Build or Replaced
Today: Someone behind the software | Dogs seem to agree with humans that 'bouba' is round and 'kiki' is spiky | RepRapMicron Episode date: 2026-10-19.
Download MP3 →JOSH: It's Monday, October 19. This is Build or Be Replaced, powered by ScanBrief.dev. I'm Josh, here with Erik Anderson. ERIK: Today's theme: if you can't see what your AI is doing, you're not running it. It's running you. JOSH: Stick around. Erik's got an AI pro tip at the end about the one log you should be keeping on every model call. [pause] JOSH: Episode 129. Erik, what did the brief catch this morning? ERIK: ScanBrief scored 73 items across 56 sources. Average relevance 53.6. Quiet Monday, but two of the top stories are really about the same thing. JOSH: Which is? ERIK: Who's in control of the model, and who can see inside it. We'll get there. [pause] JOSH: Headlines first. Linux 7.3 is out. Big release? ERIK: Record-setting. 18,765 non-merge commits from 3,021 developers. 807 of them first-timers. JOSH: That's a lot of strangers touching the kernel. ERIK: It is, and the review process absorbs it. The parts I care about are the global per-CPU BPF data and the group-scheduling work for big systems. Boring for most people. Great for anyone running a lot of services on one box. [beat] JOSH: Next. macOS 15.8.2 breaks OpenGL. ERIK: Of course it does. Apple deprecated OpenGL years ago and now a point release breaks it. If you've got an old render pipeline or a legacy tool leaning on it, pin your OS version before you update. Don't find out on a deadline. JOSH: And there's a Terry Tao story. He says he's on a war footing over the crisis in math. ERIK: It's right next to the 3Blue1Brown video, which argues now is the best time to go into math. Same field, two opposite moods. My read: AI tools are changing what math work looks like, and the people who learn to work with them win either way. [pause] JOSH: Deep dive time. Someone built a programming language that LLMs can't use. Why would you do that? ERIK: Honestly, I get the impulse. Every language right now gets judged on whether Claude can write it well. If a model can't write your language, that's a feature or a bug depending on who you are. JOSH: Is it a feature? ERIK: Sometimes. Think about it from the control side. If a model can't produce valid code in a system, then nobody can quietly paste generated code into it. For most of us that's a terrible trade. I'd rather have a language the model handles well and put guardrails around the output. JOSH: What kind of guardrails? ERIK: PrimeBus is my example. It runs the auto-review and auto-merge pipeline. The auto-merger has run 2,295 attempts since June 5th. 1,507 merged. 788 blocked by Gandalf. Zero escalated to me. JOSH: Wait, 788 blocked. That sounds like a lot of failures. ERIK: It's not failures. That's the story. Models write code all day. Gandalf is the thing that says no. The 788 are the guardrails working. I don't need a language the AI can't use. I need a gate it can't walk around. [beat] JOSH: So the answer to AI-written code isn't banning it. ERIK: Right. Make it earn the merge. A model that can write anything, plus a reviewer that blocks most of the risky stuff, beats a clever language nobody can build on. JOSH: Does that apply outside code? ERIK: Everywhere. Same rule for email, same rule for infra changes. Let the model draft, make a second system verify. Cheap to do, and it's the difference between a toy and something you leave running overnight. [pause] JOSH: Okay, that leads into the next one. Someone on Hacker News asked: if you don't run your own LLM proxy, how do you know what's going on? ERIK: Best question on the list today. And my answer is, you don't. JOSH: Say more. ERIK: Every provider gives you a dashboard. It tells you what they want you to see. Total spend, maybe a graph. It doesn't tell you which of your services made which call, which retry was a mistake, or which prompt is quietly eating tokens at three in the morning. JOSH: So what do you do instead? ERIK: Put a gateway in front of everything. Mine is PrimeRouter. Every service in the lab calls it, not the model provider directly. It decides which backend gets the call, it logs the request, it fails over when a backend is down, and it enforces caps. JOSH: Caps meaning spending limits? ERIK: Spending and call limits, yes. If something loops, the router trips a limit and stops before the bill does. I've had that save me from my own bugs. Sessions that burn credits are the thing that ruins your month. [beat] JOSH: Is it a lot of work to set up? ERIK: The first version is tiny. One endpoint, one log table. Prompt in, backend picked, response out, tokens counted. Everything else, the failover and the caps, gets added after you've seen real traffic. JOSH: And you'd say that's mandatory now? ERIK: If you have more than one thing calling a model, yes. With 139 services running on the production server right now, I can't trace each one by hand. The proxy is how I know what's happening without reading logs all day. JOSH: And the Hermes backend runs through that too? ERIK: It does. Hermes is the local agentic backend on the Mac M3, running qwen3-coder at about 70 tokens a second. The router treats it like any other backend. If it's down, the call moves to the next one. The caller never knows. [pause] JOSH: Last deep dive. Hybrid Reasoning Agents. LLM agents that think in latent space and act through text. Translate that for me. ERIK: Normal agents think out loud. Every step of reasoning is words, and every word costs tokens. The idea here is the model does most of its thinking in vectors, its internal representation, and only comes up for air to talk to the world in text. JOSH: So it thinks quietly and speaks rarely. ERIK: That's the pitch. The write-up says it cuts token usage and improves planning efficiency. I'd say it's promising, and I'd wait for independent results before I rebuild anything around it. JOSH: Why does it matter if you're a builder, not a researcher? ERIK: Tokens are the cost center of every agent loop. If an agent can plan in fewer visible steps, a multi-step job gets cheaper and faster. That changes what's worth automating. JOSH: What's the catch? ERIK: Visibility. If the reasoning lives in latent space, you can't read it. Right now when an agent does something dumb, I read the transcript and see why. With silent thinking, that transcript gets thinner. [beat] JOSH: Which loops back to the proxy conversation. ERIK: Exactly. The more the model hides, the more your outside logging matters. You log what goes in, what comes out, what it cost, and which tool calls it made. You can't see inside, so you watch the edges really closely. JOSH: That's a pretty good rule. ERIK: It's the whole show today. Gates, gateways, logs. Control the edges. [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: Log every model call in one place, with five fields. Timestamp, calling service, model, tokens in and out, and whether it succeeded. That's it. A single table or a single JSON lines file. JOSH: That's it? ERIK: That's it. After a week, sort by tokens and look at the top five callers. I guarantee one of them surprises you. Maybe a retry loop, maybe a prompt that grew every time it ran. You fix that one thing and your bill drops. JOSH: And if you don't have a gateway yet? ERIK: Wrap your model client in one function. Everything calls that function. Now you've got a gateway in ten lines, and you can grow it into a router later. ERIK: That's your tip. Use it. [pause] ERIK: That tip is straight out of The Autonomous Engineer, my book on building systems that run themselves. Grab it on Amazon. [beat] 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.