Build or Replaced
Today: Gemini 3.8 Live and 3.8 Live Extended Thinking | Introducing System One Models and Jev | We got admin access to Baseten's production GitHub Episode date: 2026-09-16.
Download MP3 →1540 words — within the 1,500-2,200 range. Here is the final script: JOSH: It's Wednesday, September 16th. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson. ERIK: Episode 108, and today the story isn't "AI wrote code" — it's "AI made a decision and nobody double-checked it." That's the whole show. JOSH: Stick around — Erik's got an AI pro tip at the end about picking the right model for the job instead of just grabbing whatever's newest. [pause] JOSH: Let's do headlines. First up — the Netherlands, a rail disruption, and the word "sabotage" right in the headline. ERIK: Physical infrastructure attack, signaling systems, not a hack. Good reminder that automation failures aren't always digital. You can harden every API you own and someone still cuts a cable in a trackside cabinet. [beat] JOSH: Next — someone built an e-ink picture frame that listens for birds outside and draws them like it's 1850. ERIK: That's a great weekend build. Acoustic detection, generative art, e-ink so it doesn't burn power sitting on a shelf. That's the whole hobbyist AI stack in one frame. [beat] JOSH: And Apple's pushing something called Reference Image — a way to verify a photo actually hasn't been touched. ERIK: That one matters more than people think. Once every photo can be faked, you need a chain of custody for the real ones. That's not a gimmick, that's infrastructure. Same idea as signed commits — you're not proving the content is good, you're proving nobody touched it after the fact. [beat] JOSH: One more — a fintech startup had their AI agent auto-approve a six-figure wire transfer with no human in the loop. Nobody caught it for two days, until the bank flagged it. ERIK: That's the whole episode in one headline. The agent did exactly what it was built to do — pattern-match and approve. Nobody put a checkpoint in front of it. JOSH: Two days is a long time to not notice a six-figure transfer. ERIK: That's the part that gets me. It's not that the AI made a call, it's that nothing downstream was watching it make that call. That's a review-loop problem, not a model problem, and it's the exact same failure mode as the Baseten story we'll get to. [pause] JOSH: Alright, let's get into the meat of it. JOSH: Google dropped Gemini 3.8 Live and a "3.8 Live Extended Thinking" version. What's actually new there? ERIK: Parallel reasoning while it's talking to you, interruption handling, and background task execution. You can talk over it like a person, and it keeps working on something else while it answers you. JOSH: Wait, background task execution — it keeps doing homework mid-conversation? ERIK: Basically. That's the piece I care about. Right now if I ask a model to do something long, I wait, or I poll it. A model that holds a live conversation and runs a background job at the same time is closer to how PrimeBus already works. Something fires an event, an agent picks it up, I keep moving. [beat] JOSH: So Google's catching up to what you already built at home? ERIK: In spirit, yeah. I've got 12 agents in the Bobaverse fleet right now — Neo, Homer, Bill, Echo, Gandalf, mix of Claude and GPT — and none of them wait for me to finish a sentence. They just work. Gemini trying to do that in one voice session is Google admitting that's the right shape. JOSH: How would you actually use it? ERIK: Voice ops. If I'm walking through the house and I say "check the Neo box," I don't want five seconds of silence while it thinks. I want it talking to me while it's pulling logs in the background. That's dispatching, not chatting. The difference between Gemini doing this in one voice session and what I run is that my dispatching happens over NATS between separate processes. Gemini's compressing that whole event bus into a single model call, which is a genuinely different engineering bet. [pause] JOSH: Next one's interesting — a company called TypeSafe AI, a model called Jev, and the pitch is it skips free-form text and just outputs structured decisions. Supposedly a hundred times cheaper. ERIK: This is the one I actually care about today. Most of what I run doesn't need prose. It needs a yes, a no, a category, a score. ScanBrief scored 93 items across 56 sources again this morning — that's not creative writing, that's classification. Relevance score, source, topic. A model built to spit out calibrated structured output instead of a paragraph is the right tool for that exact job. JOSH: So you'd swap it in for stuff you're already doing? ERIK: For the scoring layer, maybe. Right now a general-purpose model does that job, and it works, but it's overkill — like using a sledgehammer to hang a picture. And it's not just ScanBrief. email-reactor gets client website change requests over email, and step one is always the same decision: is this a typo fix, a content swap, or something that needs a human to scope it. That's a three-way classification, not an essay. Jev-shaped tasks are everywhere once you start looking for them. JOSH: If Jev's genuinely a hundred times cheaper, that's real money at volume, right? ERIK: And I run a lot at volume. PrimeBus alone has logged 2,295 auto-merge attempts since June. Multiply any per-call savings by a number like that and it stops being a rounding error. Even a small drop in per-decision cost compounds fast when the decision runs thousands of times a month instead of once. JOSH: What's the catch? ERIK: Same catch as always — "comparable intelligence" is doing a lot of work in that sentence. A hundred times cheaper doesn't mean anything if the accuracy drops even a few points on the categories that actually matter. I'm not trusting a brand new architecture with my merge pipeline on day one. Test it in the lab first, prove the win on something low-stakes like ScanBrief scoring, then move it to something that actually touches production. Same as everything else I ship. [pause] JOSH: Last one's a little scarier — somebody got admin access to Baseten's production GitHub. ERIK: And Baseten's an inference hosting company. So this isn't "someone read some code," this is "someone had the keys to how a bunch of companies serve their models." That's a supply chain problem, not a bug. JOSH: How do you even defend against that? ERIK: You assume it happens and you build so no single compromised account can do damage on its own. That's basically the whole reason Gandalf exists in my pipeline. PrimeBus doesn't auto-merge because a model said "looks good," it merges because it passed review, and if it doesn't, it gets blocked. 788 blocked out of those 2,295 attempts this cycle. Zero escalated to me, because the guardrail caught it before it needed a human. JOSH: So the blocked number is actually the good number. ERIK: That's the whole point people miss. A high block rate isn't the system failing, it's the system doing its job. If I ever need to rotate an admin key on Neo, it doesn't happen with one set of hands, and it doesn't happen on a standing credential either — short-lived tokens, scoped to exactly what that job needs, gone when the job's done. Same rule as a merge — nothing with that much blast radius moves on a single approval or a key that outlives its task. If Baseten had something in front of that admin account asking "does this actually look like normal Baseten behavior," none of this happens. [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 [beat] JOSH: Alright, what's the AI pro tip today? ERIK: Stop defaulting to your biggest model for everything. Today's headlines are literally about that — a decision model built cheap and fast for yes-no-score tasks, and a live model built for dispatch instead of essays. Look at what you're actually asking for. If it's a classification or a routing decision, use something small and fast. Save the expensive model for the part that genuinely needs reasoning. Your bill and your latency both go down. JOSH: Do you actually run that split somewhere already, or is this theory? ERIK: email-reactor already does this today. Triage — typo, content swap, or needs-a-human — runs on the cheap fast path. It only escalates to the bigger model once it's decided the request actually needs a drafted response. That split is why the whole thing runs unattended overnight and I wake up to a done queue instead of a pile of guesses. That's your tip. Use it. [pause] ERIK: If you're building toward financial independence through automation, my first book walks through the whole path. Free chapter at erikandersonbook.com. [beat] 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.