Aren't they basically safe as long as future AI models are still basically multiplying tensors of floats?