Build or Replaced
Today: Gemini 3.8 Flash and 3.8 Flash Cyber | Can I opt out of my input or output data being used for training? | Three sites made 215,128 “best software” pages for AI. Perplexity cites them Episode date: 2026-09-03.
Download MP3 →JOSH: It's Thursday, September 3. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson. ERIK: AI search is eating its own spam now, and somehow that's not even the weirdest signal today. JOSH: Stick around — Erik's got an AI pro tip at the end about making agents show their work before they touch production. [pause] JOSH: First headline. Google DeepMind released Gemini 3.8 Flash and a cybersecurity version called 3.8 Flash Cyber. Is this just another model bump? ERIK: No. Flash models matter because they're the cheap, fast ones builders actually wire into systems. If 3.8 Flash can reason better without blowing up cost, that changes what you can run on every event, every push, every alert. [beat] JOSH: Second one. Three websites apparently made over 215,000 best-software pages for AI, and Perplexity cited them a lot. That's gross, right? ERIK: That's the internet adapting to AI search. Not humans writing for humans. Machines writing bait for machines, then other machines citing it like research. Beautiful little trash fire. [beat] JOSH: Third headline. Google avoided a breakup of its ad tech business. ERIK: That's huge for the ad market, but also for anyone building on top of Google traffic. The pipes stay together, which means the incentives stay together. Regulators took a swing, but the machine is still mostly intact. [pause] JOSH: Let's start with Gemini. Why does a fast model release matter more than a giant flagship model? ERIK: Because nobody serious puts the most expensive model on every tiny task. You don't need a monster model to classify a log line, check a diff, summarize an alert, or decide whether a Selenium run needs a retry. You need cheap reasoning that can run constantly. [beat] ERIK: That's why Flash-class models are interesting. They're the working models. The ones you can afford to put in the boring parts of the system. And the boring parts are where the money is. JOSH: What do you mean by boring parts? ERIK: Triage. Routing. Enrichment. Sanity checks. Test failure clustering. Security review. Ticket updates. Stuff nobody wants to do, but the business dies slowly when it doesn't get done. [beat] ERIK: In my world, PrimeBus is pushing telemetry all day. This morning's live stats show 140 distinct projects have emitted telemetry to PrimeBus. That's not a demo app. That's a lot of little signals moving around. JOSH: And you wouldn't send every one of those to the biggest model. ERIK: Exactly. That would be financially dumb. You use model tiers like network tiers. Not every packet needs a core router brain. Some just need a fast ACL check and a clean next hop. [beat] ERIK: Gemini 3.8 Flash Cyber is interesting because cybersecurity tasks are usually pattern-heavy but consequence-heavy. You want speed, but you also need judgment. Is this dependency bump normal? Is this diff touching auth? Did somebody accidentally log secrets? Is this Terraform change about to open the front door? JOSH: So Cyber isn't just branding? ERIK: Could be branding. Always assume some branding. But training a fast model around security workflows is the right direction. Security doesn't fail because nobody bought a platform. It fails because the alert sat there, the diff looked harmless, and the tired human clicked approve. [beat] JOSH: Where would you put something like that in your setup? ERIK: Before Gandalf, around Gandalf, and after Gandalf. Gandalf is my code reviewer agent. PrimeBus auto-merger has run 2295 attempts since 2026-06-05: 1507 merged, 788 blocked by Gandalf, 0 escalated to Erik. That's the architecture I care about. [beat] ERIK: The blocked count is the story. Guardrails caught the work that shouldn't ship. Zero escalated means the machine handled the decision path without waking me up. That's the whole point. JOSH: Wait, really? Zero escalated? ERIK: Yeah. That's what the stat says this morning. And I don't say that as magic. I say it because the pipeline is built around boring controls. Tests. review gates. diff scope. confidence thresholds. event history. If the model is unsure, it doesn't get to freestyle into prod. [beat] JOSH: So a better cheap model makes the guardrails better? ERIK: It makes the middle layer better. The first layer is deterministic. Run tests. check lint. inspect files. The model layer explains weirdness, proposes fixes, compares variants, writes notes. Then the deterministic layer comes back and says, nice paragraph, let's see if it passes. [beat] ERIK: That's where builders mess this up. They ask the model to be the whole system. Bad move. The model should be one worker inside a system that knows how to say no. JOSH: That's a good line. ERIK: It's also cheaper than pretending vibes are an architecture. [pause] JOSH: The Perplexity story feels like the ugly twin of that. AI systems citing pages made for AI systems. ERIK: Yep. And it's predictable. Search created SEO. AI search creates answer engine spam. Same incentive, different surface area. [beat] JOSH: The audit said three sites made 215,128 best-software pages. Perplexity's Sonar models cited a ton of low-ranked domains. What's the builder takeaway? ERIK: Your retrieval layer is part of your product. Not an accessory. If your AI system pulls bad sources, it will sound confident while eating garbage. [beat] ERIK: The scary part isn't that spam exists. Spam has always existed. The scary part is AI answers compress the source trail. A user sees a clean answer. They don't feel the rot underneath. JOSH: That's what makes it feel trustworthy. ERIK: Right. Pretty formatting is dangerous. A bullet list with citations looks like work happened. Sometimes it did. Sometimes the model found a page generated last week by someone trying to rank for "best SOC 2 automation tools for left-handed dentists." [beat] JOSH: That's oddly specific. ERIK: Give it a week. [beat] JOSH: How do you handle that with ScanBrief? ERIK: ScanBrief doesn't just grab one source and call it done. Today it scored 99 items across 56 sources. The source mix matters. The