Transformers and diffusion models are not the same, even if they use related math
Edit: By the magic of advanced mathematics, every Turing complete structure can be rewritten into any other Turing complete structure with no loss of function or change in behavior
So the logic “remove one” will remove all of modern computing because the AI companies are not giving up their models. If they have to start pretending they aren’t neural network anymore, then they will.
And if you DON’T rewrite, then you’re back to them actually being different again
This is a dumb take. You can use a more general term like large dataset generative machine learning. Legislation needs to understand the different risks between the different implementations to do any good.
Machine B has now been fully rewritten to structures that are no longer a neural network. Everything it learned has been extracted from the model and given a new form. It doesn’t even run on a GPU anymore.
Its just decision trees and explicit function now. Same origin, same behavior, wildly different implementation and appearance. It just looks like another program. Especially to a judge.
Every rule you wrote that specified “neural”, not something more universal as “heuristic”, has now been denied by the judge, as you do not have jurisdiction against the new thing.
This isn’t some academic wordplay either. All machine implementable logic has infinitely many possible representations. If you specify your rule on appearance, not behavior, then the appearance will change and the behavior stays.
Did you seriously stop reading after the first line? Your response has NOTHING to do with what I wrote.
Keep reading
If you’re going to make a law saying “anything with a neural architecture” it’s possible to keep the behavior with alternative architectures which remove the neurons.
I very explicitly and plain as day DID NOT say they are using something different today, I said your proposed law would make them change in response. This is very clearly not the thing you think I said.
If you change the proposed law further to “any software which takes a large amount of data to extract patterns iteratively to solve for arbitrary targets” then your regulation will cover every optimizing compiler and a ton of physics solvers.
The law should cover behavior of software deployments, not architectures.
Removed by mod
Transformers and diffusion models are not the same, even if they use related math
Edit: By the magic of advanced mathematics, every Turing complete structure can be rewritten into any other Turing complete structure with no loss of function or change in behavior
So the logic “remove one” will remove all of modern computing because the AI companies are not giving up their models. If they have to start pretending they aren’t neural network anymore, then they will.
And if you DON’T rewrite, then you’re back to them actually being different again
https://infosec.pub/comment/23830616
Removed by mod
This is a dumb take. You can use a more general term like large dataset generative machine learning. Legislation needs to understand the different risks between the different implementations to do any good.
Removed by mod
https://arxiv.org/abs/2608.29530
https://link.springer.com/article/10.1023/A:1009670302979
https://proceedings.neurips.cc/paper/2020/file/e4acb4c86de9d2d9a41364f93951028d-Paper.pdf
https://arxiv.org/html/2106.03354v3
https://luke-griffiths.github.io/non-neural-networks/
Machine B has now been fully rewritten to structures that are no longer a neural network. Everything it learned has been extracted from the model and given a new form. It doesn’t even run on a GPU anymore.
Its just decision trees and explicit function now. Same origin, same behavior, wildly different implementation and appearance. It just looks like another program. Especially to a judge.
Every rule you wrote that specified “neural”, not something more universal as “heuristic”, has now been denied by the judge, as you do not have jurisdiction against the new thing.
This isn’t some academic wordplay either. All machine implementable logic has infinitely many possible representations. If you specify your rule on appearance, not behavior, then the appearance will change and the behavior stays.
Removed by mod
Did you seriously stop reading after the first line? Your response has NOTHING to do with what I wrote.
Keep reading
If you’re going to make a law saying “anything with a neural architecture” it’s possible to keep the behavior with alternative architectures which remove the neurons.
I very explicitly and plain as day DID NOT say they are using something different today, I said your proposed law would make them change in response. This is very clearly not the thing you think I said.
If you change the proposed law further to “any software which takes a large amount of data to extract patterns iteratively to solve for arbitrary targets” then your regulation will cover every optimizing compiler and a ton of physics solvers.
The law should cover behavior of software deployments, not architectures.
Removed by mod