Elektrine lite

← Feed

@bradheintz@mathstodon.xyz

2026-08-19 22:52 UTC

@tao@mathstodon.xyz The parenthood analogy has some utility, but I think you're reaching the limits of it. In the case of parenthood, a vulnerable human life is the driver of all the concerns that necessitate all the alternatives to a problematic or unavailable traditional parenting model. This does not apply to mathematical proofs. I'm just a master's student in mathematics, and no one is going to ask me to review a non-trivial proof, at least this week. That said, I have thirty years of experience building software, and I think an analogy to software engineering might be informative. A program does what it is supposed to, or not; a proof is correct, or not. Just as proofs require expert review to be accepted into the mathematical canon, most software teams (whether internal to a company, or a widely distributed open source project) require review by an expert for any change to the software to be accepted. If someone submitted a code patch for me to review, along with the information that they did not understand the code they had submitted, I would reject it out-of-hand without review. In this case, the odds that the patch introduced a subtle bug not covered by tests, or exploited an existing undetected bug to appear correct, or just generated values that passed the test, or had disastrous performance characteristics would be intolerably high. "Intolerably", because I'd be the one debugging what should have been debugged before it came to me, which wastes my time. I am far from alone in this attitude, as many open source projects inundated by superficially correct but bogus AI slop pull requests have started taking a hard line on AI usage and pull request quality. 1/

Replies (1)

  • @bradheintz@mathstodon.xyz 2026-08-19 22:56

    @tao@mathstodon.xyz In the case of a proof submitted by someone who doesn't understand it, I'd have to imagine that there's a radically increased probability that it's bogus. If the proof is formalized, it might be exploiting a bug in a formalization tool (as has happened recently). Expressed in natural language, it might be committing some hard-to-detect error of logic (e.g., employing a theorem without satisfying some subtle but necessary condition). Does it make sense to spend time reviewing a long proof that was much more likely than the mean to send you down a days-long rabbit hole of chasing Lean kernel bugs or invalid implicit assumptions or circular logic? I think you're probably right that new mechanisms are needed. I also think the model we choose to think about the problem will inform the mechanisms we choose to mitigate it. Ultimately, I think it comes down to this: If a work product isn't understod by a (qualified, expert) human well enough to explain it to another (qualified, expert) human, it's not real - or anyway, should not be treated as such. *The value is in the human understanding.* Anyone submitting a work product - software patch, proof, whatever - who doesn't understand it is just throwing up random crap and hoping it sticks. That doesn't mean there's no place for AI tools in mathematics, software, or elsewhere, but it does mean (IMO) that we should set cultural norms around not taking people seriously when they try to submit work they don’t understand, and place the burden on them to bring a product with actual value. 2/2

    Open ##5132043