Hacker News Reader

Comment by drodgers | Hacker News Reader

The best part? Weeds will still evolve resistance — just like they do for herbicides — but they'll do it by growing into shapes that the ML models misclassify.

View on HN

Updating the robots can be done much faster than it takes for the weeds to mutate. And for mutations to happen, some proportion of the weed population must survive in the environment, as no weeds can not mutate. I doubt they have a chance, unless some species figure out how to live under the earth, in which case, we won't care as those would not be "weeds" any more.

I expect John Deere copyright lawsuit will have a big impact on how free one will be to update the models. I can see weed updates as a form of IRL DLC

Yep. John Deere will collect all of the recognition data and not allow their customers to see anything except high-level statistics (for an additional fee). Then they'll use it to train next year's model and charge for the upgrade.

I'd wager that some will survive by:

1) Getting very close to the crops and intertwining with them to make weeding very hard

2) Looking almost exactly like real crops (as-in, to the point where most humans can't tell the difference and the machines need to start taking genetic samples)

You're totally right though; the machines have an adaption-speed advantage, so I'd expect this to be a very effective technique overall (assuming that the robotics pan-out).

Are we writing a vegan Jurassic Park reboot ?

No, this is more like Terminator. The weeding machines will eventually see humans as a new kind of weed, then it's all over for us.

Give it a few decades and the weeds will find a way of corroding robots. In a century, we'll have to develop robots that can handle the acid that plants will be spraying on them.

Until the weeds become actually identical to the actual crops, and we just have weirdly recreated the crops. This is hilarious, and actually precisely how Generative Adversarial Networks work (https://www.wikiwand.com/en/Generative_adversarial_networks), although obviously with gradient descent in the generator network, not biological evolution

Yet dandelions have genetically modified for lawn mowers, growing strangely mutated so as to be flat on the ground, yet the mowers haven't changed. So, sometimes the technology is not as fast as we think.

Source? If you cut the stems the plants regrow shorter, but next year if you wait a while the stems grow just as long.

My source is simple observation. I did all the mowing as a kid, acres every week. In the spring, chopping down all the new dandelions was satisfying and the grass looked great. By the middle of the summer, dandelions were back but now they were strangely 'smashed' ones that grew no taller than 4", with flattened stems that turned or curled or flowers on short, short stems that bent over.

I'm sure that gene was always there; but billions of mowings over a century of lawn care means its present everywhere. Simple selection?

Ahh, no this is similar to the process that causes Bonsai to work. Dandelions have a fairly massive root with plenty of reserves to regrow after something eat's the top. Mowing simply causes similar responses as a goat would.

Plants can't move around so they evolved to deal with a really wide range of environments and stuff eating them etc.

Nah. Dandelions don't regrow stems; this can't be something done in response to being mowed. The plant doesn't think "I got mowed! I guess I'll grow a different way now".

Its a different growth pattern in the plant. Gotta be driven by genetics.

Dandelions are actually perennials so they will regrow stems several times, and survive multiple years. They will also tend to flower most abundantly in spring, but can re-flower in the fall.

Further, it's not just the stem involved, if your mowing regularly your taking out the leaves.

PS: There are also a few different plants that all look similar. https://en.wikipedia.org/wiki/False_dandelion

Yup the flat ones were none of those. They were really weird knarly mutated plants with exactly the same structure, leaves, stem and flowers as dandelion but grown flat and warped to hug the ground.

I've always wondered about dandelion evolution and mowing. Do you know of any studies done on this?

No! And it needs to be done! I've lived in only a couple of places, and the genetic variety of dandelion (dent-de-lion?) is marvelous. When I grew up they were small modestly sized plants. Then the mowing thing, and a 2nd form of flat plants. Moved to California, they were 4' tall with greyish-green dusty leaves. Moved to an acreage in Iowa, all over the fallow field one year they were bushes about 18" high with 100 blooms - you could have sold them in a garden shop for landscaping!

Mendel should have used dandelions for his gene studies, there is lots more to see!

Meh, there are plenty more cracks in sidewalks, open fields, and poorly kept yards than there are mowed lawns. Dandelion evolution probably isn't optimizing for mowed lawns any time soon.

It's apparently not a crazy idea.

https://news.uoguelph.ca/2014/04/high-school-students-study-...

I find it difficult to believe that any optically-driven machine-learning weeding robot will be 100% successful at eliminating every weed in a garden. There will be some overlooked and they will mutate.

Are you forgetting that you're talking about two different processes that operate on orders-of-magnitude different durations?

By the time any evolutionary affects are seen, if at all, of weeds being pulled by ML, ML will have progressed far beyond where it is now, likely well beyond what could be fooled by simple diversity.

There's already evolutionary advantage for weeds to look similar to crops.

DL models can learn to classify hundreds of weed classes in less than a week. It takes centuries for a new breed/trait to appear. This is a very uneven race.

Reply on news.ycombinator.com