> You can't remotely detect that somebody is training an AI.
Energy use is energy use, training is still incredibly energy intensive and the GPU heat signatures are different from non GPU ones, it fairly trivial to detect large scale GPU usage.
Enforcement is a different problem, and is not specific to AI, if you cannot enforce an agreement it doesn't matter if its AI or nuclear or sarin gas.
Not if you are making a dirty bomb . Any radioactive material even at the levels found in power reactors can be dangerous.
The point is not the usage is harmful or not, almost any tech can be used for bad purposes if you wish to do so.
You can put controls is the point , controls here could be agent Dameons monitoring the gpus and tallying usage to heat signals, or firmware etc . The controls on what is being trained would be at a higher level than just agents process on a gpu .
That's why we should ban all large training runs without trying to distinguish good ones from bad ones. People will have to be satisfied with the foundation models we have now.
Energy use is energy use, training is still incredibly energy intensive and the GPU heat signatures are different from non GPU ones, it fairly trivial to detect large scale GPU usage.
Enforcement is a different problem, and is not specific to AI, if you cannot enforce an agreement it doesn't matter if its AI or nuclear or sarin gas.