Build or Replaced
Today: Reverse engineering my ADHD test | Dwarf Fortress is getting the mother of all magic updates | GPU World Episode date: 2026-09-01.
Download MP3 →JOSH: It's Tuesday, September 1. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson. ERIK: Cheap compute won't replace builders. It replaces people who never wired their work into a system. JOSH: Stick around — Erik's got an AI pro tip at the end about making agents leave an audit trail before they touch anything. JOSH: First headline. Hacker News is talking about GPU World, a story contest asking what happens if everyone has B300-class compute by 2040. Fantasy, warning, or both? ERIK: Both. If compute gets that cheap, the bottleneck moves from can I run the model to did I build anything worth connecting it to. JOSH: Second. Fastpotify is a native Spotify client without Chromium, and people are happy because it starts fast and uses less memory. Why does that hit a nerve? ERIK: Because we've been hiding laziness inside browser runtimes. A music player shouldn't feel like it needs a change window. JOSH: Third. Apple was reportedly caught off guard by AI demand for Mac Mini and Mac Studio. Is local AI changing what people buy? ERIK: Absolutely. The Mac is back to being a builder box. RAM, storage, thermals, local models. People aren't buying decoration. They're buying capacity. JOSH: That GPU World idea feels huge. Everyone gets frontier-level compute, all the time. What changes first? ERIK: Work stops being a thing you visit and becomes a thing that watches itself. Today, AI still feels like an errand for most people. Open Claude. Ask a question. Copy the answer. Paste it somewhere else. That is useful, but it is not the final form. ERIK: The final form is agents living inside the workflow. Watching Git. Watching logs. Watching NATS subjects. Watching tickets. Reading runbooks. Looking at service health. Making small calls all day. JOSH: Less chatbot, more infrastructure. ERIK: Exactly. A chatbot is a keyboard with opinions. Infrastructure has state. It has memory. It has guardrails. It has consequences. ERIK: That's where people get confused. They think the breakthrough is the model. The model matters, sure. But the system around the model matters more. PrimeBus is valuable because it turns events into action. Gandalf reviews code because it's subscribed to the right signals. PrimeSentinel watches stuck jobs because boring failures are still failures. JOSH: So the real question is whether your work emits signals. ERIK: Yes. If your business doesn't emit signals, your agents are guessing. That's not automation. That's a very expensive intern hallucinating in production. JOSH: Wait, really? ERIK: Really. Give an agent logs, deploy history, runbooks, alerts, Git commits, service ownership, customer tickets, and a safe action path. Now it can help. Give it a vague prompt and a blank terminal, and you've built a résumé-generating incident. JOSH: Where does this land in networking? ERIK: Networking is loaded with repeatable events. Interface down. BGP flap. Config drift. Certificate expiring. NSO package failed. Terraform plan changed. Kubernetes node pressure. None of that needs a human staring at a dashboard for nine hours. ERIK: Put the event on a bus. Have the agent gather context. Pull the last change. Check the runbook. Compare state. Propose a fix. If it's low risk, run it. If it's dangerous, route it to a human with the evidence already attached. JOSH: That's the part people worry about. Trust. ERIK: Trust the system, not the vibes. Big difference. ERIK: I don't want an agent that says, trust me bro. I want an agent that says, here is the event, here is the context, here is the command I would run, here is the rollback, here is why this is low risk, and here is the log trail. JOSH: That sounds slower. ERIK: It sounds slower until you compare it to a human waking up, finding the right Slack thread, asking who owns the service, opening the wrong dashboard, and then typing the command from memory. Humans are not exactly zero-latency infrastructure. JOSH: Fair. ERIK: The guardrails are the product. People want the magic moment where the agent fixes something. Cool. I care more about the moment where Gandalf blocks something stupid before it merges. JOSH: The blocked action is a win. ERIK: Huge win. Nobody celebrates the outage that didn't happen because it didn't make a fun screenshot. But that's the whole business. JOSH: Fastpotify feels smaller, but people clearly cared. Native app, standalone binary, no embedded Chromium. Why did that story pop? ERIK: Because everybody's expensive computer feels cheap now. You buy a good laptop, then Slack eats memory, Teams eats memory, your music app eats memory, the notes app eats memory, and some updater process is running like it owns the mortgage. JOSH: Painfully accurate. ERIK: Fastpotify hits because it's a reminder that software can still be small. One binary. Fast startup. Lower memory. Does the job. No parade. JOSH: Is that nostalgia for native apps? ERIK: Not nostalgia. Discipline. Native matters when the tool is supposed to disappear. ERIK: A music player should play music. A CLI should return fast. A dashboard should render without dragging half the internet behind it. When every small tool ships a browser, the machine starts feeling like a mall food court with fans. JOSH: How does that connect to AI systems? ERIK: Agents punish bloat harder than humans do. Humans tolerate slow tools because we get distracted. Agents call the same command again and again. They hit the same endpoint. They parse the same output. They don't forget that your CLI takes four seconds to start. JOSH: So bad tooling becomes a tax. ERIK: Exactly. If an agent calls a bloated tool hundreds of times, your waste becomes visible. Latency stacks. Memory stacks. Token usage stacks. Then somebody acts shocked that the automation bill looks