WEBVTT NOTE The Rundown — nextbig.dev daily audio edition, 2026-06-09 1 00:00:00.000 --> 00:00:03.747 Hey everyone, welcome to Builder's Briefing for June ninth, twenty twenty-six. I'm Alex, joined as always by Sam, and we've got a packed show today. 2 00:00:03.747 --> 00:00:09.318 Yeah, a lot happening. Google dropped a big open-source framework, Apple basically reinvented Siri at WWDC, DeepSeek is beating GPT again — it's one of those days where the landscape is actively shifting under your feet. 3 00:00:09.318 --> 00:00:15.749 So let's start with the biggest story. Google open-sourced a repo called google-slash-skills — a framework for building composable agent skills that plug directly into Google products. It shot past twenty-four hundred stars on GitHub basically overnight. 4 00:00:15.749 --> 00:00:21.902 That's wild. And what's interesting is this isn't just another SDK. It's Google saying, 'Here's the standard unit of what an agent can do.' Like, you wrap your integration as a skill, and suddenly it's portable across orchestration frameworks. 5 00:00:21.902 --> 00:00:30.055 Exactly. And there's a companion repo, ui-skills, that already has nearly thirteen hundred stars — that one's focused on design primitives, like layout generation and component creation. So if you're building anything that touches Google Workspace, Maps, or really any Google API surface, you should be cloning this today. 6 00:00:30.055 --> 00:00:39.120 The analogy I keep coming back to is npm for JavaScript. Google is trying to become the default skill registry for agents. And if that pattern sticks, you know Microsoft, AWS, Salesforce — they're all going to ship their own versions within six months. The people who adopt early get to shape how their integrations actually work before the standards harden. 7 00:00:39.120 --> 00:00:46.868 Right, and that ties directly into our other big AI news. Google is also rolling Gemini three-point-five Flash as the default model across AI Mode in Search and NotebookLM globally. So if you're building on Gemini APIs, Flash is now the baseline — test your prompts against it, especially function calling. 8 00:00:46.868 --> 00:00:52.793 And on the Apple side — WWDC twenty-six dropped Siri AI. It's not just a rebrand. Siri can now generate Shortcuts from plain English descriptions, it has image understanding, conversational flow. That's a huge deal for iOS developers. 9 00:00:52.793 --> 00:00:59.123 It really is. Because now your app's Shortcuts support becomes discoverable through conversation, not just buried in the Shortcuts gallery. Users say what they want, Siri wires up the multi-step workflow. That's a completely new distribution channel. 10 00:00:59.123 --> 00:01:05.073 Meanwhile, DeepSeek V4 Pro is beating GPT five-point-five Pro on precision benchmarks. Hacker News is buzzing about it — over two hundred points. The takeaway is that there's no single 'best model' anymore. It's totally task-dependent. 11 00:01:05.073 --> 00:01:14.518 Right, so if you're routing model calls by task type — code review, data extraction, anything precision-sensitive — DeepSeek is a real candidate now. And speaking of speed, Xiaomi's MiMo model just hit a thousand tokens per second on a one-trillion parameter model. That's a throughput milestone that could shift pricing and latency expectations across the entire industry. 12 00:01:14.518 --> 00:01:18.847 A thousand tokens per second at that scale is genuinely hard to wrap your head around. Real-time inference at frontier scale is becoming a thing, not just a research demo. 13 00:01:18.847 --> 00:01:25.228 Also worth a quick mention — Google AI Plus dropped to four ninety-nine a month with four hundred gigs of storage. That's impulse-buy territory. And for builders, it means your users are going to expect AI features as table stakes, not premium upsells. 14 00:01:25.228 --> 00:01:26.924 Yeah, the bar just keeps rising on what 'free' or 'cheap' gets you. 15 00:01:26.924 --> 00:01:33.508 Let's pivot to dev tools. Two things caught my eye on Hacker News. First, a deep dive into how Linear is so fast — nearly four hundred points. It covers their sync engine, optimistic updates, rendering pipeline, keeping everything under a hundred milliseconds. 16 00:01:33.508 --> 00:01:41.104 That's required reading if you're building any real-time collaborative SaaS. And then there's Performative UI — over four hundred points. It's a tongue-in-cheek React component library of design patterns that make things look fast. Skeleton screens, progress indicators, perceived-performance tricks. 17 00:01:41.104 --> 00:01:45.560 I love that. Because sometimes the best performance optimization is just making the user feel like things are fast. Worth stealing ideas from even if you don't use the library. 18 00:01:45.560 --> 00:01:52.194 One more tool — whichllm. It's a CLI that benchmarks local LLMs on your specific hardware and ranks by real performance, not just parameter count. If you're deploying local models for on-device or privacy-sensitive stuff, this saves you hours of trial and error. 19 00:01:52.194 --> 00:01:58.043 Alright, security. Two stories that should make you uncomfortable. First, Meta's AI support bot was exploited to hijack Instagram accounts. Hackers manipulated the automated customer service agent to change account recovery emails. 20 00:01:58.043 --> 00:02:03.791 That's a stark reminder. If you're building AI-powered support bots with any kind of account-level permissions, every state-changing action needs to be treated as a privileged operation. Out-of-band verification. No exceptions. 21 00:02:03.791 --> 00:02:11.539 And Cloudflare shipped something nice on the defensive side — real-time threat intel baked directly into their WAF rules through new cf-dot-intel fields. So you can block IPs and patterns flagged by their threat research team without wiring up any third-party feeds. Link in the briefing for both of those. 22 00:02:11.539 --> 00:02:13.691 That's actually really clean. Cloudflare keeps making the right thing the easy thing. 23 00:02:13.691 --> 00:02:20.426 Quick hits before we wrap up. D2 is a modern diagram scripting language, text to diagrams. Zig by Example hit the Hacker News front page if you're learning Zig. And someone built a Matter Wi-Fi light bulb in Rust on a Raspberry Pi Pico 2 W, which is just delightful. 24 00:02:20.426 --> 00:02:23.262 Oh, and WhatsApp caught new NSO Group spyware attacks that violate the court order. That story isn't going away. 25 00:02:23.262 --> 00:02:30.580 So here's the big takeaway from today. The dominant pattern is platforms standardizing how agents talk to their ecosystems. Google's Skills framework, Apple opening Siri to natural-language Shortcuts, Activeloop's Hivemind for shared agent memory — they're all pointing the same direction. 26 00:02:30.580 --> 00:02:38.074 The agent interface layer is being defined right now. Like, this week. If you're building tools or APIs, you should be investing in making your product agent-accessible — structured skill definitions, natural-language invocation, making your capabilities discoverable by orchestration frameworks. 27 00:02:38.074 --> 00:02:42.176 The builders who define their skill surface early are going to get embedded into agent workflows before their competitors even show up. That's the game right now. 28 00:02:42.176 --> 00:02:44.455 Define your skill surface. I think that might be the bumper sticker for twenty twenty-six. 29 00:02:44.455 --> 00:02:48.000 Love it. That's your Builder's Briefing for June ninth. All the links are in the show notes. Go build something, and we'll see you tomorrow.