If they're doing inference on edge devices, one challenge I see is protecting model weights. If you want to deploy a proprietary model on an edge AI chip, the w

If they're doing inference on edge devices, one challenge I see is protecting model weights. If you want to deploy a proprietary model on an edge AI chip, the weights can get stolen via side-channel attacks [1]. Obviously this isn't a concern for open models, but I doubt Apple would go the open models route. [1] https://spectrum.ieee.org/how-prevent-ai-power-usage-secrets https://spectrum.ieee.org/how-prevent-ai-power-usage-secrets