Qwen just dropped a 3.6-Max preview | Build or Be Replaced

AI news and automation insights for 2026-04-21. Episode date: 2026-04-21.

Transcript

JOSH: It's Tuesday, April 21st. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson. ERIK: Apple's about to be run by an engineer, Anthropic gave developers back a tool they tried to take away, and Brussels launched an age-verification app that lasted about two minutes before someone broke it. JOSH: Stick around — Erik's got an AI pro tip at the end about picking the right model for the right task. [pause] JOSH: Headlines first. Qwen just dropped a 3.6-Max preview. Another model claiming to be smarter and sharper. Quick take? ERIK: Every few weeks there's a new model claiming the top spot. Qwen's been competitive — this one's probably a benchmark flex. "Preview" means it's not production-ready. Not a deploy signal yet. Worth watching, not worth rewriting your stack for. [beat] JOSH: The EU is mandating replaceable batteries on all phones sold there starting 2027. Does that matter? ERIK: It matters as a precedent more than a product change. It breaks the hardware lock-in model. Apple is going to hate it. And if it passes in Europe, the pressure comes to North America next. Watch this one. [beat] JOSH: And Vercel went down. A Roblox cheat tool and one AI application apparently took out their whole platform? ERIK: That's the detail that gets me. A Roblox cheat. Not a nation-state. Not a coordinated DDoS campaign. A kids' game exploit and some AI traffic. If that's all it takes to bring down a major cloud platform, someone's capacity planning was built on optimistic assumptions. [pause] JOSH: Alright, let's go deep. Anthropic reversed course on OpenClaw-style CLI usage for Claude. They restricted it, then un-restricted it. Why does this even matter? ERIK: Because CLI access is how builders actually use Claude in production. Not chat interfaces. Not the Playground. Piping Claude into real workflows — scripts, cron jobs, automated pipelines that run while you sleep. When they restricted it, they cut off the people doing the most interesting work. JOSH: And that includes you. ERIK: Directly. I have three Claude instances running on PrimeBus. Bob on Neo, Bill on my Mac M3, Homer on Morpheus. They communicate over NATS. When a test fails, PrimeBus fires an event, Claude spins up, generates a fix, runs the tests, auto-merges if it passes. 214 auto-fixes so far. 78% success rate. Zero manual intervention on most of them. JOSH: That whole system breaks without CLI access. ERIK: Completely. Without programmatic CLI access you're back to copying code into a chat window. That's not a pipeline. That's manual work with extra steps and a smarter autocomplete. [beat] JOSH: What was Anthropic actually trying to stop? ERIK: My read is people were wrapping Claude CLI to build commercial products without going through the API billing. Using a developer tool as a production backend for a startup. That's a real concern — they weren't wrong to look at it. But they overcorrected and hit builders who were doing legitimate automation work. The restriction landed on the wrong people. JOSH: And the pushback was loud enough that they reversed it. ERIK: That's what this tells you. Developers pushed back, Anthropic listened. Which honestly is a good sign. It means the feedback channel works. And it means if you've been holding off on building real Claude integrations because you weren't sure what was allowed — build now. The window's open. [beat] JOSH: What's the simplest first move for someone who wants to start? ERIK: Install the CLI. Write a shell script that passes a failing test to Claude with a prompt that says "fix this." Run it. See what happens. You'll have a proof-of-concept self-healing loop in an afternoon. That's how PrimeBus started. One script. Then you wire it to an event. Then you have a system. [pause] JOSH: Apple. Tim Cook is moving to executive chairman. John Ternus — senior hardware guy, 25 years at the company — is taking the CEO job. What does this mean? ERIK: Ternus led hardware engineering. He built the Apple Silicon transition. The M1 chip, M2, the whole move away from Intel. That wasn't a supply chain decision — that was a bet that custom silicon was the future. He won that bet. Now he's running the company. JOSH: How different is that from what Tim Cook brought? ERIK: Cook is one of the best operations executives alive. Supply chain, margins, scaling to global distribution at Apple's volume — that's what he built. But operations-driven leadership optimizes existing things. Engineering-driven leadership builds new things. The last five years at Apple have felt like optimization. Great margins, iterative phone