Anthropic recently made headlines by retiring a challenging take-home exam after its AI model, Claude Opus 4.5, outperformed all candidates in just two hours. Tristan Hume designed the new assignment to engage candidates more effectively, moving away from generic problems. The task is centered around optimizing code for a simulated accelerator, modeled after Google’s TPUs, used mainly for AI training and inference. This assignment offers insights into performance optimization without requiring a technical background, as Hume promises to explain essential concepts like SIMD (Single Instruction, Multiple Data) and VLIW (Very Long Instruction Word). The simulator mimics a fictional processor architecture with two types of memory: slow DRAM and a fast Scratchpad. The task involves manipulating a binary tree and processing objects called "irises," which have a straightforward structure involving indices and values. The optimization challenge differs from traditional decision trees; instead of simple comparisons, it employs a hash function that determines traversal through the tree based on whether a result is even or odd. This complexity likely aims to avoid making the task memory-bound, which would reduce the challenge to a mere speed test. For execution, the assignment requires writing a kernel, a low-level program that runs on the accelerator. The kernel consists of simple commands that the hardware can process in parallel, making it crucial for performance. Hume provides details on the structure of the necessary code and the nature of the instructions to be executed. This task not only tests programming skills but also deepens understanding of how to leverage specific hardware capabilities for optimization.