scoring matters. Deduping matters. And honestly, rejection matters more than collection. [beat] ERIK: Everybody wants more data. No. You want better refusal. You want the system to say, this looks thin, this source is weird, this claim only appears in one place, don't promote it. JOSH: So source quality becomes an engineering problem. ERIK: It always was. AI just made it visible. If your app summarizes vendor blogs, affiliate farms, and random generated pages, your app is a blender with Wi-Fi. [beat] JOSH: How should a small team defend against this? ERIK: Start with source classes. Primary source beats commentary. Official docs beat listicles. Release notes beat hot takes. Court filings beat someone's thread about court filings. Then add cross-source confirmation before the answer gets treated as fact. [beat] ERIK: Next, store provenance. Not just the final summary. Store where the claim came from, when it was fetched, and what else agreed with it. If the answer changes tomorrow, you need to know whether reality changed or your retrieval got poisoned. JOSH: That sounds heavier than just calling a search API. ERIK: It is. Welcome to building real systems. Calling the API is the easy part. Knowing whether the answer should be trusted is the product. [beat] JOSH: Does this change how people should use Perplexity or AI search personally? ERIK: Yeah. Use it like a scout, not a judge. Let it find leads. Then check the source. Especially for software recommendations, legal stuff, medical stuff, money, security, hiring. Anything where being wrong costs you. [beat] ERIK: And if you're building with AI search, don't hide citations in tiny gray text. Make source quality part of the interface. Let users see why the answer exists. JOSH: That's less magical. ERIK: Good. Magic is what people call systems they don't understand right before they break. [pause] JOSH: Let's talk Google and ad tech. Google avoided a breakup. Why should builders care? ERIK: Because distribution is oxygen. If Google keeps the ad tech stack mostly together, the same machine keeps deciding who gets traffic, who pays for it, who measures it, and who gets priced out. [beat] JOSH: Is that bad? ERIK: It's reality. Bad or good depends where you're standing. If you're a giant advertiser, integrated pipes can be convenient. If you're a small builder buying leads, you better know your unit economics cold. [beat] ERIK: That's why I care more about owned channels than rented attention. ScanBrief goes to inboxes. Prime Automation Solutions has direct outreach. PAS website-pipeline builds actual assets. YouTube is useful, but I don't want my whole business sitting inside someone else's auction. JOSH: So the lesson isn't "Google bad." ERIK: Correct. The lesson is don't build a company where one policy change deletes your week. [beat] JOSH: How does automation fit into that? ERIK: Automation lets you test channels without turning your life into admin work. If I want to test postcards, email, a landing page, a scan report, and a follow-up sequence, I don't want six manual spreadsheets and a prayer. [beat] ERIK: I want events. Lead found. site checked. preview generated. email drafted. human approves. send. reply tracked. outcome logged. That's a business process. Put it on a bus and it becomes something agents can work on. JOSH: That's PrimeBus again. ERIK: Everything becomes PrimeBus eventually. That's not a slogan. It's just what happens when you get sick of duct tape. [beat] JOSH: Where do Google ads fit in that kind of system? ERIK: Paid ads are one input. Useful, but not sacred. If cost per click jumps, the system should know. If a landing page converts badly, the system should know. If organic leads beat paid leads this week, the system should route effort there. [beat] ERIK: Builders need dashboards that tell the truth. PrimeDash shows me what's running in the lab. Live stat this morning says 152 services running on the production server right now. That number matters because I can't manage that by memory. Nobody can. JOSH: That's a lot of services. ERIK: It's a lot of little workers. That's the point. Small tools, clear jobs, telemetry everywhere. When something breaks, it should tell the bus. When something works, it should tell the bus. Silence is the enemy. [beat] JOSH: Does that apply to ad systems too? ERIK: Absolutely. Every campaign should emit events. Spend changed. conversion changed. form broke. lead quality dropped. competitor entered auction. Don't wait until Friday to discover Monday was on fire. [beat] JOSH: And Google staying intact means people still need to play that game. ERIK: Yes. But play it with instruments. Not vibes. Ads are math with a billing account attached. [pause] JOSH: There was also that pre-release of Polars 2.0 in the ScanBrief list. Quick builder note on that? ERIK: Polars matters because data work is usually where automation gets slow and annoying. Fast dataframes mean your local tools can process more without reaching for a warehouse every five minutes. [beat] ERIK: If you're scoring feeds, parsing logs, cleaning CSVs, or doing market scans, performance changes architecture. A faster local tool means a smaller bill and fewer moving parts. I like fewer moving parts until I need more. JOSH: Very network engineer answer. ERIK: Correct answer, then. [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 pre-flight record before it changes anything. [beat] ERIK: Not a long essay. A structured record. Goal. files touched. commands it plans to run. risk level. rollback plan. confidence. Then your system checks that record before execution. [beat] ERIK: For example, if Claude wants to edit Terraform, hit production, or change auth code, that should trip a higher gate automatically. If it's fixing a typo in a test name, fine, let it move. [beat] ERIK: This is how you stop agentic workflows from becoming agentic accidents. The prompt is simple: "Before making changes, output a JSON pre-flight plan with intended files, commands, risks, and stop conditions. Wait for approval unless risk is low." Then wire your runner to enforce it. [beat] ERIK: Don't trust the model to remember the rule. Make the system enforce the rule. 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.