The first thing AI thins out is not jobs. It's subject boundaries.

你刚刷到这条消息，本来准备顺手划走，但又怕自己错过了真正会影响下一步判断的那一点。

最容易做错的，是An OpenAI model has disproved a central conjecture in discrete geometry；代价往往是如果只盯表面热闹，你很容易在错误方向上花掉时间、预算和注意力。；我先给一个保守判断：先被AI打薄的，不是岗位，是学科边界。。

'An OpenAI model has disproved a central conjecture in discrete geometry.'

The easy mistake is to file that under math news for other people. The cost of that mistake is real: you spend time, budget, and attention on the wrong lesson.

My conservative take: the first thing AI thins out is not jobs. It is disciplinary boundaries. The part worth not scrolling past is not that AI solved a geometry conjecture. It is that, in OpenAI's account, the key move came from algebraic number theory rather than from discrete geometry's usual toolkit.

That detail is what turns this from a flashy headline into a useful signal. The first result only established a bound of n^(1+delta) for some delta > 0. Will Sawin later made it explicit at n^1.014, and a 9-author paper rewrote the argument into a human-checkable form. Based on OpenAI's May 2026 write-up and arXiv:2605.20579 / 2605.20695, the story is not 'AI got better at one subject.' It is 'ideas crossed a border.'

That is why this feels like an Erdos-conjecture story that died by number-theory trespass. That does not mean 'be broad and shallow.' Those tools are deep, and this is not a license for hand-wavy interdisciplinarity. It does mean cross-domain transfer is getting more valuable than staying perfectly inside one well-named specialty.

真正该讨论的是：An OpenAI model has disproved a central conjecture in discrete geometry