most promising cases land on a real mathematician’s desk. In a pilot run in combinatorics, public reporting says it narrowed down a pool of open problems to a smaller set for review.
If you’ve ever watched how math actually gets done, you know why this is tempting. Mathematicians don’t just “solve conjectures.” They spend a lot of time choosing which conjectures to even care about. And that choice is messy. It’s shaped by taste, fashion, mentors, fear of wasting years, and what feels doable. FAR is basically saying: stop pretending that selection is some pure art. Treat it like a process we can automate.
I buy the problem. I’m not sold on the solution.
Because the moment you automate selection, you’re not just speeding up math. You’re changing what math becomes. The conjectures that get attention will be the conjectures that score well under whatever filters FAR uses: problems that are easy to find in text, easy to formalize, easy to take a swing at with automated tools, and easy to measure “promise” on. That’s not the same thing as “important.” It’s “machine-legible.”
Imagine you’re a young researcher deciding what to work on. In the old world, you might chase a weird question because you can’t shake it, or because a pattern keeps showing up in your work, or because a senior person shrugs and says, “No idea, but it smells real.” In the new world, you’ll be nudged toward whatever comes out the end of the FAR pipeline. Not because you’re lazy. Because attention is scarce, careers are fragile, and it’s hard to justify spending three years on something that an algorithm didn’t flag as worthwhile.
That shift will reward a certain kind of math brain: the one that is great at executing within a pre-approved menu of “promising” problems. It will punish the stubborn, the eccentric, the people whose best work comes from wandering. And yes, wandering wastes time. But wandering is also how fields change direction.
The defenders of this approach will say: come on, it’s just triage. The system doesn’t “decide truth.” It just narrows the list. Humans still choose. Sure. But anyone who has worked inside a real pipeline knows how this goes. When you get a list of 10 “top” items and 200 “low priority” items, you don’t treat them equally. The list becomes reality. The filter becomes taste.
And the filter is built out of someone’s assumptions, even if they never write them down. What counts as “progress” in an attempted proof? What counts as “worth investigating”? Does the system favor conjectures where small cases can be checked quickly? Does it favor problems that connect cleanly to existing methods? Does it penalize ideas that need a new language or a new viewpoint? Those choices are philosophy wearing a lab coat.
There’s also a social consequence that people will pretend is not there: gatekeeping gets easier. If a department or a grant panel can say, “We’re focusing on problems that are FAR-selected,” that’s a neat excuse to ignore work that doesn’t fit the pipeline. It’s not even malicious. It’s convenient. It’s cover. And convenience is one of the strongest forces in academia.
Now, to be fair, the current system is not some romantic meritocracy. A lot of good questions die because the right person didn’t notice them, or because they were posted in the wrong place, or because they weren’t packaged in the right style. If FAR can surface overlooked problems, that’s genuinely exciting. The same goes for reducing busywork: scanning tons of papers for “open problems” and organizing them is the kind of thing humans are bad at and hate doing. If the tool helps mathematicians spend more time thinking and less time sorting, that’s real value.
But even that comes with a trade. When selection gets easier, the volume of “possible” problems explodes. And when volume explodes, people lean harder on ranking systems. The tool that was supposed to support judgment starts replacing it, because nobody has time to fight the list. You can end up with a field that looks productive—lots of problems attempted, lots of “promising” leads—while becoming more narrow and more risk-averse underneath.
And I can’t shake one more worry: the system might be very good at picking problems that are almost solvable with today’s tools, and very bad at picking problems that force tomorrow’s tools to exist. It might optimize math for short-term wins. That sounds great until you realize that some of the most valuable work in math is exactly the stuff that looked pointless right up until it wasn’t.
So yes, I want mathematicians to have better ways to find good problems. I want them to waste less time. I even want automated attempts at scale, if it helps spot patterns humans miss. But I don’t want the definition of “worth it” to quietly slide from human taste to machine scoring without a fight, because that’s how you get a discipline that moves fast and forgets how to choose.
If FAR gets widely used, who should be responsible for deciding what “promising” means?