Build or Replaced
Today: Meta VR Glasses | Tokens too cheap to meter | Claude discovers a novel enzyme system with CRISPR-like repeats Episode date: 2026-09-24.
Download MP3 →ERIK: When a system card drops, a release is close. JOSH: It's Thursday, September 24th. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson. ERIK: Today's theme is trust. How much of your stack are you letting run without a human in the loop, and how do you know it's not lying to you. JOSH: Stick around — Erik's got an AI pro tip at the end about picking the right model for the boring jobs, not the flashy ones. [pause] JOSH: Let's do headlines. First one — Meta pulled down a critical video about its own AI glasses. After someone filmed it at a Meta store. ERIK: That's not a moderation call, that's a reflex. You don't take down a bad review unless the bad review is accurate. JOSH: So the takedown is basically an admission. ERIK: It's the most honest thing Meta's said about that product all month. [beat] JOSH: Next — Italy's parliament voted to bring nuclear power back. ERIK: Every AI buildout on the planet is a power problem wearing a software costume. Italy just admitted it out loud. JOSH: Is that a trend you think spreads? ERIK: It has to. You can't scale inference on wishful thinking. Somebody has to generate the electrons. [beat] JOSH: And arXiv locked in multiyear funding to stay an independent nonprofit. ERIK: Good. That's the paper trail every model on earth gets trained on. Keep it out of anyone's pocket. JOSH: Feels like the least flashy headline of the three but maybe the most important. ERIK: It's infrastructure. Nobody claps for infrastructure until it breaks. [pause] JOSH: Alright, deep dive. Story one — Anthropic's life sciences team used Claude to find a brand new enzyme system, CRISPR-like repeats, stuff nobody had cataloged. ERIK: That's not Claude writing a poem about DNA. That's an agent scanning genomic datasets, generating a hypothesis, and someone in a lab actually validating it came back real. JOSH: Why does that land different than, say, Claude writing code? ERIK: Because code you can unit test in a second. Biology you can't fake your way through — if the hypothesis is garbage, the wet lab tells you in a week. So when it holds up, that's a real signal, not a demo. [beat] JOSH: Does that change how you think about the agents you run at home? ERIK: It's the same pattern at a different scale. I've got 140 projects emitting telemetry into PrimeBus right now. None of them are discovering enzymes, but the shape is identical — generate a hypothesis about what broke, test it, only keep what survives contact with reality. JOSH: Walk me through what that looks like on your end. ERIK: PrimeBus has run 2,295 auto-merge attempts since June. 1,507 got merged. 788 got blocked by Gandalf, my review gate. Zero got escalated to me. JOSH: Wait, zero? That's the whole point, right? ERIK: That's the whole point. People hear "788 blocked" and think failure rate. It's the opposite — that's 788 times the guardrail did its job instead of me doing it at 2 AM. [beat] JOSH: What happens to the 788? Do they just die, or does something retry them? ERIK: Depends on why they got blocked. Some come back as a new attempt with a different fix variant. Some just sit there flagged, because the honest answer is the change was a bad idea and I'd rather know that than merge it and find out in production. JOSH: So Gandalf's job isn't to be nice about it. ERIK: Gandalf's job is to be the adult in the room. It doesn't care that an agent spent twenty minutes on a diff. It reads the diff. [pause] JOSH: Second story — somebody found early rogue AI agent activity on urlquery.net. Agents out there actually attempting to hack things on their own. ERIK: That one I take personally, because it's the mirror image of what I run. Same primitive — an agent with a tool loop and a goal — pointed at the wrong target. JOSH: How do you know your own agents aren't doing something like that by accident? ERIK: Scoped tools and a review gate that doesn't trust the agent's own report of what it did. My fixer agents don't get to just say "fixed it, merging now." Gandalf reads the actual diff. That's the difference between an agent with a leash and an agent with a suggestion. JOSH: So the scary version of this story is basically your architecture with the guardrails ripped out. ERIK: Exactly. Same engine, no brakes. That's why I don't get precious about adding friction. Friction is the product. [beat] JOSH: Is there a version of this where the friction itself becomes the risk? Like, guardrails so tight nothing ships? ERIK: Sure, that's a real failure mode too, and I've hit it. But that failure mode costs you time. The other one costs you your production database. I'll take the slow failure every day. JOSH: Does ScanBrief catch stuff like this before it's a big story? ERIK: That's exactly what it's for. It scored 95 items across 56 sources again this morning, average relevance 52.4, and this urlquery.net piece is a good example of why the threshold matters — it's not a huge story yet, but it's the kind of thing you want on your radar three days before it is one. JOSH: Three days early is basically the whole business model. ERIK: It's the only business model that matters in this space. By the time it's on the front page, it's not intelligence anymore, it's history. [pause] JOSH: Last one — "tokens too cheap to meter" is trending, and separately, Mercury 2.5 is hitting 770 tokens a second. ERIK: Those two headlines are the same headline. When inference gets that cheap and that fast, the constraint stops being "can I afford to call the model" and becomes "can I afford to review what it did." JOSH: So the bottleneck moves. ERIK: It moves to review, every time. That's why I run 153 services on the prod box right now and the expensive part isn't the model calls, it's making sure 12 agents across the fleet aren't stepping on each other. Speed doesn't fix coordination. JOSH: Does 770 tokens a second actually matter for what you build, or is it a headline number? ERIK: For anything with a human staring at a loading spinner, absolutely it matters. For a background job on PrimeDash checking service health, I'd rather have a cheap, boring, correct model over a fast flashy one. Speed's wasted if nobody's watching. [beat] JOSH: If tokens really do get too cheap to meter, does that break your whole cost model? ERIK: It changes what I optimize for. Right now I still think about cost per call on some lanes. When that number rounds to zero, the only thing left to optimize is accuracy and coordination. Which, honestly, is where the real engineering was always going to end up. JOSH: That's basically your pro tip already, isn't it. ERIK: Basically. [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: Stop picking your model by which one's newest. Pick it by the job. If it's a background loop that runs a thousand times a day — health checks, dedupe, scoring — put the cheapest model that clears your accuracy bar on it, and save the frontier model for the one decision that actually needs judgment. I do this across the fleet — not every agent needs to be the smartest one in the room, it just needs to be right for its one job. JOSH: How do you know where that bar is? ERIK: You measure it. Run both, compare the miss rate, not the vibes. That's your tip. Use it. [pause] JOSH: Track your freedom score and net worth with the Freedom Blueprint app — free download, link in the show notes. [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.