like a car payment. JOSH: What's the fix? ERIK: Build boring tools that start fast and return structured output. JSON when possible. Exit codes that mean something. Logs that don't read like a panic attack. If a human and an agent both need the same tool, design it for both. JOSH: Give me an example from your side. ERIK: ScanBrief is a good example. I don't want a pretty web app first. I want reliable ingestion, scoring, ranking, and a clean output surface. The UI can come after the machine knows what happened. Same with BookForge behind InkEngine. The editorial backend has to track state cleanly or the writing agent loses the thread. JOSH: So native versus web is not really the point. ERIK: Right. The point is friction. Native can be bad. Web can be good. But we have normalized heavy software for tiny jobs, and agentic workflows expose that fast. JOSH: Apple being surprised by AI demand for Mac Mini and Mac Studio fits this too. Why are people suddenly looking at desktop Macs differently? ERIK: Local AI made hardware interesting again. For years, a lot of people just needed browser tabs and battery life. Now builders want memory. They want quiet sustained performance. They want local inference. They want a box that can run Claude Code, local models, Docker, Kubernetes, Selenium tests, and a pile of background workers without crying. JOSH: And the Mac Mini hits that? ERIK: For a lot of builders, yes. Small footprint. Good power draw. Enough performance to run real work. Mac Studio is the bigger hammer when you need more memory and more headroom. JOSH: Why local, though? Cloud models are still better for a lot of tasks. ERIK: Cloud models are great. I use them constantly. But local gives you speed, privacy, and control for the right class of work. Summarize logs. Classify events. Draft test names. Run small coding checks. Clean structured data. You don't need a frontier model for every thought. JOSH: That's a model selection point. ERIK: Big time. People use one giant model for everything, then complain it costs too much. That's like using a core router as a doorstop. Technically works. Weird life choice. JOSH: How do you split it? ERIK: Use the smallest model that can do the job reliably. Frontier model for architecture, gnarly debugging, final review, anything where judgment matters. Smaller local model for routing, summarizing, labeling, extracting, checking patterns. Deterministic code for anything that should not involve creativity. JOSH: Deterministic code still matters. ERIK: More than ever. AI doesn't replace normal engineering. It makes sloppy engineering hurt faster. If your schemas are a mess, your agent is going to trip. If your tests are weak, your agent will happily make nonsense pass. If your runbooks are vague, your agent will become vague at machine speed. JOSH: That's a little bleak. ERIK: It's good news if you build well. The boring engineering finally pays rent. JOSH: Where does Apple fit strategically? ERIK: Apple has the hardware people want, but the AI developer story still has to feel obvious. Builders don't want a keynote adjective. They want memory options, local model support, tooling that doesn't fight them, and machines that ship when demand shows up. JOSH: You think they underestimated builders? ERIK: I think the market moved from content creation to system creation. Video editors still matter. Designers still matter. But now the person buying a desktop box might be running agents, containers, test suites, local inference, and automation pipelines all day. JOSH: That's a different customer. ERIK: It's the same customer with a much heavier backpack. JOSH: What should builders take from all three stories together? ERIK: Compute is getting cheaper. Tools need to get lighter. Hardware is becoming local infrastructure again. Put those together and the advantage goes to the person who can wire systems together. ERIK: Not the person with the fanciest prompt. Not the person collecting AI tools like baseball cards. The person who turns events into action and leaves a trail. JOSH: That sounds like the whole show. ERIK: Pretty much. Build the bus. Put useful events on it. Add agents with limited permissions. Make them show their work. Let the boring stuff run itself. Then go build the next thing. JOSH: Before the tip, sponsor read. 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 JOSH: Alright, what's the AI pro tip today? ERIK: Make every agent write an action receipt before it touches anything. ERIK: Not a summary after the fact. Before. The receipt should include five fields: trigger, evidence, proposed action, risk level, rollback. JOSH: Why before? ERIK: Because before is where bad ideas are still cheap. After is where you're explaining to somebody why the database is shaped different now. JOSH: What does that look like in practice? ERIK: Say PrimeSentinel catches a stuck job. The agent doesn't just restart it. It writes: trigger is stuck job detection, evidence is no heartbeat since the last check, proposed action is restart worker, risk is low, rollback is stop the new worker and preserve logs. ERIK: Then your system decides. Low risk can run. Medium risk might need Gandalf review. High risk goes to a human. The point is the agent has to make its reasoning visible in a fixed format. JOSH: So it's not just prompt logging. ERIK: No. Prompt logs are archaeology. Action receipts are control surfaces. You can search them, score them, block them, audit them, and improve the next run. ERIK: Use JSON. Store it next to the event. Make the agent sign its work with model name, tool version, repo SHA, and timestamp. That's your tip. Use it. JOSH: We also drop daily market picks and automation tips on YouTube — search Build or Be Replaced. 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. ERIK: Build or be replaced. JOSH: If you want these signals in your inbox every morning, scanbrief.dev. See you tomorrow.