JOSH: It's Friday, May 22. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.

ERIK: Hardware got expensive, robots hate puddles, and old servers still punish lazy engineers.

JOSH: Stick around — Erik's got an AI pro tip at the end about making agents prove their work before they touch production.

[pause]

JOSH: First headline. Consumer electronics may be getting more expensive because memory is getting tight. Erik, why should builders care?

ERIK: Memory is the quiet tax on everything. Phones, laptops, GPUs, edge boxes, NAS builds, all of it. If RAM and flash reprice, cheap local AI gets less cheap.

JOSH: That hits the home lab crowd too?

ERIK: Absolutely. Everyone wants local models until the bill shows up for memory, storage, power, and cooling. The cloud invoice is annoying. The hardware invoice is a commitment ceremony.

[beat]

JOSH: Second headline. Someone asked whether a 48 thousand dollar GPU server was worth it. Six RTX 6000 Ada cards. Apartment power constraints. Real mad scientist energy.

ERIK: That's my favorite kind of post because it's numbers, not vibes. Local inference can win, but only if you keep the machine busy. If it sits idle, congrats, you bought a very expensive space heater.

JOSH: That's direct.

ERIK: True though.

[beat]

JOSH: Third headline. Waymo paused service in Atlanta after robotaxis kept driving into floods.

ERIK: That's the one. Autonomy is easy in a demo and brutal in weather. Water on a road is not a corner case if you run cars in real cities.

[pause]

JOSH: Let's start with memory and GPU costs, because these feel connected. Is local AI getting harder right as everyone wants more of it?

ERIK: Yeah. That's the punchline. Everyone got used to this fantasy where compute keeps getting cheaper, storage keeps getting cheaper, and you can just keep adding context, RAM, swap, GPUs, whatever. Then memory gets tight and the bill punches you in the mouth.

JOSH: What changes first?

ERIK: Consumer devices. Laptops with decent RAM. Mini PCs. Workstations. Network appliances with more flash. Anything with high-bandwidth memory around GPUs. You don't need every part to spike. One constrained component can move the whole build.

JOSH: So builders should wait?

ERIK: No. Builders should buy with intent. Big difference. Don't buy hardware because some thread said local AI is the future. Buy hardware because you have a workload you can name.

[beat]

JOSH: Give me an example.

ERIK: ScanBrief scored 36 items across 54 sources today. That's a workload. Fetch, dedupe, rank, summarize, publish. I can decide what belongs local, what belongs in Claude, what belongs in GPT, and what belongs in boring scripts. That's architecture.

JOSH: Not just "throw a model at it."

ERIK: Exactly. If the job is scraping sources, filtering duplicates, and scoring relevance, I don't need a giant local model for every step. Some of that is Python. Some of that is a database query. Some of that is a model call after the cheap filters already did their job.

JOSH: That's where people waste money?

ERIK: Constantly. They use the most expensive brain on the dumbest part of the task. It's like hiring a principal engineer to rename files. Funny once. Bad system.

[beat]

JOSH: How does that tie back to the 48 thousand dollar GPU server?

ERIK: That build makes sense if the machine has a real queue. Six RTX 6000 Ada cards can move a lot of inference. But the question isn't "is it fast?" Of course it's fast. The question is utilization.

JOSH: How busy is it?

ERIK: Right. If it's running research jobs all day, batch inference, embeddings, evaluation, video indexing, agent tests, then maybe it pays. If it's used two hours a week because cloud pricing feels emotionally offensive, that's not engineering. That's retail therapy with PCIe slots.

JOSH: Wait, really? Even with cloud GPU prices?

ERIK: Yes. People underestimate the boring stuff. Power circuits. Heat. Noise. Driver issues. Failed disks. Backups. Networking. Remote access. Monitoring. Physical security, which in an apartment usually means "please don't trip over the machine."

JOSH: Very official.

ERIK: Certified enterprise carpet deployment.

[pause]

JOSH: You run your own systems though. Where do you draw the line?

ERIK: I run the lab because it teaches me things the cloud hides. Neo and Morpheus make me deal with real limits. CPU, memory, disk, NATS, Docker, systemd, cron, bad deploys, noisy logs. That's valuable.

JOSH: But not everything goes local.

ERIK: Correct. My Bobaverse fleet has 5 agents right now: Neo, Homer, Bill, Echo, and Gandalf. Claude plus GPT. PrimeBus ties them together. The point is not "local good, cloud bad." The point is putting the right job in the right place.

JOSH: What does PrimeBus do in that setup?

ERIK: PrimeBus is the nervous system. It processed 1949 automation events across 8 projects today. Events come in, agents subscribe, jobs fire, reviews happen, guardrails block bad changes. That's not a chatbot. That's infrastructure.

JOSH: And Gandalf is the reviewer?

ERIK: Gandalf is the reviewer and the bouncer. Overnight, 8 code changes were automatically reviewed by Gandalf and merged to production. But the more important number is the blocking. Since April 17, the auto-merger has run 1864 attempts: 730 merged, 1122 blocked by Gandalf, 12 escalated to me.

[beat]

JOSH: The blocked count is higher than the merged count.

ERIK: Good. That's the system working. People hear "blocked" and think failure. Nope. Blocked means the guardrail caught something before production had to teach us the lesson.

JOSH: That's a different way to look at automation metrics.

ERIK: It has to be. If your AI agent merges everything, you don't have automation. You have a liability with an API key.

