Build or Replaced
Today: MiMo v2.6 | I don't want to read what you didn't write | AI coding has made CI a bottleneck, so we reworked ours to keep up Episode date: 2026-09-22.
Download MP3 →ERIK: ScanBrief scored ninety-five items across fifty-six sources before either of us was up. Today's theme is speed outrunning the process built to check it. JOSH: It's Tuesday, September 22nd. This is Build or Be Replaced, powered by ScanBrief dot dev. I'm Josh, here with Erik Anderson. ERIK: Episode one twelve. Still standing. [beat] JOSH: Stick around, Erik's got an AI pro tip at the end about why routing everything to your biggest model is actually costing you. ERIK: And it's a cheap fix once you see it. [pause] JOSH: Let's do headlines. First up, Hacker News, MiMo v2.6 just dropped. Over nine hundred upvotes, four hundred comments. What's driving that? ERIK: That's a community that's hungry for an open model that ships real follow-ups fast. A reaction that size means people are already running it in production, not just reading the changelog. [beat] JOSH: Next, "I said no and Apple said yes." Cryptic headline. What's your take? ERIK: Apple's reversing a policy it held for years. When a platform flips in public like that, somebody inside lost an argument they used to win. JOSH: That's a strange hill to lose on. ERIK: Those hills are usually load-bearing. Nobody flips a years-old policy over nothing. [beat] JOSH: Third, "Can gzip be a language model?" That sounds like a joke. ERIK: It's not really a joke. Compression and prediction are the same math underneath. If you can compress text well, you can predict the next token well. People forget that's the same problem. [beat] JOSH: Okay, one more before we go deep. A robotics lab posted benchmark numbers that beat their own claims from three months ago. ERIK: That pace is the story more than the numbers. Three months used to be a rounding error in robotics. Now it's a full generation. [pause] JOSH: Let's go deep on the one that's basically your whole job. "AI coding has made CI a bottleneck, so we reworked ours to keep up." ERIK: That's my day, every day. When Claude or Codex can throw ten pull requests at you before lunch, the bottleneck stops being who writes the code. It becomes who verifies it. JOSH: So how do you handle that on your end? ERIK: Gandalf. It's the review gate that sits between whatever an agent writes and anything landing on Neo. Since June fifth, the auto-merger's run twenty-two ninety-five attempts. Fifteen-oh-seven merged clean. Seven eighty-eight got blocked. JOSH: Wait, seven hundred eighty-eight blocked? That sounds bad. ERIK: That's not a failure number, that's the guardrail working. Sixty-five point seven percent merge rate isn't the headline. Seven hundred eighty-eight times the pipeline caught something before it touched prod, and zero of those got escalated to me. I never even saw them. [beat] JOSH: So the bottleneck everybody else just hit, you already dealt with? ERIK: I didn't get rid of it, I moved it. It used to be me reading diffs all day. Now it's an agentic reviewer reading files on demand, and that thing has its own failure mode. It can oscillate, pass a diff, then flag the same diff on a second look. So the rule on my end is small diffs, one change at a time. A small diff, the reviewer can hold the whole thing in its head. A giant diff, it starts guessing, and guessing is how you get a false block or worse, a false pass. JOSH: How does that compare to a normal engineering team without something like PrimeBus running review? ERIK: A normal team either throws more people at review or quietly lowers the bar as volume goes up. I throw more automation at it and keep the bar exactly where it was. Neo's got a hundred fifty-three services running right now, and Gandalf holds the same standard on all of them. JOSH: No exceptions for the small stuff? ERIK: Especially not the small stuff. Small-looking changes are exactly where the costly incidents come from. A one-line config edit can take down a chain just as easy as a rewrite. [pause] JOSH: Next deep dive, and this one's unsettling. "Spymarks, Not Watermarks." Google's SynthID. ERIK: A normal watermark says "AI made this" out loud, everybody can see it, that's disclosure. SynthID embeds a hundred thirty-six bit signal you can't see or hear, and it can carry a sixty-four bit ID tied to a specific user. That's not a watermark. That's a tag. JOSH: What's the practical difference? ERIK: A watermark tells you something was generated. A tag like that tells someone who generated it. If every image or clip you make carries an identifier back to your account, that's tracking infrastructure, not a disclaimer. JOSH: Does that touch anything you're running? You score a ton of content every morning. ERIK: ScanBrief pulls from fifty-six sources, and more of that content every month is AI generated on the source side, not just summarized on my side. If platforms start embedding user-level tags at scale, dedup and scoring get a lot more interesting. Suddenly you can tell which outlet actually wrote a piece versus which one ran it through a model and reposted it. JOSH: That connects to "I don't want to read what you didn't write" too, doesn't it? ERIK: Same problem, two angles. One's about whether you disclosed it. The other's about whether someone can prove it after the fact without your permission. SynthID's in the second camp, and that camp is the one nobody voted for. JOSH: So which camp are you in? ERIK: Disclosure, always. I'll tell you what's automated. I won't build something that fingerprints you without asking. [beat] JOSH: Okay, last one, and I think this is just funny to you personally. "HERMES radio enables voice and data communication over vast distances." ERIK: I've got a Hermes running on a Mac M3 in my own lab. Different Hermes, same instinct though. Get a message through over distance, without one point of failure taking the whole thing down. JOSH: Wait, you named something Hermes before this story even existed? ERIK: Yeah, mine's a local agentic backend, qwen3-coder, running about seventy tokens a second, no cloud call needed. But the concept in that headline is exactly what I care about with PrimeRouter. A message needs to land even when the first path is down. JOSH: How does PrimeRouter actually do that day to day? ERIK: Priority-tier routing with failover. If a call needs Claude and that slot's busy or unreachable, PrimeRouter doesn't just error out, it walks the chain to the next backend. Same idea as a radio grabbing a different frequency when the first one's jammed. I've got a hundred forty projects emitting telemetry into that mesh right now, so when something goes quiet, I see it fast. JOSH: A hundred forty projects. That's a lot to keep straight in your head. ERIK: That's why the fleet exists. Twelve agents, Claude and GPT both, split across Neo, Homer, Bill, Echo, and Gandalf doing review. Nobody's staring at a hundred forty dashboards by hand. JOSH: So the radio headline and your actual stack basically rhyme. ERIK: They do. Distance, interference, failover — swap radio waves for API calls and it's the same engineering problem people have been solving for a century. We just do it in software now. [pause] JOSH: Quick word from today's sponsor. 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: Stop sending everything to your biggest model. I route by priority tier in PrimeRouter — quick classification jobs go to a smaller, faster backend, and only the calls that actually need deep reasoning walk the chain up to the top-tier slot. JOSH: How much does that actually save you? ERIK: It's not just cost, it's latency too. A summarization task doesn't need the same horsepower as a code review. If you're running every request through your most expensive model by default, you're paying premium price for work that a cheaper model handles fine. Match the model to the job, and reserve the expensive one for when it actually matters. ERIK: That's your tip. Use it. [beat] JOSH: We also drop daily market picks and automation tips on YouTube — search Build or Be Replaced. [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.