Felix is right. And it goes further: We often treat units of energy used or units of compute supplied as fungible, as if they were all the same. But it matters what you do with them, not every use of energy for example is the same. It makes a massive difference whether you use a kWh of energy on running a train, powering an MRT , confirming Bitcoin transactions or generating some "AI art".
Energy and compute are often seen as ethereal, as invisible but they have a material basis, a material cost. And we more than ever need to make sure to invest our material resources less wastefully, more consciously and with a stronger focus on the well being and dignity of everyone on this planet.
@tante A nice idea that sounds impossible to implement.
If I run a Python script on one of my machines, who is supposed to check if it's a media player for my RPi, or a Bitcoin miner, or if it trains a language model? And how would you check it? And who will adjust the price of consumed energy accordingly?
It would also discourage people from building open models. If I have to pay a premium for training my model, I'd rather keep it for myself, so I can reap the benefits and make up for the initial higher costs, rather than giving it away for free to people who haven't paid the same initial costs as me - thus putting myself in a position of competitive disadvantage.
Also, "AI needs to cost more" will have so many bad repercussions. Researchers of all kinds use all kind of AI for protein folding models, text analysis, climate predictions etc., and probably those shouldn't be taxed as much as the guy who uses Dall-E to generate memes.
I think we need to stick to "computing neutrality" just like we stick to network neutrality, because the drawbacks and implementation hurdles far outweigh the benefits.
And maybe, instead of going so granular, a tax on the overall amount of energy consumed by a data center that exceeds the expected baseline would be more effective. It taxes directly the metric that we want to optimize (wasted energy) rather than the specific application, and it puts the burden on data centers (instead of individual developers and researchers).
Bonus: data centers may just offload the costs to customers if they are renting the hardware, or try and minimize them if they're building their models in-house. That would give the right incentives (which are currently missing) to build more efficient algorithms - it'd make PoS much cheaper than PoW without further regulation needed, and it'd drive people to train AI models in more efficient ways.
@fabio If AI really has the benefits claimed, paying some extra would be worth it. It's just a wasteful tech that should be
disincentivized to make "fucking around" no longer worth it (which would wipe out most of our current "AI" industry.
@GratianRiter that was the Crypto argument as well. The problem is that this renewable energy then isn't available for other things so we have renewable "AI" but dirty hospitals.
@tante One of the many problems with contemporary tech culture is that it takes the amount of energy used as a sign of civilizational progress, meaning using more energy is a good thing in and of itself.
This is, ultimately, based on the Kardashev scale, a method of measuring a civilization's level of technological advancement based on the amount of energy it is capable of using.
Add comment