Build or Replaced
Today: Herdr is joining Y Combinator. The runtime stays open | Humans missed 1 in 3 threats approving AI agent commands across 40k game runs | AMD acquires Taalas to boost inference performance by etching models in silicon Episode date: 2026-08-07.
Download MP3 →JOSH: It's Friday, August 7. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson. ERIK: Taste is the new compiler, and bad taste is now a production outage. JOSH: Stick around — Erik's got an AI pro tip at the end about making agents prove their work before they touch your repo. [pause] JOSH: First up. Humans missed one in three threats while approving AI agent commands across forty thousand game runs. Erik, that's not a small miss. ERIK: No, that's the whole point. Human approval is not a security model if the human is tired, rushed, or doesn't understand the command. You need policy, sandboxing, test gates, and logs the agent can't edit. JOSH: Next. AMD bought Taalas, and the pitch is model weights etched into silicon. Why should builders care? ERIK: Because inference latency is the tax on every AI product. If model-specific chips get real, the cheap local agent gets a lot less weird and a lot more useful. JOSH: Third headline. OpenAI says GPT-5.6 Sol is improving in ChatGPT, and Luna access is expanding to free users. Does that change anything? ERIK: It changes the floor. When free users get better models, the excuse budget goes down. Your competitor with no team and no funding just got a better intern. [pause] JOSH: ScanBrief scored ninety-six items across fifty-six sources today. The one that jumped out to me was the agent safety study. ERIK: Same. Humans missing one in three threats is exactly what I expect when you ask a person to approve a wall of terminal commands. JOSH: Wait, really? ERIK: Yeah. Think about a command like npm install, then a postinstall script, then a curl, then a token in an env var. The approval prompt says, do you allow this. Most people are looking for obvious evil. rm -rf slash. Bitcoin miner. Something with a black hoodie. JOSH: And the real danger is boring. ERIK: Exactly. The danger is normal-looking. The danger is a dependency update that changes a script. Or a test command that writes outside the workspace. Or an agent that says, I need your GitHub token to finish the job. JOSH: That sounds like every AI coding setup right now. ERIK: Pretty much. And that doesn't mean don't use agents. I use them all day. I have twelve agents in the Bobaverse fleet right now. Neo, Homer, Bill, Echo, Gandalf, Claude, GPT. They're useful because I don't trust them blindly. JOSH: What's the difference between using them and trusting them? ERIK: Guardrails. PrimeBus doesn't say, hey Erik, does this look fine. It emits events. Tests pass or they don't. Gandalf blocks risky merges. The auto-merger has run 2295 attempts since June 5: 1507 merged, 788 blocked by Gandalf, 0 escalated to Erik. That's the number I care about. JOSH: Zero escalated? ERIK: Zero. And the blocked count is not failure. That's the system doing its job. If an agent proposes a sketchy fix, I want it caught before it gets near prod. Same way a network change hits lab, validation, diff review, and rollback checks before it touches a router. JOSH: How would that look for someone without your whole setup? ERIK: Start small. Put the agent in a folder with no secrets. Give it a fake repo first. Make it run tests. Make it show the diff. Then make another model review the diff with a nasty prompt. Tell it to find data loss, auth bypass, missing tests, and commands that should never run. JOSH: A nasty prompt? ERIK: Yeah. Don't ask, is this good. Ask, how does this break at 2 AM when nobody is watching. Totally different answer. [pause] JOSH: The article said humans were approving commands in game runs. Some people will hear that and say, well, that's not real production. ERIK: That's backwards. Games are useful because you can run forty thousand trials and measure behavior. Production just hides the same failures behind incident calls and calendar invites. JOSH: That's bleak. ERIK: It's honest. Humans are great at judgment. They're bad at being a firewall. If your agent safety plan is Josh staring at a command and saying vibes are good, congratulations, you invented a sleepy SOC analyst with worse coffee. JOSH: Fair. ERIK: Use humans for policy. Use systems for enforcement. That's the split. [pause] JOSH: Second deep story. AMD acquires Taalas. The summary says model-specific chips with weights directly in silicon, no HBM, and up to about seventeen thousand tokens per second. ERIK: That number is the bait, but the architecture is the story. HBM is expensive, power-hungry, and it keeps inference tied to memory movement. If you can burn the weights into silicon for a known model, you cut a huge part of the trip. JOSH: But doesn't that make the chip less flexible? ERIK: It does. That's the trade. A GPU is a general-purpose monster. Great when models change every week. A model-specific chip is more like an appliance. Less flexible, maybe much faster, maybe much cheaper per token if the workload is stable. JOSH: Where does that matter first? ERIK: Repetitive inference. Routers. Customer support. Code review. Document extraction. Monitoring. Anything where you're hammering the same model with similar requests all day. JOSH: So not frontier research. ERIK: Not first. This is production plumbing. Nobody gets excited about plumbing until the bill shows up. JOSH: You run PrimeRouter as an LLM gateway. Would this change your routing? ERIK: Absolutely. PrimeRouter already thinks in tiers. Cheap model for boring work, strong model for judgment, fallback provider when something acts up. If a fixed-function inference box gives me low latency for classification, routing, summarization, or policy checks, I don't care if it can't write poetry about Kubernetes. JOSH: It just needs to do one job. ERIK: One job, fast, cheap, repeatable. That's how infra wins. The fancy demo gets the retweet. The boring box saves money every minute. JOSH: Is this bad for Nvidia? ERIK: Not today. Nvidia is still the default answer because CUDA is gravity. But AMD buying Taalas says the market is splitting. Training, frontier