The Generative AI Race Has a Dirty Secret

https://www.wired.com/story/the-generative-ai-search-race-has-a-dirty-secret/

by electic • 3 years ago

18 52 3 years ago

Energy source > Energy consumption.

Energy consumption is not a bogeyman and we'll need to consume vastly more energy if we want to make people richer, healthier and happier.

We need to pivot to cleaner sources, so the focus should be on that.

Would be great if the journos could stop demonizing the wrong thing.

lovemenot5• 3 years ago

> we'll need to consume vastly more energy if we want to make people richer, healthier and happier

I have frequently heard this argument made - often as here axiomatically.

No argument that changing energy sources is crucial. But I would not dismiss a plan to also reduce overall energy consumption while maintaining or improving quality of life indicators.

LewisVerstappen3• 3 years ago

I have frequently heard this argument made - often as here axiomatically.

Okay, you're dismissing the argument without proposing any counter arguments.

J. Storrs Hall lays out the case for the relationship between energy usage & quality of life.

What are your specific rebuttals?

graeme5• 3 years ago

That’s all well and good to say, but we currently have no system to do that.

What we ought to have is a carbon tax. Then if some activity such as bitcoin or ai or driving or flying uses carbon, you pay extra. Whereas if you manage to fuel your activities with nuclear/wind/solar you don’t.

There are loads of complexities to this but simply saying “consumption is necessary for riches” while ignoring that we have not managed to lower our carbon usage is not quite hitting the mark.

kragen10• 3 years ago

we do currently have a system to pivot to cleaner sources; it's the photovoltaic module industry, which has undercut the price of fossil-fuel energy by almost a factor of two in much of the world, though not in polar regions

colechristensen• 3 years ago

Carbon taxes aren’t necessary to give solar and wind an advantage. More than half of ~55GW new power generation capacity in the US this year is going to be solar.

while ignoring that we have not managed to lower our carbon usage

We have in the US, both total and per capita. Carbon per person is down nearly 25% since 2000.

People seem to make up statistics in their heads to match what they think is going on.

albertopv3• 3 years ago

That doesn't change the fact that first and cleaner renewable energy source is consumption reduction.

LewisVerstappen3• 3 years ago

That doesn't change the fact that first and cleaner renewable energy source is consumption reduction.

1) That's not a fact

2) That sentence doesn't make any sense grammatically

3) Consumption reduction = lower quality of life & health for poor people (the rich have 0 interest in consumption reduction... do you see the davos crowd turning in their private jets?)

renewiltord• 3 years ago

This is the abstention-only of climate management.

tomjen3• 3 years ago

Doesn't matter. Humans are not going to accept the the massively sustained decrease in living standards that it would require, which means it is not a politically acceptable solution, and therefore not a solution.

Whereas spending some money to get us all nuclear might be. If not, we will have to spend more to metigate the effects of global warming.

If we really wanted to solve this quickly and efficiently, we could kill all humans.

KarlKemp• 3 years ago

This used to be true, but changed dramatically.

The relevant term is „energy intensity“, measuring how much energy is needed per unit of GDP. It has been improving for quite a while now, see, for example https://www.aceee.org/blog/2015/12/what-energy-intensity-and...

water-your-self• 3 years ago

Until the sources are clean I will blame the consumers that do so without regard for the source.

nl• 3 years ago

I'm 100% in the "let's make clean energy too cheap to meter" camp. Get rid of anti-dumping legislation and let China subsidise as much solar panel production as they can! Build wind farms everywhere! Long distance HV transmission between climate zone should be so common that people don't keep thinking it isn't in use now! (And yes, it would be great if we can get fusion working)

But we aren't there yet. Until we are we need to both increase efficiency and increase renewables.

Google in particular should be praised for their work in subsidizing renewable energy generation[1] by guaranteeing energy purchases from renewable projects[3]. This guaranteed income stream is often what gets them off the ground, and the higher carbon-free energy usage from Google's data centers than the energy markets they are in is great to see[1].

However, that shouldn't preclude journalists from keeping that energy usage in people's minds.

[1] https://cloud.google.com/blog/topics/sustainability/5-years-...

[2] https://www.gstatic.com/gumdrop/sustainability/2021-carbon-f...

[3] https://cloud.google.com/blog/topics/sustainability/clean-en...

jrockway• 3 years ago

I think people are always looking to dismiss emerging technologies. AI is especially scary to many people that read newspapers (and this site); it's going to listen to press releases and write newspaper articles, it's going to illustrate them, and it's going to write the FactoryFactoryInterfaceAdaptor that serves the content to the public. So pretty much everyone involved in this discussion has to fight it or lose their career. (I think that's pretty far away, but I can't blame people for being paranoid here.)

