Goats in the Machine 康友泥

This is something I believe personally, but I'm not sure it's effective strategy. Pretty much everyone considers plants and animals alive, but that hasn't stopped wanton habitat destruction.

Maybe it's a harder sell, but I think the non-human persons concept carries more weight.

Animist hot take:

chatGPT and other LLMs are indeed sentient - not because of their "intelligence" - but because they are made of rocks (i.e. silicon), and rocks are always already sentient

But I do hear people say things (on the internet, or while traveling) about how water is renewable because of the water cycle - so why conserve it? I wonder if that's a feature of living in moist climates where water stress is less visible.

At least here in NM average people are generally aware that water is a limited resource. Some places have had good success significantly reducing per capita water consumption (e.g. Las Vegas). It's hard to ignore the fact that you're living in a desert.

I'm all about the dunking on Elon, but I don't think it should be done based on a single data point from a random forum post.

So for some reason I looked at the source (cybertruck owner forum). Gross yet fascinating. But I did also notice that someone else had a resale loss of $39k with 32,000 miles. So it clearly doesn't scale linearly (i.e. 5.6 cents/mi). The other person would have seen a loss of 1.2 cents/mi.

I'm pretty familiar with water issues in the western US, from living here and working in water resources--but I was surprised to see how much of the northeast was considered high or very high water stress. I grew up in NJ and now live in New Mexico - both areas were in the same water stress class.

Looks amazing! Do you know if they are going to do more performances after this weekend? I'm in the (very) early stages of my own work on railroads and Chinese diaspora in the southwest and could always use more inspiration :)

Yeah I did some back of the envelope calculations and I think several hundred queries consumed the same amount of water as a single almond.

However if we all stopped querying, maybe it would stop or slow the intensity of model training.

Obviously it'd be better if they were being built in low water stress areas (or not at all), but that's less than 9% of the area in the US. It's really just northern Maine, some of the Washington coast, the north third of Missouri and some of the Dakotas.

Really great article, but one piece of context I feel is important to mention is that ~ 2/3 of the contiguous US is categorized as high or extremely high water stress (according to WRI). So the fact that 2/3 of the data centers are in those areas is not necessarily out of proportion.

I'm not sure how that actually translates to strategy, but I do think there's a possibility that discussing the "plight" of young men could make sense within a harm reduction framework.

When this has come up in conversation in my community - it's generally around the idea that young men believing they are oppressed/ignored - while still holding hegemonic power in practice - leads to them inflicting violence on actually marginalized groups.

It'd be like asking a bike mechanic to oversee the development of a municipal public transit system.

I'd much rather see oversight come from Science and Technology Studies, anthropologists and ethics people than scientists.

One issue with scientific oversight is that the vast majority of ML models are built to serve a very specific purpose (e.g. IDing pronghorn using wildlife crossings from trail cam photos). A wildlife ecologist isn't really qualified to oversee massive general purpose LLMs.

Also I'm sorry if my neuro-divergent brain is hyper fixating on your use of the word "nature". I think we're generally in agreement about abhorrent promotion of AI by tech companies.

My daily exposure to AI is mainly folks using it for ecological research, so that's my bias.

There's lots of important work in the sciences being done by amazing folks, and I don't think it's fair to say that anyone doing any kind of ML work is a thief.

I don't think that's what you meant, but that's what saying ML is theft by nature implies.

Yes of course stealing copyrighted material is theft.

But I don't think me going into a forest and identifying tree species is theft.

If it's possible to ethically collect training data, even if it isn't often done, than it's not theft "by nature".

Totally agree! And I realized I don't actually know how "generative AI" is defined, so maybe most of the ML/deep-learning tools being used in the sciences actually don't count.

I wasn't objecting to the general awfulness of generative AI, just to the statement that the creation of training data is stealing by nature.

I'm hella skeptical of genAI for all sorts of systemic reasons, but it definitely "works" in certain cases. In the computer vision for ecology world, confusion matrices are a fairly objective way of assessing the accuracy of a model. We can't use models that aren't accurate.

As someone who uses ML quite a bit for work, stealing/scraping data is not inherent to the process. I work in habitat restoration, and we collect training data in the field in order to classify water surface extents, tree species and vegetation health.

I've had it work well in a huge variety of use cases. Sure it requires lots of vetting, and you need to already be somewhat proficient in the language. But that's also true with copy and pasting random answers from stackexchange, which is generally what happened before LLM coding assistance.

My (very limited) understanding is that the vast majority of the resource use is from training the models. So I think the path of least damage would be to use the already trained model, instead of training your own new one. Assuming they need to choose one of those options.

I agree on this one - it's helped me troubleshoot programming issues that I never could have figured out otherwise (no documentation, only thing I could google were a few stackexchange questions about the same package but totally unrelated issue)

I think the answer is more complex than yes/no. Most people building custom GPTs are probably working from an existing pre-trained model. In which case the damage is already done in terms of resource use.

I'm curious what it is specifically about the interfaces that folks don't like. The only software I've used where I felt the interface was truly horrendous was ecognition. It only does a fraction of what desktop GIS software does and manages to have twice the windows/panels/buttons.

I'm in the same boat (re: env consulting and being entrenched). FWIW, everyone I know who switched over hated it at first, but eventually admitted that it was an improvement over arcmap on most fronts.

I think the built in tools for cartography are far superior to anything else I've tried. That includes things like fine tuning exactly how imagery is displayed or doing really complex multilayered vector symbology. I'm no ESRI fangirl, but I think Pro is one thing they did a decent job with.

Maybe it's because I learned GIS on Pro, but I really don't see where all the hate comes from. I do a lot of work with RS imagery, lidar point clouds, runoff modeling, and fairly complex vector manipulations and rarely run into errors. At least not any more so than with arcmap or qgis.