Build or Replaced
Today: Turn off Apple Intelligence on macOS 27 and get its disk space back | Powerless F1 drivers frustrated by Bahrain F1 software glitch | A browser-native classic Visual Basic VB6 IDE Episode date: 2026-10-05.
Download MP3 →ERIK: It's Monday, October 5th. This is Build or Be Replaced — powered by ScanBrief.dev. JOSH: And I'm Josh, here with Erik Anderson. ERIK: Episode 120. And this morning ScanBrief chewed through 68 items across 56 sources before either of us had coffee. JOSH: Stick around — Erik's got an AI pro tip at the end about when to stop paying for a model you don't need. [pause] JOSH: Okay, headlines. First up — Apple Intelligence on macOS 27 is apparently eating your disk space, and there's a toggle to turn it off and get the gigabytes back. ERIK: Which tells you everything about how that feature shipped. Bundled on-device models, no opt-in prompt, just "surprise, you're hosting a small AI stack now." Good that it's a toggle. Bad that it defaulted to on. JOSH: Next — Denmark had a data breach, 8.8 million people's personal data exposed. ERIK: That's almost the entire population of Denmark. When the breach number rounds to "everyone in the country," the story isn't the hack, it's whatever database let that much live in one place. [beat] JOSH: And there's a write-up on self-hosted HTTP tunnels using SSH and Nginx. ERIK: I like this one because it's not new technology, it's just someone writing down what a lot of us already duct-tape together. SSH reverse tunnel plus Nginx is how half my home lab talks to itself. JOSH: One more — a report that some staffing platforms are letting an AI agent run the first round of technical interviews now. No human until round two. ERIK: That's going to produce bad rejections and bad hires in roughly equal numbers. An agent interviewing for depth needs a rubric a human actually trusts. "Did it sound confident" is not a rubric. [beat] JOSH: And there's a thread going around about companies quietly cutting their cloud GPU budgets this quarter. ERIK: That's the tell. Nobody announces "we're spending less on inference," they just stop renting the big instances. JOSH: How do you even spot that from outside, if nobody's saying it out loud? ERIK: You watch the secondary market. When a quarter's worth of reserved capacity suddenly shows up cheaper on rental marketplaces, that's somebody's unused reservation hitting the market. I watch pricing like that before I commit to anything live, same way I'd want a receipt before I trust a vendor's own number. [pause] JOSH: Let's go deeper on a few of these. Start with the Qwen story — a 125-billion-parameter model running on a single RTX 4090, something like 100 trillion-token-per-second-equivalent throughput. That's a consumer card doing frontier-era numbers. ERIK: The throughput number's the hook, but I don't fully trust vendor throughput numbers until I've run them myself. Too many of those benchmarks pick the one prompt that makes the model look good. What I do trust is the trend line. A year ago, 125 billion parameters meant a rack. Now it's a card you can buy at Micro Center. That direction doesn't lie even if the exact number's inflated. JOSH: So what does that actually change for someone running a fleet like yours? ERIK: It changes where the cheap work goes. Not every job needs a frontier call. A lot of what hits my pipeline is classification — is this error real or noise, does this item even deserve a score. That's grunt work, and grunt work is exactly the tier a local card like that is built for. JOSH: So you've already moved stuff over? ERIK: No, and I want to be straight about that. I've tested it, it's promising, and promising isn't deployed. Same thing always holds me back — I don't trust a new tier until I've broken it on purpose first. Reliability under load, what it does when the prompt's weird, whether it fails quiet or fails loud. JOSH: Give me an example of what "already stress tested" looks like, versus not. ERIK: ScanBrief ranked 68 items this morning off an average relevance score of 52.9, and that pipeline runs clean because I've already beaten on it. I'm not putting a model I tested for a weekend into a job that runs before I wake up and nobody's watching it fail. JOSH: That's a more careful take than "just switch and save money." ERIK: Because that pitch is how you end up debugging a silent failure at 2am. The savings are real on paper. Paper doesn't page you. [pause] JOSH: Let's talk about what's actually running while you sleep. PrimeBus — that's the self-healing pipeline, right? ERIK: Right, it's the event bus underneath everything. Since June 5th it's run 2,295 auto-merge attempts. 1,507 got merged clean. 788 got blocked by Gandalf. Zero — zero — came to me. JOSH: Wait, 788 blocked sounds like a lot of failures. ERIK: That's the part people get backwards. Blocked isn't failure, blocked is the guardrail working. A merge that gets stopped before it ships broken code did exactly what it was built to do. The number I'd actually worry about is zero blocks, because that means either everything's perfect, which it isn't, or nothing's checking. JOSH: So the real metric isn't the merge rate. ERIK: It's the ratio holding steady. 1,507 clean, 788 caught, and I didn't have to open my laptop for any of it. The whole point of the system isn't that it never fails. It's that failure gets caught somewhere that isn't production. [beat] JOSH: Last deep dive — two stories that are really the same story. There's a piece on backup discipline called "The Tao of Backup," and separately someone automated their 35mm film scanning pipeline. Feels like a stretch, but I bet you connect them. ERIK: Same pattern, different domain. Backup isn't a checkbox, it's a discipline, and the "Tao" of it isn't the backup, it's the restore. I run PrimeRestorer to pull databases and system state offsite on a schedule, but the backup isn't what I trust — testing the restore is what I trust. Anybody can copy a file. Fewer people actually bring it back and check it's not corrupted. JOSH: And the film scanning thing? ERIK: Same shape. Somebody had a manual repetitive task — load film, scan, color correct, export — and built a pipeline so a human doesn't babysit every frame. That's identical to what email-reactor does on the business side. Client sends a change request, the system reads it, routes it, acts. Different domain, same move — find the repetitive step, put it on a pipeline, let it run while you're not watching. JOSH: Does that ever feel like overkill for a one-person shop? ERIK: No, it matters more solo. I'm the only recovery plan there is. If PrimeRestorer goes quiet for a week and I don't notice, that's on me, nobody else is watching it. Scale doesn't make the discipline matter more. It just means more people share the blame when nobody tested the restore. [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: Stop paying frontier prices for triage. Before you send a call to your most expensive model, ask whether the task is actually hard or just frequent. Classification, routing, "is this worth looking at" — that's volume, not difficulty, and volume belongs on the cheapest model that's reliable enough to trust. JOSH: How do you figure out where that line is? ERIK: I'm not going to hand you a clean ratio, because I don't have one I'd bet money on yet. What I'd actually do today is pull your own logs, tag ten calls by hand as "actually hard" versus "just routine," and look at how lopsided that split is. For most people it's uglier than they think — mostly routine, a handful hard. Price your model tier to that split, not to your gut. ERIK: That's your tip. Use it. [beat] ERIK: That tip is straight out of The Autonomous Engineer — my book on building systems that run themselves. Grab it on Amazon. [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.