What happens in these cases is that people look for reasons why the emerging technology is objectively bad, and then throw out arguments and see which ones stick. Electricity usage implies climate change, and that's a popular fear to play on right now. "This thing that's going to kill your career is also why hurricanes are stronger now!" Of course, that is only partially true; maybe AI isn't the problem and driving 60 miles a day in stop-and-go-traffic to go to work is what we really need to crack down on. But, that sort of thing is 70 years old and people are comfortable with it, so there isn't much persuasion left to do; people know it's bad and do it anyway. Meanwhile, climate change was a huge problem before the phrases "Bitcoin" or "deep learning" were ever coined, so it's kind of weird to blame them.

This sort of complaining is also easy. Sure, with more electricity users coming online, we worry about the cost going up. The cost going up is good for killing energy intensive activities that you don't like; eventually it won't be profitable to mine Bitcoins or train AIs. But the cost going up also affects other industries that we DO like; poorer people heating their homes can't absorb the cost, or industries like aluminum smelting can't absorb the cost. So now we want to distribute resources with some system other than "the highest bidder wins", and you know what that means! Time to involve the government! But the government is terrible at reacting to things like this, and even though regulation is obviously the globally optimal path forward, it moves too slowly to really make a difference. By the time we charge data centers more for electricity, AI will have taken your jobs and the East Coast will be underwater.

(Meanwhile, many data centers are powered by cheap renewable hydroelectricity anyway and so aren't causing climate change. They can build their AI training data centers near a hydroelectric dam, but people aren't going to move their family so they can get cheaper electricity or spend less energy commuting to work or whatever. I'm not sure where this all leads, but it does feel like appointing a boogeyman to blame instead of actually fixing the problem. This goes back to my opening argument about seeing which argument sticks. If you don't look into things too deeply, you have a nice set of rules that leads to an obvious action. I don't like Bitcoin because the existing system works OK. Bitcoin uses a lot of electricity. Some energy use changes the climate. The changing climate kill us all. So we'd better kill Bitcoin. A simple path from "I don't like it" to "kill it before it kills us". Get out your pitchforks!)

bee_rider• 3 years ago

ML training also seems like it ought to be fairly dispatchable, right? Pause the training if there isn’t excess wind or solar. We’ll want to overbuild on renewable power capacity, so that if it not at peak, the lights can stay on. Having a use for that excess power seems like it ought to help economically incentivize the building.

yowzadave• 3 years ago

Has someone already made a cryptocurrency that requires that miners train an AI to demonstrate proof of work?

NBJack• 3 years ago

sigh Web4 hype train, here we come....

deafpolygon• 3 years ago

We're not done with Web3

brucethemoose22• 3 years ago

I kid you not, I read a scheme such a thing a year ago.

But now I cant find the source... video encoding mining seems to have gained more traction :/.

janalsncm• 3 years ago

Could be interesting. One way to do it would be to have a training set and a hidden test set, and mining a coin means arbitrarily high performance on the test set. In order to keep the test set hidden from miners (otherwise they could just memorize it) they would need to upload their weights to be validated.

The main issue I see is that many of these models are massive. Dozens of gigabytes. You’ll quickly DDOS your validator nodes before getting anything useful.

siliconc0w• 3 years ago

I kinda hate these articles about the carbon cost of compute- it's utterly short sighted, gpt3 is infrastructure that creates huge efficiencies - the cost is a bargain. Not to mention this is like early days and we're driving down the costs everyday.

palata8• 3 years ago

Sources for the "creates huge efficiencies" part?

lovemenot5• 3 years ago

Isn't it the case that training a ML model consumes vastly more power than using the trained model for inference?

If so, the problem may be somewhat overstated. Training is a dispatchable power load. It can be scheduled for periods of surplus power in regions where surplus power is periodically abundant. Training a ML model could literally follow the sun.

brucethemoose22• 3 years ago

At Big Tech scale, the same model is used for inference so many times that most of the energy is probably from inference.

palata8• 3 years ago

Note that if you have surplus of power, maybe you want to use it more wisely :).

lovemenot5• 3 years ago

Perhaps, you are right about utility, but I can think of few power loads as well suited to dispatch.

Bauxite has to be shipped to Iceland to take advantage of their abundant geothermal and hydro power for Aluminium smelting. Data moves with much less friction.

The Generative AI Race Has a Dirty Secret

Related Stories