[pause]

JOSH: Let's move to Waymo. Robotaxis driving into floods sounds like the kind of story people share because it feels absurd.

ERIK: It's absurd, but it's also the real autonomy problem. Roads are not static. Rain changes the surface. Reflections change perception. Water hides lane markings. Depth gets weird. Humans are not perfect, but humans see a flooded street and usually think, "Nope."

JOSH: Usually.

ERIK: Usually. There's always one guy in a lifted truck trying to become a local news segment.

[beat]

JOSH: What's the engineering lesson?

ERIK: Autonomy needs refusal. That's the lesson. The system has to know when not to act. Not just "can I complete the route?" but "should I stop, reroute, call for help, or hand this off?"

JOSH: That sounds a lot like AI agents.

ERIK: Same pattern. Agents fail when builders only reward completion. "Finish the task. Merge the code. Send the email. Close the ticket." That's how you get confident garbage.

JOSH: So the missing feature is hesitation?

ERIK: Not hesitation. Judgment. Hesitation sounds random. Judgment is policy plus telemetry plus escalation. HumanRail exists because some work should route to a human when the model is not confident or the risk is too high.

JOSH: Give me a real version of that.

ERIK: Website change request comes in through email-reactor. If it's "change this button text," the system can handle it. If it's "replace the pricing page and update all legal language," that's not a blind commit. It needs review. Maybe client approval. Maybe a diff. Maybe a human.

JOSH: Because the cost of being wrong changed.

ERIK: Exactly. Same with a robotaxi. Driving one more block is cheap until the block is underwater.

[pause]

JOSH: How should builders design for that?

ERIK: Start with failure modes. Not prompts. Not model choice. Failure modes. Ask what a bad action looks like. Then put a circuit breaker in front of it.

JOSH: What's a circuit breaker in an AI workflow?

ERIK: Could be a confidence score. Could be a test suite. Could be a second model review. Could be a policy check. Could be requiring a human approval when files touch billing, auth, legal, Terraform state, Kubernetes manifests, NSO service templates, anything that can hurt you.

JOSH: That's a long list.

ERIK: Production is a long list.

[beat]

JOSH: And this is where people get agentic workflows wrong?

ERIK: They give the agent hands before they give it rules. I don't care how good the model is. Claude is great. GPT is great. They're still tools inside a system. The system is what makes the work reliable.

JOSH: So you trust the workflow more than the model.

ERIK: Every time. Models change. Prompts drift. APIs have bad days. A workflow with logs, tests, reviews, rollbacks, and escalation survives that. A magic prompt does not.

[pause]

JOSH: Third deep dive. The Ubuntu 16.04 blog migration to FreeBSD. This sounds less flashy than robotaxis, but you had a face when you saw it.

ERIK: Because that one is the real world. Someone ran a blog on Ubuntu 16.04 for 10 years, then moved it to FreeBSD. That's not a headline about the future. That's a headline about the mess sitting under half the internet.

JOSH: Old servers everywhere?

ERIK: Everywhere. Old LTS releases. Forgotten droplets. One-off cron jobs. Mystery nginx configs. Python apps with dependencies nobody wants to touch. A box named "web-old-final-2" that secretly runs revenue.

JOSH: That name hurt.

ERIK: Because it's real.

[beat]

JOSH: Why does this matter for AI builders?

ERIK: Because AI doesn't remove maintenance. It makes more of it. Every agent you create has dependencies, keys, queues, logs, retries, state, prompts, model versions, permissions. If you don't have maintenance discipline, your AI system becomes another old server with a nicer README.

JOSH: That's not the pitch deck version.

ERIK: Good. Pitch decks don't carry pagers.

[pause]

JOSH: The same ScanBrief batch also had Python 3.15 features and complaints about uv's package management UX. Is that part of the same story?

ERIK: Very much. Tools get faster, but operations still matter. uv is fast. Great. But if dependency maintenance is awkward, teams still avoid doing it. Python 3.15 can add nice features, but you still have to migrate, test, deploy, and watch for breakage.

JOSH: So speed doesn't replace hygiene.

ERIK: Correct. Fast install is not the same as healthy software. I want dependency visibility. I want outdated package reports. I want reproducible builds. I want tests that fail loudly. Then I want an agent that can propose the update and a reviewer that can reject it.

JOSH: Gandalf again.

ERIK: Gandalf again. Boring name would have been "policy enforcement layer." I chose better.

[beat]

JOSH: Where does FreeBSD fit?

ERIK: FreeBSD Jails are a reminder that boring isolation still works. Containers did not invent the idea of separating workloads. Good systems thinking has been around. Jails, chroot, VMs, VLANs, ACLs, runbooks, backups. AI doesn't make those obsolete. It gives you more reasons to use them.

JOSH: That's probably not what people want to hear.

ERIK: They need to hear it. If you can't patch a server, you shouldn't let an agent patch production unsupervised. Get the basics right first.

[pause]

JOSH: What does "basics right" mean in your world?

ERIK: Logs first. Every automation emits an event. Every event has a project, status, timestamp, and useful context. Then tests. Then rollback. Then review. Then metrics.

JOSH: That's the PrimeBus model.

ERIK: Yeah. PrimeBus didn't start as some giant master plan. It started because programs needed to talk. Then agents needed work. Then failures needed routing. Then merges needed review. Build one small loop, prove it, add guardrails, then put more weight on it.

JOSH: And now 81 distinct projects have emitted telemetry to PrimeBus.

