So I continue to argue that energy consumption as a PRIMARY criticism of AI is going to be unpersuasive, as incentives drive down the energy costs with time, and all critique of the energy costs is just as applicable to conventional cloud computing, which we have already accepted into our social bargain
Just this week, OpenAI launched a new model that's cheaper to operate
Google announced a new chip that's 5x the efficiency of their previous data center processors
so the longer this goes on the clearer it becomes we can't just recycle the cryptocurrency bullshit playbook on AI because the economic incentives are fundamentally OPPOSITE between these two computing approaches
meanwhile, AI poses unique harms that a superficial critique will not address, especially in the face of people USING it
if you build your soapbox on AI being “bad for the environment”
your platform will be eroded quarter by quarter as the technologies are subjected to the benefits of Moore's Law and ongoing R&D that lowers their costs
and meanwhile people will tune you out because, well, we built an entire global economy on fucking the environment, and you’ll need to make clear what's bad about this specific variant of it
@danilo While I agree with your overall point, the Jevons paradox applies. The cost of energy is the primary limiting factor for both AI and cloud compute. As its efficiency goes up, its power consumption won't go down; instead they'll just do more AI, and the total power consumption will almost certainly rise. So arguments about the per operation cost won't hold up, but arguments about total energy and water consumption will. So I guess it will depend on how people think about the total utility of AI versus other energy-using endeavors.
@danilo Sounds about right. And that number keeps rising faster than other consumers of electricity, so it will be equivalent to ever larger chunks of human activity over time. Which is fine, if people consider the utility of those datacenters to be worth it.
And by "worth it" I mean not just the price of the electricity but the externalities that come with it. One of the great things about datacenters from a multinational corporation's perspective is that you have some flexibility in where to put them. Flexibility to avoid, for example, places that do a better job of pricing in the externalities of electricity production. Fortunately, Google and Facebook (and I think Microsoft as well?) have also tried to choose places that have cheap electricity from sources like hydroelectric, but that can still displace other users into dirtier energy sources.
@freakazoid that makes sense to me, and I think my big takeaway from this is that we have to price the externalities of energy consumption PERIOD, and thus the distinction between “AI” and conventional computing is 100% unproductive red herring
AI has specific hazards which should be addressed specifically
Computing of ALL sorts has other hazards, especially energy consumption, which should be addressed to their own specificity
Personally, I never found the energy arguments against crypfocurrency to be particularly compelling. I figured eventually transaction fees would have to pay for all energy consumption due to the transaction, or mining wouldn't be profitable. It's still too soon to tell, but I think we're starting to see that with this latest Bitcoin halving, but it's hard to tell because it seems like transaction fees are driven a lot more by demand than by miners. The total hashrate seems to be dropping, though.
@freakazoid@danilowanders in to catch strays I mildly disagree because I do believe the energy consumption arguments against crypto does work. By design proof of work consumed more energy as adoption increased. To wit one of the early pro arguments was that bitcoin was a “store of energy”.
Proof of stake for sure was meant to burn something other than energy to generate trust. But other arguments against other “proof of’s” exist beyond energy.
@freakazoid@danilo more simpl: the arguments work because “success” for proof of work based blockchains necessitated more and more energy. The argument around energy was strong enough it eventually pushed competing blockchains to ditch proof of work.
@danilo Incidentally, what year is that figure from? Europe's, and especially Germany's, energy consumption have been dropping due to high energy prices. I seem to recall Google had a few datacenters in Germany. I'm sure they'll operate them until the equipment has depreciated, but I wouldn't be surprised if they shut them down and take the loss on the buildings after that, if they haven't already.
@danilo Maybe we shouldn’t tune out because we’ve built an economy on fucking over the environment. Maybe we should actively invest in each other and our future.
@danilo “Writing 3,500 words on the ills of AI” doesn’t represent apathy. It represents sharing information with others about dangers. What is apathetic is telling people not to make certain arguments because you believe them to be futile. As I mention in one of my other replies, calling out power and water issues is important. It matters.
But what I am saying is not just futile but actively counterproductive and self-defeating is making up a Good Kind of Computer and a Bad Kind of Computer and attributing sin or virtue to a watt of energy or liter of water according to which it supports
that’s a separate, necessary conversation of pricing externalities for all resource consumption which we SHOULD do
and which is not at all helped by the special-casing of “AI”
@danilo I agree with your list (esp the industrial theft) but I'd add that the energy issue is also important even if it is temporary... because the near-term surge in energy usage is critically harmful for climate impact. Also: because if energy costs are driven down, usage will only go up (jevon's paradox). So realistically, energy usage will surge in the near-term and then continue to grow in the further-away-term.
@danilo I agree with it as a poor central strategy. I made the same argument early on re: quality, when people were basing their arguments against AI because it was lower quality than human work. If you hinge your argument there, it could fall apart, when the REAL argument is moral (theft; and all the other things you list). I'd say the energy argument is also moral, but you're correct that it's one of many moral components of the argument against AI.
if you make the argument it’s not as good as human labor you’re just inviting the industry to create and game a benchmark for human labor quality (which is exactly what the AI coder startups are busily doing)
We’re going to need folks savvy on the better levers and if we treat this like something we can deride like Juicero, the game is over and the Altmans win
Add comment