Take the first sentence.
> As recently discussed on The Carbon Copy with Brian Janous, utilities are seeing major forecasted demand growth for the first time in decades, and almost entirely from data centers.
Utilities aren't seeing major forecasted demand growth for the first time in decades. It isn't entirely from DCs.
The author appears to have a startup, pitching a Power to NatGas plus direct air capture system because "Why not?™" [0]. I'm offended on behalf of chemical engineers everywhere.
[0]https://terraformindustries.wordpress.com/2023/06/26/the-ter...
Should we as humanity stop right here? Draw a line in the sand and not do anything new? Maybe we should move backwards, skipping the Industrial Revolution and go back to each individual growing their own food?
This is a serious question. What do you propose? Because you are not able to pick and choose what humanity tries out, so shall we stop entirely?
> Because you are not able to pick and choose what humanity tries out, so shall we stop entirely?
I don’t see anything that dictatorial in their suggestion.
There’s clearly something wrong in the current LLM based ecosystem. They take gigawatt-hours to train and digest the entire corpus of the internet to produce a model that writes at, what, the level of an erudite college freshman?
> not focusing on ever increasing the size of ai models, so that we don't need to build all those power-hungry ai training datacenters?
I read the italic bit as not a command to stop, but a suggestion to come up with better algorithms. Which researchers are presumably working on. Hopefully chatGPT & friends don’t suck up all the oxygen.
I think these kinds of things will be solved by market forces.
And there's no reason to ever be bullish on liberal (classical) society's capacity for intentional change.
Human agency doesn't mean shit.
Maybe someone with the knowledge of hindsight in a few generations will say that, yes, we should stop and draw a line in the sand.
We are still used to the idea that we can treat externalities to the systems we build as infinite and boundless, while it is getting increasingly clear that they are not. I am not saying we should stop working on LLMs. I just say we probably should factor in (that means: assign economic value to) downstream consequences of said high energy consumption if we don't wanna destroy our own habitats. And that then would probably lead to a world where LLMs are used for actually important problems instead of furthering the goals of the few who truly profit from surveillance-capitalism.
But hey, let's continue wasting a years worth of energy on training LLMs on cat pictures, they are cute after all.
There will be a point when the answer is "yes".
For example, when we can produce hobby 3d printers for biological viruses. Or DIY kits for nuclear weapons. Or cameras that are the size of a grain of sand and cost only 1 cent. Stuff that some individuals might want, but society will never be ready for.
The argument against current AI progress isn't anti-technology, it's pointing out that these things can have a hugely negative impact that's mostly being shouted down or swept under the rug in the name of "progress". The climate crisis is pretty serious business, and the AI power consumption is making it worse, not better. The risk of huge amounts of unemployment are being mostly handwaved away as "yeah, that's not gonna happen, we always make jobs out of somewhere".
Just yesterday at work I asked it for a Powershell script to upload a local file to S3 and it hallucinated a method. The whole script is like 10 lines. How could it mess that up, but meanwhile it’s about to change the world and be mass adopted?
This is Claude 3 the best model for coding…
After using these models for awhile I think they’re really helpful to get off the ground in terms of coding and writing and to help review coding and writing, but that’s about all I’ve seen so far. It feels like we’re also getting diminishing returns with the current paradigm so we’ll see if I’ll be eating my hat next year, but I really don’t see mass adoption like a search engine.
The only reliable long term growth curves so far are world population and global average temperature.
Unfortunately the chips are all to expensive so everybody wants to run them full-out all the time.
DC are run by companies with a shitload of money.
Not a hard problem.
And we can use the heat to heat houses too!
If AI training is being that power hungry, maybe we shouldn't be doing it ... right now.
It's all fueled by speculative VC funding. This isn't a slam dunk economics calculation, certainly no more than regulation to shut this nonsense down would be.
"Wafer prices stable-to-soft on market oversupply"
https://www.pv-magazine.com/2024/03/15/wafer-prices-stable-t...
"Wacker Chemie’s sales, earnings fall in tough market"
https://www.pv-magazine.com/2024/01/30/wacker-chemies-sales-...
"China polysilicon prices fall 51.8% year-on-year amid supply glut"
https://www.pv-magazine.com/2024/01/19/china-polysilicon-pri...
"China solar cell prices decline on sluggish downstream demand"
https://www.pv-magazine.com/2024/01/05/china-solar-cell-pric...
What we need the most is better storage technologies. PV, wind and hydro are highly variable which means you need something to cover the gaps between production and demand. Gas and coal plants are actually super useful for this because you can literally scale production up or down as needed by increasing or reducing fuel. If you want to eliminate gas and coal, you need a way to store excess energy and release it on demand - and to be able able to do so at the same scale as with fossil.
And nuclear power is actually worse in this regard because you effectively can't turn a nuclear power plant on or off. If it's on, it must remain on because taking it offline can take days and the same goes for turning it back on. But the problem right now is not lack of production but stability. And stability can only be achieved with better storage technologies - or fossil fuels.
I wouldn't use what Bill Gates is investing in as any indicator of what we should be doing.