ERIK: Right. That number matters because it shows coverage. If a project doesn't emit telemetry, it doesn't exist to the automation layer. It's just a thing you hope is fine. Hope is not monitoring.

[beat]

JOSH: That's the line.

ERIK: It should be printed on a rack door.

[pause]

JOSH: Pulling these together, today's theme feels like limits. Hardware limits, weather limits, maintenance limits.

ERIK: That's exactly it. Builders like capability. More tokens. More GPUs. More agents. More sources. But limits are where engineering starts. The memory shortage says cost matters. The GPU server says utilization matters. Waymo says refusal matters. The old Ubuntu box says maintenance matters.

JOSH: And the answer isn't "use less AI."

ERIK: No. Use better systems. Put AI where it wins. Let scripts do script work. Let queues carry work. Let tests catch bad output. Let humans approve risky actions. Then keep shipping.

JOSH: That's the show in two words.

ERIK: Build systems.

[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 proof bundle before it gets write access.

JOSH: What's in the bundle?

ERIK: The exact files it plans to change, the reason for each change, the tests it will run, the rollback path, and the risk level. Make it write that first. Then let another model or a simple policy script review the plan.

JOSH: Before the code changes?

ERIK: Before the code changes. For small edits, this takes seconds. For risky edits, it catches nonsense early. If the agent can't explain what it will touch, it doesn't get to touch it.

JOSH: That's simple.

ERIK: Simple is the point. You can do this today in Claude Code, Codex, GitHub Actions, whatever. Tell the agent: "Create a plan with changed files, commands, rollback, and risk. Stop after the plan." Then run review. Then allow the patch. That's your tip. Use it.

[pause]

JOSH: Binge all five episodes this weekend plus our YouTube shorts — links at buildorbereplaced.dev.

[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.

JOSH: It's Thursday, May 21. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.

ERIK: The developer toolchain is now an attack surface with a nice icon and five-star reviews.

JOSH: Stick around — Erik's got an AI pro tip at the end about making agents prove their work before they touch production.

[pause]

JOSH: Headlines. First up, GitHub confirmed a breach tied to a malicious VS Code extension. Erik, that's every developer's favorite nightmare, right?

[beat]

ERIK: Yep. Trusted tool, trusted editor, bad payload. GitHub says 3,800 repositories were exposed, and the lesson is simple: your IDE is part of your production environment now.

JOSH: Second, Anthropic is reportedly expanding into Colossus2 and using GB200 chips. What's the signal there?

ERIK: Compute is the product. Everyone talks about model intelligence, but the companies with chips, power, cooling, and scheduling win the next round.

JOSH: Third, an OpenAI model reportedly disproved a central conjecture in discrete geometry.

ERIK: That's not a chatbot writing a poem. That's machine reasoning pushing into math. If that holds up, the AI conversation gets a lot less cute and a lot more serious.

[pause]

JOSH: Let's start with the GitHub breach. A malicious VS Code extension gets installed, thousands of repos get exposed. How bad is that?

ERIK: Bad in the way that matters. Not dramatic movie bad. Boring infrastructure bad. Somebody installs an extension because it looks useful. Maybe it formats code. Maybe it helps with Git. Maybe it has a logo that looks like twelve other tools. Then it gets access to tokens, workspace files, terminal sessions, environment variables, maybe secrets in plain text because people still do that. Now your private repos aren't private anymore.

[beat]

ERIK: That's the part people miss. A repo breach isn't just source code. It's architecture. It's comments. It's test fixtures. It's CI files. It's deployment patterns. It's old credentials somebody forgot in a sample config. It's the map of the building.

JOSH: Wait, really? Source code alone is that useful?

ERIK: Source code is the least interesting part sometimes. If I can see your Terraform, your Kubernetes manifests, your GitHub Actions, your service names, your S3 bucket naming, your internal package names, I'm building a plan. I know where to poke. I know what assumptions your team makes. And if your developers have extensions running with broad access, that's a really nice doorway.

JOSH: How do you defend against that without making developers miserable?

ERIK: You treat extensions like dependencies. Approved list. Version pinning where you can. Permission review. Separate workspaces for sensitive repos. No random extension gets to walk into the bank vault because it has dark mode.

[beat]

ERIK: And you watch behavior. PrimeBus processed 2231 automation events across 10 projects today. That matters because events tell you when something changes. New process. New outbound call. New repo clone. New secret access. New extension install. If you don't have telemetry, you're asking humans to remember what normal feels like. Humans are bad at that. We have meetings.

JOSH: So this isn't just a GitHub problem.

ERIK: Nope. GitHub is the headline because it has the brand name. The real issue is developer supply chain. npm packages, browser extensions, VS Code plugins, GitHub Actions, Docker images, AI coding tools, MCP servers, all of it. Developers are now running little automation agents all day. Some are useful. Some are a tire fire with a README.

JOSH: Where do AI coding tools fit into that?

ERIK: Same bucket. If Claude, Cursor, Copilot, Windsurf, or any agent can read your repo and run commands, it needs boundaries. Not vibes. Boundaries. Read access is one thing. Write access is another. Shell access is another. Production deploy access is a completely different animal.

[beat]

ERIK: In my setup, Gandalf reviewed and merged 8 code changes to production overnight. That's not because I trust automation blindly. It's because the pipeline has checks, tests, diff review, and blocking rules. The PrimeBus auto-merger has run 1800 attempts since 2026-04-17: 714 merged, 1074 blocked by Gandalf, 12 escalated to Erik. That blocked number is the story. Guardrails did their job.

