Fabio Manganiello on Nostr: npub1vygwc…apzlt npub1yhj4u…h3h3t even if we're talking about data centers, a ...
npub1vygwc7m2as9al2qjdlmh5j2m5yv7wfcxghmsl6yemf6rfxlu5elqtapzlt (npub1vyg…pzlt) npub1yhj4ua6580h3w0kucwwrtwl4uv9pg7m2drxzplq5x3364jczp36s4h3h3t (npub1yhj…3h3t) even if we're talking about data centers, a solution like this needs to go granular to investigate the use-case of the code that is being run, if we want to apply the tax consistently.
Does importing PyTorch or Tensorflow, or using the CUDA libraries, make my code liable for a higher energy tax, since those frameworks are used to train AI models?
Or maybe the tax should be on high-end GPUs used to train large models?
Or maybe on the size of the model itself?
What about the use-case of the model? Does a model that predicts temperature variations in the Atlantic in order to better understand climate change have to pay the same energy tax as a model used to generate memes?
It's a slippery slope and it doesn't take long before it compromises the principle of computing neutrality - which is as important as network neutrality. It's a solution akin to taxing blank CDs and USB drives because they could be used to distributed pirated content - it targets the intermediary technology rather than the actual problem, and it misses the distinction between use and abuse.
That's why I believe that a more effective solution may involve a tax on the excess of consumed energy, once reasonable baselines have been defined, rather than the technology itself. It doesn't penalize legitimate use-cases, while it forces everyone to think of more sustainable ways of writing and running their software.
After all, the source of all problems isn't AI (or even Blockchain) itself, but the misguided assumption that we live in a world with unlimited resources, and that those resources are cheap, and "scaling up" is just a matter of throwing more computing power at the problem, because it's cheaper than throwing more engineering resources to think of better ways of solving the problem. Taxation only needs to change this broken system of incentives, not be opinionated on what technologies are sustainable and which are not.
Does importing PyTorch or Tensorflow, or using the CUDA libraries, make my code liable for a higher energy tax, since those frameworks are used to train AI models?
Or maybe the tax should be on high-end GPUs used to train large models?
Or maybe on the size of the model itself?
What about the use-case of the model? Does a model that predicts temperature variations in the Atlantic in order to better understand climate change have to pay the same energy tax as a model used to generate memes?
It's a slippery slope and it doesn't take long before it compromises the principle of computing neutrality - which is as important as network neutrality. It's a solution akin to taxing blank CDs and USB drives because they could be used to distributed pirated content - it targets the intermediary technology rather than the actual problem, and it misses the distinction between use and abuse.
That's why I believe that a more effective solution may involve a tax on the excess of consumed energy, once reasonable baselines have been defined, rather than the technology itself. It doesn't penalize legitimate use-cases, while it forces everyone to think of more sustainable ways of writing and running their software.
After all, the source of all problems isn't AI (or even Blockchain) itself, but the misguided assumption that we live in a world with unlimited resources, and that those resources are cheap, and "scaling up" is just a matter of throwing more computing power at the problem, because it's cheaper than throwing more engineering resources to think of better ways of solving the problem. Taxation only needs to change this broken system of incentives, not be opinionated on what technologies are sustainable and which are not.