I have no idea why this is downvoted because it's exactly right. Unlike antibiotic resistance where the consequences can be measured in human lives, it just doesn't matter for weed killers: and the iteration time on new compounds is much faster.

It's also inevitable: there are weeds which have substantially changed their appearance to more closely resemble crops as an adaptive strategy just to human driven control measures: https://en.wikipedia.org/wiki/Vavilovian_mimicry

Which is a problem which mechanical weed control measures will exacerbate probably in bizarre ways (e.g. the weed is no longer selecting against the human vision system but instead a machine vision model)

Edit: though probably worth noting that encouraging weeds to compete against a machine vision model opens up interesting possibilities - e.g. encoding a failure mode for something which the active model can't spot, then running it competitively against a model trained to sport the adaptation and then switching back over when your hit rate falls below a certain level - trap the weed in a controlled local minima. You can't replace human image recognition and new compounds are hard, but updating software is easy.