JOSH: That's the opposite of just letting an agent rip through a repo.

ERIK: Exactly. Agents should be powerful, but they should be treated like junior engineers with root access and unlimited coffee. Useful, fast, occasionally terrifying. You make them show diffs. You make them run tests. You make them explain risk. You don't give them a production token because they asked nicely.

[pause]

JOSH: The second deep story is Anthropic expanding to Colossus2 and using GB200. Everyone hears chip names and data center names and tunes out. Why should builders care?

ERIK: Because model access is going to feel like bandwidth. Some days fast. Some days expensive. Some days rate-limited. Some days the model you want is not the model you get. The companies building massive AI infrastructure are trying to make that less painful for themselves first, and then maybe for us.

JOSH: GB200 is Nvidia's big AI platform, right?

ERIK: Right. Grace Blackwell. CPU and GPU tied together for AI workloads. High memory bandwidth, fast interconnect, built for giant model training and inference. The details matter to hardware people. For builders, the useful translation is this: bigger clusters mean bigger models, longer context, faster inference, and more competition around price.

JOSH: Does that mean everything just gets cheaper?

ERIK: Not automatically. Compute never gets cheap in a straight line. Demand eats every improvement. Faster chips show up, then model labs train bigger models. More efficient inference shows up, then people run agents all day instead of one prompt. You give developers more tokens, they fill the bucket by lunch.

[beat]

ERIK: PrimeRouter exists because I don't want one provider deciding whether my systems work today. It routes LLM calls by priority tier, tracks fleet telemetry, and fails over across providers. That's not fancy. That's plumbing. But plumbing is what keeps the house from becoming a swamp.

JOSH: So your take is, don't bet your automation on one model?

ERIK: Correct. Don't bet the business on one model, one endpoint, one context window, one vendor account, or one pricing page. Use Claude where it wins. Use GPT where it wins. Use smaller models when the task is boring. Use local or cheap models for classification if the quality is good enough. Builders need routing, evals, and budgets. Not brand loyalty.

JOSH: How does token speed fit into that? ScanBrief had a story today about how fast tokens per second actually feels.

ERIK: That story is underrated. People say 500 tokens per second like it's a clean number. But tokens aren't words. Code is dense. JSON is dense. Prose feels different. A model generating 500 tokens per second on a short summary feels instant. The same rate on a code patch with tests and logs can still feel like you're watching paint apply for a permit.

JOSH: That's painfully specific.

ERIK: Because it's real. In agent systems, latency is not just one model response. It's plan, read files, call tools, edit, run tests, parse errors, retry, summarize, wait on CI. If each step is small, token speed matters less than orchestration. If each step is huge, token speed matters a lot.

[beat]

ERIK: That's why I care about event buses. NATS, PrimeBus, queues, workers, telemetry. You don't want one giant agent doing everything in one giant context. You want small jobs with receipts. ScanBrief scored 51 items across 54 sources today. That works because the pipeline breaks the job apart: fetch, dedupe, score, summarize, rank. Each piece is measurable. If one piece gets weird, I can see it.

JOSH: That's the pattern listeners should steal?

ERIK: Yes. Break the work into events. Give every event an ID. Log inputs and outputs. Make retries visible. Keep human escalation as a first-class path. That's how you build automation that survives bad days.

JOSH: And the AI labs are racing to give us more compute for that.

ERIK: They are. But compute doesn't fix sloppy architecture. Bigger models make bad systems fail with better grammar.

[pause]

JOSH: Third deep story. An OpenAI model disproves a central conjecture in discrete geometry. This one sounds academic, but you said it changes the AI conversation.

ERIK: It does if the result holds up. A model helping disprove a conjecture means we're not just using AI for summarizing emails and writing unit tests. We're using it to search a hard problem space and find something humans missed.

JOSH: Is that reasoning, or is it just brute force?

ERIK: Sometimes brute force is reasoning with better shoes. Math discovery has always used computation. The interesting part is whether the model can guide the search, form useful structures, reject dead ends, and produce something mathematicians can verify. It doesn't need to be mystical. It needs to be productive.

JOSH: What does that mean for engineers?

ERIK: It means the ceiling is moving. Today it's geometry. Tomorrow it's protocol design, compiler passes, network failure models, test generation, formal verification, routing policy analysis. The stuff that used to require a very expensive human staring at a whiteboard for two weeks can start with an AI generating candidate proofs, counterexamples, or simulations.

[beat]

ERIK: In networking, this matters. I've spent years around Cisco NSO, service models, rollback behavior, config drift, device templates, all the fun stuff that makes your eyes do pushups. AI doesn't have to replace the senior engineer in one shot. It can generate the test cases the senior engineer would've written after three cups of coffee and mild regret.

JOSH: That's where the "build or be replaced" part gets uncomfortable.

ERIK: It should be uncomfortable. If your job is only to type what someone else already decided, you're in trouble. If your job is to define the problem, design the system, set the guardrails, verify the output, and ship, AI makes you harder to replace.

JOSH: So the builder's advantage is judgment?

ERIK: Judgment plus systems. People say "learn AI" and then they mean prompts. Prompts are fine. But the real advantage is turning AI into a worker inside a controlled system. Inputs, tools, permissions, tests, rollback, alerts. The model is not the product. The system around it is the product.

