Build or Replaced
Today: H.264 Is Magic (2016) | We're releasing the weights of Mistral Large 4 | Binance-hacker-1029 – $5.6B stolen from Binance Episode date: 2026-10-30.
Download MP3 →JOSH: It's Friday, October 30. This is Build or Be Replaced, powered by ScanBrief.dev. I'm Josh, here with Erik Anderson. ERIK: Open weights, stolen money, and agents playing video games. Today is about who controls the model. JOSH: Stick around. Erik's got an AI pro tip at the end about testing a new model without trusting the benchmark. [pause] JOSH: Episode 138. Erik, what did ScanBrief find this morning? ERIK: Ninety-eight items scored across 56 sources. Average relevance 51.5. Nothing screamed, but three stories are worth your time. JOSH: Headlines first. [beat] JOSH: There's a post called "H.264 Is Magic." It's from 2016, and it's back on the front page. Why does it keep coming back? ERIK: Because it's the best explanation of how video compression works. Most of what you watch is a clever lie about what changed between frames. Engineers love it because it's the same trick as good automation. Don't resend what didn't change. JOSH: Next one. Someone got a tiny world model running games on a Raspberry Pi Pico 2 W. ERIK: That chip costs a few dollars and has about 520 kilobytes of RAM. A learned model predicting game frames on it is a stunt. It's also a preview. The floor for running intelligence keeps dropping. JOSH: And there's one for the network crowd. "If IPv4 Were Land." ERIK: Perfect Friday read for anyone who's done address planning. Four billion addresses, and we still treat them like a city with no zoning. Land analogies work because everybody understands scarcity and squatting. Every NOC engineer has argued about a /24 at some point. [pause] JOSH: Okay. Big one. Mistral released the weights of Mistral Large 4. A one-trillion-parameter multimodal model. What happened? ERIK: Mistral open-sourced the weights. The post says coding, defensive security and reasoning all jumped. It also has three inference modes, so you pick your tradeoff between speed, cost and depth. JOSH: Wait, a trillion parameters. Can regular people run that? ERIK: Not on your laptop. You need serious hardware, or a host that already has it. But the weights being public changes the negotiation. If one vendor gets pricey or goes down, there's a credible alternative on the table. [beat] JOSH: So does that matter to someone building with AI? ERIK: It matters most if you don't depend on one provider. I run PrimeRouter as the gateway for the whole fleet. Every call has a priority tier and a chain of backends. If the first one is down or capped, it walks to the next. The app doesn't know and doesn't care. JOSH: So a new model is just another entry in the chain. ERIK: Right. That's the point of a gateway. I don't rewrite anything when a new model shows up. I add a backend, keep it dark, and watch it before it touches anything important. A model getting good on a leaderboard is not the same as it being safe in my pipeline. JOSH: Why not just switch if it's better? ERIK: Because "better" is an average. My workload is specific. Reviewing diffs, fixing failing tests, summarizing feeds. A model can win on a benchmark and still be bad at the one thing I need. [pause] JOSH: Let's talk about Binance. A post claims five point six billion dollars was stolen. Is that real? ERIK: Careful here. It's a Hacker News post, and the claim reads as alleged. Five point six billion would be the biggest theft in the history of the industry. Until an exchange or a credible outlet confirms it, treat it as a headline, not a fact. JOSH: Good point. But if it is true? ERIK: Then the lesson is the same one it always is. Concentration kills you. One exchange, one set of keys, one failure domain. I've held Bitcoin since 2014 and I run a full node. Not because I'm a purist. Because I don't want to depend on anybody's front door being locked. JOSH: Does that apply outside of crypto? ERIK: Everywhere. Backups are the same story. PrimeRestorer handles my database and system backups and ships offsite to S3. Local copies are convenience. The offsite copy is the one that matters when something goes wrong. [beat] JOSH: So the boring answer is the right one. ERIK: It usually is. Boring means redundant, tested, and separated. Exciting means a single point of failure with good marketing. [pause] JOSH: Last deep dive. Someone ran AI agent swarms in Minecraft and Mindustry. Why does this belong on a show about work? ERIK: Games are cheap test labs. You get a messy environment, shared resources, and goals that need coordination. If agents can't divide labor on a base-building map, they won't do it on your infrastructure. JOSH: And do they work together? ERIK: Swarms are only as good as the coordination layer. The failures are boring. Two agents grab the same task. One waits on something that never finishes. Nobody owns the result. JOSH: Sounds like an office. ERIK: It does. The fix is the same as in an office. Clear ownership and a shared record of what happened. My fleet has 12 agents, and they talk over PrimeBus. Events go on the bus, and whoever owns that lane picks them up. Nobody freelances. JOSH: Does it ever go wrong? ERIK: Constantly, which is why the guardrails exist. The auto-merger has made 2,295 attempts since June 5th. 1,507 merged. Gandalf blocked 788. Zero escalated to me. People look at the blocked number and think that's failure. It's the reviewer doing the job. JOSH: So the blocked count is the story. ERIK: Yes. A pipeline that merges everything isn't confident. It's careless. I'd take a gate that says no 788 times over a fleet that ships whatever it writes. [beat] JOSH: And 140 services running on the prod server. ERIK: 140 services, 140 projects emitting telemetry. At that size you don't watch it by hand. PrimeDash shows me everything on the LAN, and the bus tells me when something's off. My job is deciding what's allowed to happen, not watching every process. [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: Build a private benchmark of ten real tasks from your own work. Not puzzles from the internet. Actual things you did last week. A failing test you had to fix. A messy email thread you had to summarize. A config diff you had to review. ERIK: Save the input and what a good answer looks like. When a new model drops, like Mistral Large 4 today, run the same ten through it and compare against what you use now. Score pass or fail. Takes twenty minutes. ERIK: Do it through a gateway or a single wrapper script, so swapping the model is one line. If the new one passes eight of ten on your tasks, you've got evidence. If it passes five, you just saved yourself a bad migration. Public benchmarks tell you how a model does on someone else's homework. ERIK: That's your tip. Use it. [pause] JOSH: Binge all five episodes this weekend plus our YouTube shorts — links at buildorbereplaced.dev. 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.