experimentation, high-end flexible compute over here. Repeated inference workloads over there. Those are different businesses. JOSH: And builders should watch pricing? ERIK: Watch latency, power, and tokens per dollar. Not press releases. If your app waits eight seconds for an answer, users blame you, not your provider. If your monthly AI bill eats the margin, same story. [pause] JOSH: How do you decide what model gets what job in your own stack? ERIK: I separate tasks by risk. ScanBrief can use cheaper passes for scoring and dedupe, then stronger models for summaries that actually go out. PrimeBus uses stricter review for code changes because that can touch production. PrimeRestorer is even more boring on purpose. Backups are not a place to be creative. JOSH: That's a sentence. ERIK: Backups should be boring. Restores should be rehearsed. If an agent wants to modify backup logic, it gets treated like a guy walking into the data center with bolt cutters. JOSH: That escalated quickly. ERIK: Good. It should. Automation without blast-radius thinking is just faster damage. [pause] JOSH: Third story. The essay title was Taste Is All That's Left. AI can generate the artifact, so the human value moves to picking, judging, and correcting. Do you buy it? ERIK: Mostly. The idea-to-artifact gap collapsed. That's real. You can describe a service, get a repo, run it, and have something alive before lunch. But taste is not vibes. Taste is knowing what should exist, what should not exist, and what will hurt later. JOSH: Give me an example. ERIK: An agent will happily build a dashboard with twelve charts, three filters, a floating assistant, and a settings panel nobody asked for. Looks productive. Then the user opens it and can't answer the one question they came for. JOSH: So taste is restraint? ERIK: Restraint plus judgment. In networking, I learned this with NSO. You can automate a bad process perfectly. That doesn't make it good. It just makes the bad process faster and harder to stop. JOSH: That's uncomfortable. ERIK: Good. Builders need to feel that. AI makes output cheap. It does not make decisions cheap. The decision is still the hard part. JOSH: Does that mean coding skill matters less? ERIK: Syntax matters less. Engineering matters more. Knowing how systems fail matters more. Knowing where to put a queue, when to use NATS, when a database transaction matters, when Selenium is fine and when it's a liability. That's still experience. JOSH: Where do juniors fit in that world? ERIK: They need to build more, not less. The old path was read docs, write small code, wait for review. The new path is build ten versions, compare them, break them, explain why one survives. If you can't judge the output, the model owns you. [pause] JOSH: That sounds like the title of the show. ERIK: Yeah. Build or be replaced is not anti-AI. It's anti-passenger. If you are waiting for a perfect tool, you're already behind. If you're building systems around the tools, you're dangerous in the good way. JOSH: How does that show up in your day? ERIK: Today, PrimeBus processed fourteen automation events across one project. That's not glamorous. But every event is a chance to remove a manual step. ScanBrief pulled the signal. PrimeRouter decides where model work goes. PrimeBus moves events. Gandalf blocks bad merges. Build or Be Replaced turns the work into a daily record. JOSH: Episode eighty-three. ERIK: Episode eighty-three. That's the part people skip. They want the architecture diagram. Cool. But the real system is daily shipping. Small commits. Small tests. Small fixes. Then one day you look up and there are 143 services running on the production server. JOSH: That sounds like a lot. ERIK: It is a lot. But it's not chaos if telemetry exists. The problem is not having many services. The problem is not knowing what they're doing. JOSH: That's where most teams are? ERIK: Yep. They have alerts, but no story. Logs, but no ownership. Runbooks in a wiki nobody opens. Then they buy an AI assistant and ask it to save them. JOSH: And it doesn't. ERIK: It gives them faster confusion. You need events. You need naming. You need health checks. You need rollback. You need a human who cares enough to make the machine boring. [pause] JOSH: Coming back to the OpenAI model update. Better free access means more people can build. Is that mostly good? ERIK: Yes. More builders is good. More noise too, but that's fine. The floor going up forces everyone to compete on taste, distribution, reliability, and speed. JOSH: Not just prompts. ERIK: Prompts are table stakes. The wrapper app era where someone calls an API and adds a textarea is basically cooked. The next useful stuff is agentic workflow with real constraints. It connects to your repo, your tickets, your logs, your CRM, your network controller. It does a job and leaves evidence. JOSH: Evidence keeps coming up. ERIK: Because evidence is the difference between automation and magic tricks. If an agent says it fixed something, show me the test. Show me the diff. Show me the event. Show me the rollback path. Otherwise it's a chatbot wearing a hard hat. [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: Make your agent produce a proof packet before you accept its work. Not a paragraph. A packet. One, the command it ran. Two, the files it changed. Three, the tests it ran and the result. Four, the risk it thinks remains. Five, the rollback command. JOSH: That's pretty concrete. ERIK: It has to be. Add this to your coding prompt today: before final answer, give me changed files, verification commands, test output, and rollback steps. If any part is missing, do not claim done. JOSH: And for higher-risk changes? ERIK: Use a second model as a reviewer. Tell it to assume the first model missed something. Make it hunt for auth issues, data loss, race conditions, secrets, and missing tests. Then only merge after the proof packet and the review both pass. JOSH: That's usable today. ERIK: Yep. Don't ask AI to be trustworthy. Build a workflow where trust is earned by evidence. That's your tip. Use it. [pause] JOSH: Binge all five episodes this weekend plus our YouTube shorts — links at buildorbereplaced.dev. [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.