JOSH: That sounds like your home lab talking.

ERIK: It is. The Bobaverse fleet has 5 agents: Neo, Homer, Bill, Echo, Gandalf. Claude plus GPT. That's not because I enjoy naming machines, although I clearly do. It's because different agents do different jobs. One reviews. One builds. One watches telemetry. One handles routing. Gandalf blocks bad merges. Very on brand.

[beat]

JOSH: What happens when the model is wrong?

ERIK: Same thing that happens when a human is wrong. The system catches it or you have a bad day. That's why I don't care about demos where an agent writes a todo app. Cool. Can it handle secrets? Can it roll back? Can it explain why it touched a file? Can it stop when confidence is low? Can it ask a human instead of pretending?

JOSH: That's the HumanRail idea, right? Human when the model isn't confident?

ERIK: Exactly. Humans shouldn't be in every loop. That's how you build a slow robot. Humans should be in the right loop. Weird permission change. Failing test with ambiguous cause. Security-sensitive diff. Production database migration. The boring stuff can pass. The sharp stuff gets a checkpoint.

JOSH: Does the math story make you more aggressive about using AI?

ERIK: More serious, not reckless. If models can contribute to math, they can absolutely contribute to infrastructure. But infrastructure has blast radius. A wrong proof annoys mathematicians. A wrong automation deletes a database. PrimeRestorer exists because I like sleeping. Backups, system state, S3 offsite. Not glamorous. Very useful when somebody's clever agent gets clever in the wrong directory.

[pause]

JOSH: Quick thread tying these together. Malicious extensions, giant AI compute, models doing math. What's the builder's take?

ERIK: The builder's take is simple. Tools are getting stronger on both sides. Attackers get better automation. Builders get better automation. Model labs get more compute. Your editor gets more powerful. Your agents get more access. The gap is governance. Not corporate theater governance. Real engineering governance. What can run? What can read? What can write? What gets logged? What gets blocked? Who gets paged?

[beat]

ERIK: If you're a solo builder, do the small version. Separate tokens. Use least privilege. Keep secrets out of repos. Review extensions. Run agents in a sandbox. Require tests before merge. Log every automation action. If you're a company, do the same thing with fewer excuses and more money.

JOSH: And if someone says that's too much overhead?

ERIK: Then they haven't been breached yet.

[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 agents produce evidence, not just output. Every agent task should end with three things: what changed, how it verified the change, and what it refused to do. If it edited code, make it show the files and test results. If it read logs, make it cite the exact error. If it skipped something, make it say why.

[beat]

ERIK: This is how you stop fake confidence from entering your system. Add a final instruction: "Before finishing, list the commands you ran, the checks that passed, the checks that failed, and any risky assumptions." Then wire that into your review flow. Don't accept "done." Done is a claim. Evidence is the receipt.

[beat]

ERIK: 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.

JOSH: It's Wednesday, May 20. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.

ERIK: Guardrails are having a moment. Google just put them on search. An 8B model went from 53% to 99% because of them. My pipeline shipped 18 code changes to production overnight because of them. That's the theme today.

JOSH: Stick around — Erik's got an AI pro tip at the end about routing workloads to the right model and actually cutting your costs.

[pause]

JOSH: Okay, first headline — Google is calling this the biggest search overhaul in 25 years. Gemini 3.5 Flash is now the default AI model for search. What's the short version?

ERIK: The search box is an agent now. Not smarter autocomplete. An agent. You can give it multi-step tasks and it executes them. That's a real architectural shift, not a UI refresh.

[beat]

JOSH: Next — GitHub is investigating unauthorized access to their internal repositories.

ERIK: Internal repos at GitHub means potentially source code, pipeline configs, infrastructure logic. Details are thin but scope could be significant. Worth watching.

[beat]

JOSH: And OpenAI is now using Google's SynthID watermarking system for AI-generated images.

ERIK: Two competitors sharing infrastructure to solve the same problem. When that happens, the problem is real and neither of them wants to be the one left without a solution.

[pause]

JOSH: Let's go deep on Google first. Twenty-five years is a long time to not change something.

ERIK: It is. And the reason search didn't change is it didn't have to. Type keywords, get a list of links, click through. That loop worked. The problem is that loop was built for a world where information was hard to find. Now information is everywhere and the bottleneck is synthesis — turning ten sources into one answer. That's what Gemini 3.5 Flash does inside search now. It doesn't just retrieve. It reasons.

JOSH: What does that actually look like for someone using it?

ERIK: You can ask it something like "find me an open-source Python library for PDF extraction that handles scanned pages and has active maintenance" and it comes back with a curated answer, not ten blue links you have to go read yourself. The complexity of the query used to determine how long it took you to get an answer. Now the model absorbs that complexity. Your time to answer drops.

JOSH: Is Gemini 3.5 Flash actually good enough for that?

ERIK: That's the smart question. They're saying 4x faster than competing models at frontier-level intelligence. Speed matters at Google's scale — they're serving billions of queries. You need fast and good. Haiku-tier speed with Sonnet-tier reasoning is roughly the target. If they hit that, it works.

[beat]

JOSH: ScanBrief flagged this one high today?

ERIK: Pulled 51 items across 54 sources this morning. This one scored near the top. And it's episode 26 for us — the model story has been showing up almost every week since we started. The pace isn't slowing down.

JOSH: How does this affect what you're building?

