Dive into Core AI model authoring and optimization

A Podkey summary of Dive into Core AI model authoring and optimization, from WWDC 2026.A lot of this week’s Core AI story is about taking what used to be fiddly, specialist work and making it much more direct. There’s a new through-line here: describe what you need, convert and run models from Python, debug where compression actually breaks things, and then reshape the model so the fast path is the one you use most. And the SAM3 examples are especially useful because they show both the promise and the limits of squeezing big models down for phones.Core AI skills as a planning layerA cleaner Python conversion pipelineCustom Metal kernels inside the modelWhat happened when SAM3 got crushed to int4Using the debugger to find the real damageThe small part of SAM3 that mattered mostReauthoring the model so repeated work disappearsWhy these pieces fit togetherThis podcast was created with Podkey. Make your own at https://podkey.fm