Build or Replaced
Today: Sherline Tools Is Going Out of Business | I am often wrong | The Effect of CRTs on Pixel Art (2024) Episode date: 2026-09-21.
Download MP3 →ERIK: Build or be replaced. JOSH: It's Monday, September 21st. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson. ERIK: Monday means the guardrails did their job all weekend without me touching a keyboard. JOSH: Stick around — Erik's got an AI pro tip at the end about the one thing most people skip before they trust an agent with real work. [pause] JOSH: Alright, headlines. First up — Samsung's doubling output on HBM4 memory next year. Why do I care about RAM? ERIK: Because every model you like runs on it. More high-bandwidth memory supply means cheaper inference eventually. It's boring until you realize it's the reason token prices keep dropping. JOSH: Next — Qwen Image 2.1 dropped from Alibaba. ERIK: Open-source image model, real jump over the last version. Every time China ships something like this in the open, it resets the floor for what "good enough" looks like for free. JOSH: There's also a new Cisco advisory out — a critical IOS XE vulnerability, already being exploited in the wild. ERIK: That one's personal. Ten years of my career lives on top of NSO talking to boxes running that code. When an advisory like that drops, it's not a headline for me, it's a Friday-night patch window for somebody's NOC team. JOSH: How do you even keep up with that at your scale? ERIK: I don't read every CVE by hand. I've got a runbook that flags anything touching gear in my config drift reports, and it routes straight onto PrimeBus so a human sees it before it turns into an incident instead of after. JOSH: And this one's a little sad — Sherline Tools is going out of business. ERIK: Small precision machine tool maker, decades of work, done. That's the other side of automation nobody talks about on this show enough. Not every closure is AI's fault, but the pace of everything right now doesn't help legacy manufacturers. JOSH: Is there a version of that story where automation saves a company like that instead of burying it? ERIK: Sure — if they'd modernized ordering and inventory ten years ago instead of running it on paper, they'd have had more runway to survive a slow quarter. Automation doesn't kill companies. Waiting until you're already dying to adopt it does. [pause] JOSH: Okay, let's go deep. Google dropped something called AX — an open agentic orchestrator. Break that down for me. ERIK: So AX sandbox-isolates AI agents and lets you declare their code, network policy, and secrets up front, then runs each one as a lightweight actor that can checkpoint, suspend, and resume. That's the whole game right there — suspend and resume. JOSH: Why does that specific detail matter to you? ERIK: Because that's exactly the problem I solved myself with PrimeBus, just without Google's name on it. I've got 140 distinct projects emitting telemetry onto that bus right now. When an agent kicks off a fix, it can't just die if something restarts — it has to pick back up where it left off. JOSH: Wait, 140 projects all on one bus? ERIK: Yeah, and PrimeRouter sits in front of it doing priority-tier routing so I'm not burning a frontier model on a task a small one can handle. AX is basically Google saying "yeah, this pattern is real," at Google scale. I've been living it at home-lab scale since before it had a name. JOSH: Does that make you nervous — Google building the same thing you built? ERIK: Not nervous. Validated. When a company with infinite compute ships the architecture you already committed to, that's not competition, that's confirmation you weren't wasting your time. [beat] JOSH: What's the actual failure mode if you don't have checkpoint and resume built in? ERIK: You get zombie work. An agent starts a merge, the host reboots for a kernel patch, and now nobody knows if that merge actually happened. I've had that exact scenario on PrimeBus before I had proper state tracking — an agent still thought a fix was in flight three hours after the process had died. JOSH: How'd you fix that? ERIK: Every job writes its state before it touches anything, not after. So a restart just means the next check picks up exactly where the last one stopped instead of guessing. JOSH: So the sandbox part and the resume part are actually two different problems. ERIK: Right. Isolation keeps a bad agent from touching what it shouldn't. Resume keeps a good agent from lying to you about what it already did. Google bundling both into one primitive tells me they've been burned by the same thing I have. [beat] JOSH: Next story — "MCP was always a bad idea?" That's a real headline. ERIK: And it's a real question. MCP made it easy to hand a model a tool and just hope. No contract, no guardrail, just "here's a function, go." JOSH: So what's the fix — don't use MCP? ERIK: No, the fix is you don't trust the tool call blindly. Every agent in my fleet — Neo, Homer, Bill, Gandalf, twelve agents total right now — goes through review before anything merges. PrimeBus auto-merger has made 2,295 attempts since June. 1,507 got merged. 788 got blocked by Gandalf. JOSH: Zero escalated to you? ERIK: Zero. That's the part people miss when they read "MCP is bad." The tool isn't the risk. The absence of a reviewer is the risk. 788 blocked isn't a failure rate — that's 788 times the guardrail did exactly what it was built to do instead of me finding out at 2 AM. JOSH: So critics are aiming at the wrong target. ERIK: They're aiming at the gun instead of the guy with no safety training. Give an agent a tool without a review gate and yeah, MCP was a bad idea. Put Gandalf in front of it, and it's just infrastructure. JOSH: Does that scale, though? Twelve agents is one thing — what happens at a hundred? ERIK: The review gate doesn't care how many agents are asking. It cares about the diff. Whether it's twelve agents or a hundred, every change still goes through the same checkpoints before it touches anything that matters. More agents just means more traffic through the same door, not a different door. [pause] JOSH: Last one — this one's a mood piece. "The LLMentalist Effect." What is that? ERIK: It's the theory that a lot of what feels like an LLM "understanding" you is the same trick cold readers and mentalists use — vague enough that you fill in the meaning yourself, and it feels uncanny. JOSH: That's a little unsettling. ERIK: It should be. It's a good check on hype. I run ScanBrief every morning — scored 77 items across 56 sources today, average relevance around 52.7 — and half the AI headlines in there read like the model discovered something. Most of the time it just gave you a mirror. JOSH: How do you keep yourself honest about that with your own systems? ERIK: I don't grade my agents on how impressive their output sounds. I grade them on whether the merge held, whether the test passed, whether the number in the database matches reality. 153 services are running on my production server right now, and not one of them cares if the code that touches it sounded smart. It either works or Gandalf blocks it. JOSH: That's a pretty grounded way to read a spooky headline. ERIK: Cold reading works on humans because humans want to believe. Build systems that don't run on belief. [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: Before you let any agent take an action with side effects — a merge, a send, a deploy — give it a fail-closed path, not a fail-open one. If a key's missing, a service is down, or a check can't run, the default has to be "do nothing," never "do the risky thing anyway." I write that rule into every chain before the chain does anything else. It's the difference between 788 blocked merges and one bad one that finds you at 2 AM. JOSH: That's a lot more specific than "be careful." ERIK: "Be careful" isn't a design pattern. Fail-closed is. 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.