WWDC 2026: The Swift "Glow" and the Transition to Core AI

Guest: Thanks, Alex! It’s great to be here. And man, that invite… I think my heart rate actually spiked when I saw that glowing logo. It’s been a long time coming, hasn't it? Host: It really has! I mean, "Coming Bright Up"—they aren't even trying to hide it. Before we dive into the deep technical stuff, what was your immediate "gut reaction" to the invitation? Guest: Honestly? Relief. (laughs) Because, look, we’ve been trying to shoehorn these massive Transformers and Large Language Models into Core ML for a couple of years now, and… it’s been a struggle, Alex. Core ML was brilliant for its time—great for image classification or basic regressions—but it wasn’t built for the "weight" of modern AI. Seeing that "Glow" tells me Apple is finally saying, "Okay, we’re rewriting the foundation." It’s a signal that the compiler itself is becoming AI-aware. Host: That’s such an interesting way to put it—the "compiler becoming AI-aware." You mentioned that Swift might be becoming an "AI-native" language with the upcoming Swift 8.0. What does that actually look like for a dev sitting at their desk? Guest: So, okay, think about how we handle memory right now. We have these strict concurrency models and ownership rules. In Swift 8.0, we’re expecting to see what I call "first-class primitives" for tensors. Imagine a `NeuralTensor` struct that is `@frozen` and maps directly to the NPU—the Neural Processing Unit—memory. Host: Oh, wow. So instead of us having to manually manage how a model sits in memory, the language handles the "heavy lifting" of the unified memory architecture? Guest: Exactly. It’s about maximizing that M-series or A-series chip without the developer needing a PhD in hardware engineering. We’re looking at a world where you can offload an attention head of an LLM to the NPU while your GPU is still focused on, say, rendering a complex SwiftUI animation. It’s all about throughput. Host: That brings us to the "Core AI" rumor. We’ve heard Core ML is essentially being put out to pasture. What was the "breaking point" for Core ML? Why the total overhaul? Guest: It’s the "static" problem. Core ML was built for static models. You’d train a model, convert it to a `.mlmodel` file, and it was… well, it was done. But the LLM era? That demands dynamic weights, KV caching, and real-time quantization. Host: Interesting! So, my app gets to be "smart" without the 2GB download size? That would be a game-changer for the App Store. Guest: Huge. And it’s a privacy play, too. Apple is building this "Privacy-First" moat. If the inference happens on-device using these system-wide models, the data never leaves the silicon. No round-trip to a server in the cloud. That’s how they beat everyone else—by making it fast, free for the dev, and private for the user. Host: I was reading on MacRumors that we might see new APIs like `CoreAI.TextGenerator` and `CoreAI.SemanticIndexer`. It sounds like they’re trying to make multimodal AI as easy as adding a button. Guest: It really does. I actually saw some speculative code recently—actually, I think it was in your notes, too—where you just call `LLMModel.loadSystemModel` and then pass it an image and a text prompt in one go. No more complex pre-processing of images into pixel buffers manually. Just: "Here’s a photo, here’s a prompt, tell me what’s happening." Host: (laughs) As someone who has spent way too many hours debugging pixel buffer formats, that sounds like a dream. But Marcus, let’s talk about the "struggle" side of this. For the developers listening who have existing Core ML implementations… what does the migration look like? Is this going to be a "Swift 2 to Swift 3" level of pain? Guest: (sighs) Honestly, Alex? It might be. Apple usually provides conversion tools in Xcode, so we’ll likely see a "Transpile to Core AI" button. But the real work is architectural. If your app is built on old, synchronous patterns, you’re going to have to refactor for the streaming world. Generative AI doesn't just "finish" a task; it streams tokens. Our UIs have to become more fluid to handle that. Host: It’s like the "Glow" is everywhere—in the language, the framework, and the IDE. It feels like a very "Apple" way of closing the loop. Guest: Exactly. They aren’t just playing catch-up in the AI race. They’re redefining the rules. They’re saying, "AI isn't a feature you add to an app; it's the foundation the app is built on." Host: That is a perfect place to wrap this up. Marcus, thank you so much for coming on and helping us decode the "Glow." I feel like I need to go rewrite all my projects now! Guest: (laughs) Wait for the keynote! Let’s see what they actually announce in June. But yeah, start brushing up on your async/await, because it’s about to get very busy.