ERIK: It doesn't threaten ScanBrief. Google search and ScanBrief are solving different problems. Google brings you what's indexed and popular. ScanBrief scores by relevance to what I actually care about — automation, infrastructure, AI tooling. It's curation, not retrieval. What does affect me is that Gemini 3.5 Flash is now in the mix as a routing option. PrimeRouter already handles multi-provider failover. I'll evaluate it on speed and token cost against what I'm running now.

[beat]

JOSH: You already have that infrastructure built to slot it in.

ERIK: PrimeBus processed 3,891 automation events across 12 projects today. Any model with a better cost-to-speed ratio gets tested against that load. That's just how the fleet runs.

[pause]

JOSH: Okay, the guardrails story. Forge. An 8-billion parameter model at 53% on agentic tasks — that's not great. Then guardrails push it to 99%. How?

ERIK: The number is the whole story. 53% on agentic tasks means the model is failing roughly half the time on anything multi-step. That's not a capability problem — that's a reliability problem. Guardrails don't make the model smarter. They intercept the outputs before they execute and check them against known failure patterns. Wrong tool call, hallucinated parameter, a loop that won't terminate — these are predictable failure modes. You pattern-match against them and block the bad output before it does damage.

JOSH: So the model doesn't change at all?

ERIK: Same weights, same training. You're adding a validation layer on top. The model still generates the bad output — you just never let it leave the system. That's why the jump is so dramatic. You go from "model fails half the time" to "model's failures are caught before they matter."

JOSH: Is that how Gandalf works in your pipeline?

ERIK: Exactly how. Since April 17, the auto-merger has run 1,641 attempts. 621 made it to production. 1,008 got blocked by Gandalf.

[beat]

JOSH: So more than half get blocked. That sounds like a lot.

ERIK: It's supposed to. Every blocked PR is a guardrail working. Gandalf is checking for a specific set of failure patterns — the ones that burned us before. Asyncio task garbage collection, SQL-level race conditions, mixed import signatures, fire-and-forget calls that eat exceptions. Those aren't random failures. They're predictable. So they're blockable.

JOSH: What about the 12 that got escalated to you?

ERIK: Those are the ones where the fix was technically correct but the approach was wrong. You can't guardrail judgment calls. If a service is trying to solve a concurrency problem and the auto-fix works locally but creates a worse race condition at scale — Gandalf can't see that. That comes to me. Everything else runs without me. Eighteen code changes shipped to prod overnight.

JOSH: You wake up in the morning and there are already eighteen changes in production.

ERIK: And I didn't touch a terminal. That's 81 distinct projects emitting telemetry to PrimeBus right now. The fleet is running. My job is the judgment calls, not the execution.

[beat]

JOSH: The Forge paper basically proves what you built empirically.

ERIK: The academic version, yeah. They ran the benchmarks. I ran it on 95 live services. Same conclusion: guardrails are the delta between a model that's almost useful and one you can trust with production code. The 8B model story matters because it means you don't need frontier-scale compute to build something reliable. You need the right architecture around whatever model you're running.

JOSH: That's a big deal for cost.

ERIK: Massive. If a guardrailed 8B model performs at 99% on your specific task, you have no reason to pay for a 70B. The conversation in AI right now is about model capability. The real variable is guardrail quality. Nobody's writing papers about that fast enough.

[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 defaulting everything to the most capable model you have access to. Every task has a cost-to-quality curve and most builders are ignoring it. Simple classification tasks — does this text match this category, yes or no — don't need Opus. They need the fastest model that gets it right 95% of the time. Save the expensive calls for the hard problems: architecture decisions, ambiguous error logs, anything where being wrong costs more than the API call itself. I route explicitly in PrimeRouter. Routine summarization goes to the fastest tier. Complex multi-step analysis escalates up. When I stopped defaulting everything to the top model, API costs dropped by more than half. Latency dropped too. Know what your workload actually requires, then match your model to that requirement. Not to your anxiety about whether the cheaper model is good enough. Test it. Most of the time it is. 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.

[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.

JOSH: It's Tuesday, May 19. Episode 25 of Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.

[beat]

ERIK: Anthropic bought the company that makes it easy to hand AI agents a wrench. That's what today's about.

JOSH: Stick around — Erik's got an AI pro tip at the end about which model to use for which task, and how to stop paying premium prices for work that doesn't need it.

[pause]

JOSH: Let's run headlines. Elon Musk's lawsuit against OpenAI is dead. Jury ruled against him.

ERIK: Claims were time-barred. California statute of limitations killed it before it got legs. The charity-to-for-profit argument might have had merit in a different timeline. He waited too long.

JOSH: Does it matter for the industry?

ERIK: Not really. OpenAI keeps building. The lawsuit was a distraction, not a threat.

[beat]

JOSH: Next — 314 npm packages got compromised in what looks like a coordinated supply chain attack.

ERIK: Typosquatting. Someone registered package names one keystroke off from things you'd actually install. People auto-install, malware runs. This is not a new attack. Pin your dependencies, audit your lockfile, use a private registry if you're in prod.

[beat]

JOSH: And Cursor shipped Composer 2.5.

ERIK: Every time Cursor ships, someone who wasn't using AI coding tools yesterday starts today. The gap between people using these tools and people waiting is widening every week. That's the only headline that matters there.

[pause]

JOSH: Alright, let's go deep. Anthropic acquired a company called Stainless this week. Before this morning I had never heard of them. What do they do?

