In the second phase in regards to LLM environmental impact it really depends and can already be mitigated. I’ll try not to make a huge comment because I don’t want to write an essay, but the source’s claims need scrutiny. Everything consumes energy - even we as human bodies release GHG. Going to work requires energy and using a computer for work requires energy too. If AI can do in 10 seconds what takes a human 2 hours, then you are certainly saving energy, if that’s the only metric we’re worried about.

So it has to be relativized which most AI environmental articles don’t do. A chatGPT prompt consumes five times more electricity than a google search, sure, but that amount is close to 0 watts. Watching Youtube also consumes energy, a minute of youtube consumes much more energy than an LLM query does.

Some people will say that we need to stop watching Youtube, no more treats or fun for workers, which is obviously not something we take seriously (deleting your emails to make room in data centers was a huge thing on linkedin a few years ago too).

And all of this pales in comparison to the fossil fuel industry that we keep pumping money into in the west or obsolete tech that does have greener alternatives but we keep forcing on people because there’s money to be made.

edit - and the meat and animal industry… Beef is very water-intensive and polluting, it’s not even close to AI. If that’s the metric then those that can should become vegan.

Likewise for the water usage, there was that article about texas telling people to take fewer showers because it needs the water for data centers… I don’t know if you saw it at the time, it went viral on social media. It was a satirical article against AI, that people used as a serious argument. Texas never said to take fewer showers, these datacenters don’t use a lot of water at all as a share of total consumption in their respective geographical areas. In the US a bigger problem imo is the damming of the Colorado River so that almost no water reaches Mexico downstream, and the water is given out to farmers for free in arid regions so they can grow water-intensive crops like rice or dates (and US dates don’t even taste good)

It also has sort of an anti-civ conclusion… Everything consumes energy and emits pollution, so the most logical conclusion is to destroy all technology and go back to living like the 13th century. And if we can keep some technology how do we choose between AI and Youtube?

Rather I believe investments in research make things better over time, and this is the case for AI too (and we would have much better, safe nuclear power plants too if we kept investing in research instead of giving in to fearmongering and halting progress but I digress). I changed a lot of my point of view on environmentalism when back in 2020 people were protesting against 5G because “microwaves” and “we don’t need it” and I was on board (4G was plenty fast enough) until I saw how in some places they use 5G for remote surgery and that’s a great thing that they couldn’t do with 4G because there was too much latency. A doctor in China with 6G could perform remote surgery on a child in the Congo.

In China electricity is considered a solved problem; at any time the grid has 2-3x more energy than it needs. The west has decided to stop investing in public projects and instead concentrate all surplus value in the hands of a select few. We have stopped building housing, we stopped building roads and rail, but we find the money to build datacenters that could be much greener, but why would they be when that costs money and there’s no laws that mandate it?

Speaking of China they use a lot of coal still (comparatively speaking) but they also see it just an outdated means of energy production that can be replaced by newer, better alternatives. It’s very different, they’re doing a lot of solar and wind - in the west btw chinese solar panels are tariffed to hell and back, if they weren’t every single building in europe would be equipped with solar panels - and even pioneering new methods of energy production and storage, like the sodium battery or gravity storage. Gravity battery storage (raising and lowering heavy blocks of concrete over the day) is not necessarily Chinese but in Europe this is still just a prototype. In China they’re already building them as part of their energy strategy. They don’t demonize coal as uniquely evil like liberals might, but rather that once they’re able to, they’ll ditch coal because there’s better alternatives now.

In regards to AI in China there’s been a few articles posted on the grad and it’s promising. They are careful about efficiency because they have to be. I don’t know if you saw the article from a few days ago about Alibaba Cloud cutting the number of GPUs needed to host their model farm by 82%. The test was done on NVidia H20 cards which is not a coincidence, it’s the best China can get by US decree. The top of the line model is the H100 (the H20 having only 20% of the capabilities) but the US has an order not to export anything above the H20 to China, so they find creative ways to stretch it. And now they’re developing their own GPU industry and the US shot itself in the foot again.

Speaking of model farm… it’s totally possible to run models locally. I have a 16GB GPU and I can generate realistic pictures (if that’s the benchmark) in 30 seconds, the model only needs 5GB Vram but the architecture inside the card is also important for speed. For LLM generation I can run 12B models, rarely higher, and with new efficiency algorithms I think over time that will stretch to bigger and bigger models, all on the same card. They run model farms for the cloud service because so many people connect to it at the same time, but it’s not a hard requirement for running LLMs. In another comment I mentioned how Iran is interested in LLMs because like 4G and other modern tech that lags a bit in the west, they see it as a way to stretch their material conditions more (being heavily sanctioned economically).

There’s also stuff being done in the open source community, for example LORAs are used in image generation and help skew the generation towards a certain result. This means you don’t need to train a whole model, loras are usually trained by people on their machines with like 100 images. As for training time it can be done in 30 minutes to train a lora. So what we see is comparatively few companies/groups making full models (either LLM or image gen, called checkpoints) and most people making finetunes for these models.

Meanwhile in the West there’s a 500 billion $ “plan” to invest in the big tech companies that already have a ton of money, that’s the best they can muster. Give them unlimited money and expect that they won’t act like everything is unlimited. Deepseek actually came out shortly after that plan (called Stargate) and I think pretty much killed it before it even took off lol. It’s the destiny of capitalism to con the government into giving them money, of course they were not going to say “no actually if we put some personal investment we could make a model that uses 5x less energy”, because they would not get 500 billion $ if they did. They also don’t care about the energy grid, that’s an externality for them - the government will take care of it, from their pov.

Anyway it’s not entirely a direct response to your comment because I’m sure you don’t believe in all the fearmongering, but it’s stuff I think is important to keep in mind and I wanted to add here. And I ended up writing an essay anyway lol.