My iPhone 16 Pro Max produces garbage output when running MLX LLMs (journal.rafaelcosta.me)

A developer recently encountered significant issues while testing machine learning models on their iPhone 16 Pro Max, which yielded incorrect and nonsensical output. Despite the same code executing flawlessly on an iPhone 15 Pro and a MacBook Pro, the new device's Neural Engine appeared to malfunction, producing tensor outputs that were orders of magnitude off from expected values. This anomaly raised concerns about potential hardware defects in the iPhone 16 Pro Max, as the same software and models functioned well on other devices. This situation is particularly relevant for the AI/ML community as it highlights potential discrepancies in hardware performance, especially with specialized components like Apple's Neural Engine. The developer's clarity on the error demonstrates the importance of thorough debugging across both software and hardware layers. With the rapid proliferation of AI-powered applications, the incident underscores the need for developers to consider hardware integrity during troubleshooting, particularly as AI models become more integrated into consumer devices. Ultimately, when addressed, it revealed that the iPhone 16 Pro Max was defective, cementing the critical nature of reliability in machine learning operations on mobile platforms.