ERIK: Stainless takes an API spec — an OpenAPI document — and generates fully typed client libraries in multiple languages. Python, TypeScript, Go, Ruby, Java. They also build MCP servers, which are connectors that expose any API as a tool call that Claude can use natively.

JOSH: And Anthropic just bought them outright.

ERIK: Bought them. Which means this is not a partnership. Anthropic is saying we are bringing this capability inside the tent permanently.

[beat]

JOSH: Why? Claude can already use tools.

ERIK: Claude can use tools that someone already built. The hard part is building the connector in the first place. Right now if you want an agent to call your company's internal API, you write a wrapper, define the schema, document the edge cases, test it. That's hours of work per API. Stainless compresses that to near zero. You give it the spec, it hands you the connector.

JOSH: So it's lowering the cost of plugging Claude into real systems.

ERIK: That's the whole game. The intelligence race is mostly done — the models are good. The next battlefield is integration breadth. Whoever has the most tools Claude can reach wins the enterprise layer. Stainless is a direct bet on that.

[beat]

JOSH: Does this change anything you're building?

ERIK: I have five agents on the fleet right now — Neo, Homer, Bill, Echo, and Gandalf. They talk over PrimeBus using NATS messaging and structured events. What I've hand-rolled every time is the connector layer — when an agent needs to reach an external system, I've built that bridge myself. If Stainless tooling ships as part of the Claude SDK, that changes what's easy to ship in a weekend.

JOSH: What does Gandalf do in that crew?

ERIK: Code reviewer. Every PR the auto-merger generates hits Gandalf before it lands. He approves, blocks, or escalates to me.

JOSH: How's the pipeline running right now?

ERIK: Since April 17 — 1,620 merge attempts. 603 merged. 1,005 blocked by Gandalf. 12 landed on my desk.

JOSH: Wait. Only 12 escalated to you personally?

ERIK: Twelve. That's the whole point. Everything else the agents handle. The blocks are not failures — they're the guardrails doing exactly what they're supposed to do. Gandalf catches commit scope problems, missing test coverage, pattern violations. In the last seven days, 237 blocked, 121 merged, zero escalated to me.

[beat]

JOSH: Zero escalated to you in a week.

ERIK: Zero. I didn't touch the merge queue at all last week. It just ran.

JOSH: That's the goal — you as the exception handler, not the default path.

ERIK: That's always the goal. If I'm the default path, I've built a job, not a system.

[pause]

JOSH: Second deep dive. A team let AI agents autonomously run four radio stations. No humans. Each agent started with twenty dollars and had to become profitable. Music, ads, scheduling, listener engagement — all of it. What's your reaction?

ERIK: First reaction — that's not a novelty project. That's a product blueprint. Each station is a goal-bounded agent with a starting resource and a target outcome. It either figures out how to generate revenue or it doesn't. That's a real constraint most AI demos don't have.

JOSH: Did any of them make money?

ERIK: Results weren't clear in what I read. But the architecture is more interesting than the P&L. This is what I'd call budget-aware orchestration. The agent has to reason about cost versus outcome — do I negotiate with this advertiser or move on? That's a fundamentally harder decision than do I play this song next.

[beat]

JOSH: Why is that harder?

ERIK: Because it requires the model to weight tradeoffs over time, not just in the current context. Most agents today are context-local. They make a good decision for this moment. Budget-aware agents have to make a good decision for this moment given everything they've already spent and everything they still need to produce.

JOSH: How close are you to running something like this?

ERIK: The infrastructure is already there. PrimeBus processed 2,348 automation events today across eight projects. Ninety-three services running on prod right now. The missing piece isn't compute or connectivity — it's what I just described. Agents that track what they've spent, what they've produced, and stop themselves when the math stops working.

JOSH: That's the guardrail problem again.

ERIK: It's always the guardrail problem. The capability exists. What's missing is governance. Agents that self-limit when something goes sideways instead of continuing until a human notices three hours later.

JOSH: How do you handle that in your stack right now?

ERIK: Event contracts on PrimeBus. Every agent publishes structured events with expected schemas. If an agent publishes something anomalous — wrong event type, unexpected values, silence when there should be activity — another agent sees it and flags it. ScanBrief this morning scored 48 items across 54 sources. There's a scoring agent, a dedup agent, a delivery agent in that pipeline. If scoring goes flat, the delivery agent knows before I check my phone.

[beat]

JOSH: So the answer to AI running unsupervised is more AI supervising it.

ERIK: That's exactly it. You don't solve the autonomous agent problem by adding human checkpoints everywhere — you solve it by making agents watch each other. Humans in the loop for judgment calls. Agents in the loop for everything else.

JOSH: What would you build with this architecture if you had a week?

ERIK: Autonomous content pipeline. An agent monitors a topic space, identifies gaps, drafts content, routes through review, publishes, tracks performance, and feeds the signal back into the monitoring agent. No human touches it unless something breaks pattern. PrimeRouter already handles the model routing and failover. PrimeBus has the event mesh. The part still needing wire is closing the feedback loop — making the performance signal actually change the next editorial decision.

[beat]

JOSH: You're describing the radio station experiment, but for content.

ERIK: I'm describing the radio station experiment with four years of infrastructure behind it instead of a twenty dollar starting budget.