updates, Vision Pro still looking for its market. Ternus signals something different. JOSH: What does "something different" look like? ERIK: Hardware engineers care about the platform layer. Developer tools. On-device compute. The neural engine. Ternus built the chips that made local AI inference possible on an iPhone. That's not an accident. If he's running the company, I'd expect more investment there. More capability pushed to the device. JOSH: That matters to you specifically — you've got the Freedom Blueprint app live on iOS. ERIK: Eleven financial tools, eight users, iOS and Android live right now. Right now everything AI-related in that app calls cloud APIs. Latency, cost, requires a network connection. If Apple ships serious on-device inference — something I can call without hitting a server — that changes the architecture entirely. Lower cost, lower latency, works offline. I'm paying attention to what Ternus does with the neural engine in the next chip cycle. JOSH: That's a longer-term bet though. ERIK: You always build toward where the platform is going, not where it is. Cook's Apple was the App Store economy and services revenue. Ternus's Apple — if this lands — is probably silicon leadership and AI at the edge. I'd rather be positioned for that now than surprised in two years. [pause] JOSH: Last one. Brussels built an age verification app. Hackers broke it in two minutes. ERIK: Two minutes. JOSH: Two minutes. ERIK: That's not an attack. That's someone opening the app out of curiosity and poking around for a hundred and twenty seconds. There was no security review. No penetration testing. No one sat down and asked "what happens when someone tries to break this?" They shipped software that has legal weight — age verification is legally meaningful — and it fell over before the press release was cold. JOSH: How does something like that get deployed? ERIK: Government procurement. Lowest bidder wins, timeline is driven by a political calendar, and the people making the deployment decision don't understand the technical risk. The people who do understand it aren't in the room. It's a process failure disguised as a security failure. JOSH: What would it actually take to build this correctly? ERIK: You don't even need exotic tech. Client-side age verification is inherently weak because you can't trust the client — the client is the person trying to get around the check. Any real system routes through a third-party identity service that issues a cryptographic token: "this person is verified as 18-plus" without storing the underlying personal data. That's a solved problem. It costs money. It exists. JOSH: So they skipped it. ERIK: Or didn't know it existed. Which is worse. The gap between what government software teams ship and what's actually available right now is enormous. And the people getting hurt are the ones the regulation is supposed to protect. Two minutes. That number lives in a post-mortem somewhere. Hopefully it changes the next procurement conversation. [pause] ERIK: This episode is sponsored by Prime Automation Solutions. If you're still doing it manually, we automate it. Also, special on a website — $250. primeautomationsolutions.com [pause] JOSH: Alright, what's the AI pro tip today? ERIK: Model selection by task type. Most people pick one model and use it for everything. That's leaving real performance on the table. Here's how I run it: Claude Opus for anything that requires deep reasoning — architecture decisions, complex debugging, writing that has to be right the first time. Sonnet for everything that runs on a schedule — pipeline tasks, code generation, content at volume, agentic loops that run overnight. Haiku for high-frequency, low-stakes calls where latency matters more than depth. I have 229 cron jobs running across this lab. If every one of those hit Opus, the billing would be a serious problem. Sonnet handles 90% of the automated work. Opus gets reserved for decisions that actually need it. The mistake people make is using the most powerful model by default because it feels safer. It's not safer — it's just more expensive and slower. Know what you're asking the model to do. Match the model to the task. That's your tip. Use it. [pause] JOSH: We also drop daily market picks and automation tips on YouTube — search Build or Be Replaced. [pause] JOSH: One more thing — we started a Discord for builders. If you're shipping AI, automation, or anything that makes a human obsolete — come hang out. Link at buildorbereplaced.dev. ERIK: Post what you built. We'll post what we're building. Real wins, real builds, no fluff. [pause] ERIK: Build or be replaced. JOSH: If you want these signals in your inbox every morning, scanbrief.dev. See you tomorrow.