[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: Model selection. Most people run everything through their smartest model and wonder why the bill is high and the pipeline is slow. The rule I follow: Sonnet for execution work — implementation, refactors, tests, standard changes. Opus for architecture decisions, retry logic, ambiguous debugging, or anything where two attempts already failed. PrimeRouter does this automatically — Sonnet is the default route, and escalation to Opus fires when the task pattern matches something that warrants it. The result is you get Opus judgment exactly when it counts and you stop paying Opus prices for work that doesn't need it. One change to your routing logic can cut your AI costs in half without touching your output quality. That's your tip. Use it.

[pause]

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.

JOSH: It's Thursday, May 14. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.

ERIK: AI isn't a tool anymore. It's staff. If you don't wire it into real work, somebody else will.

[beat]

JOSH: Stick around — Erik's got an AI pro tip at the end about building a model router that saves money without slowing down.

[pause]

JOSH: Three quick ones first. Anthropic rolled out Claude for Small Business. Is that actually a big shift, or just a nicer landing page?

ERIK: Big shift. When the model ships with connectors and ready-made workflows, it stops being a demo and starts acting like a junior operator inside finance, sales, and support.

[beat]

JOSH: YellowKey hit BitLocker with a zero-day and the proof of concept is already public. How bad is that?

ERIK: Bad enough that every admin should care today. Encryption is great until the boot chain hands the keys over too easily.

[beat]

JOSH: Cisco posted record quarterly revenue, then said it's cutting fewer than four thousand jobs to push harder on AI. What's the read?

ERIK: Same story across tech. Revenue up does not protect repetitive work. Budget is moving toward AI infra, security, and the people who can actually build with it.

[pause]

JOSH: Start with Anthropic. Why does this one matter more than the usual product launch noise?

ERIK: Because it changes where the battle is. For a while, AI was a tab. You opened Claude or ChatGPT, typed in a box, got some text, maybe pasted it into another system. That's useful, but it's not operational. Claude for Small Business moves it into QuickBooks, HubSpot, Google Workspace, Microsoft 365, PayPal, DocuSign. That's not "help me think." That's "help me do work where the records already live."

[beat]

JOSH: What's the difference for a normal business owner?

ERIK: Blank box versus installed worker. That's the whole thing. Most owners don't wake up wanting to prompt engineer. They want the lead email answered, the CRM updated, the invoice checked, the calendar sorted, the overdue stuff surfaced. Early AI products dumped people into a text field and acted surprised when adoption stalled. A bakery owner or a small MSP doesn't want a research project. They want a lane.

JOSH: So the lane is the product.

ERIK: Exactly. Not the model. The lane. Prebuilt workflow, limited permissions, clear finish line. That's what sells.

[beat]

JOSH: People have been pitching integrations forever, though.

ERIK: Yeah, and most of them were duct tape with nice branding. What's different now is the packaging got opinionated. Anthropic isn't saying "connect whatever you want and hope for the best." They're saying, here's finance, here's customer service, here's sales, here's ops. That's important because trust matters more than IQ once AI can touch money and records.

JOSH: Less freestyle.

ERIK: Way less. Freestyle is fun in a demo. Freestyle in accounting is how you buy the same software license twelve times.

[beat]

JOSH: Where does this land for people building custom systems?

ERIK: It squeezes the shallow wrappers. Hard. If your startup is "we connected a model to Gmail and HubSpot," the platform vendors are already eating that category. You survive by owning the ugly edges. Approvals. weird data. audit trails. domain rules. all the stuff demos skip because it isn't sexy.

JOSH: That's where your stuff lives.

ERIK: Exactly. PrimeRouter exists because one model is not the answer to every task. A cheap extraction job doesn't need the same brain as a change plan touching production network configs. Routing matters. Permissions matter. Logging matters. If a model is drafting a reply, that's one thing. If it's updating a CRM record tied to revenue, that's another. If it's writing an NSO service change, now we're in a completely different risk class.

[beat]

JOSH: What do you mean by routing, in plain English?

ERIK: Match the job to the brain. Small model for classification. Bigger model for synthesis. Local model for private low-risk stuff if you can get away with it. Human checkpoint before any action that can cost money or break trust. That's what PrimeRouter does in my world. It sends the right job to the right provider with failover and telemetry, instead of pretending every prompt deserves the expensive path.

JOSH: And that saves real money?

ERIK: Yep. People burn money because they treat all tokens like they're sacred. They're not. Most work is boring. Sort this. extract that. tag this. compare those fields. My Hermes box on the Mac M3 pushes around 70 tokens a second locally. That's great for certain classes of work. Then when I need a stronger model, PrimeRouter bumps it up. That is a systems problem, not a prompt problem.

[beat]

JOSH: Where do small teams mess this up?

ERIK: Permissions first. They get excited and hand the model broad access because the first run looked clean. That's reckless. Give it read access before write access. Make it stop before send. Make it stop before payment. Make it stop before merge. PrimeBus taught me that one fast. A guardrail that blocks a bad action is not friction. That's the product working.

JOSH: People hate hearing "add a checkpoint."

ERIK: They hate it right until the wrong invoice gets approved or the wrong customer gets emailed. Then suddenly checkpoints look pretty smart.

[beat]

JOSH: If somebody listening runs a twenty-person company, what's the first use case?

ERIK: Pick one ugly workflow with a boring finish line. New lead comes in, enrich it, score it, draft the response, create the CRM entry, stop. Or incoming invoice lands, extract fields, compare against PO, flag mismatches, stop. Don't start with "AI for the business." That's how you get a slideshow and no outcome.

JOSH: Narrow lane. Clear stop sign.

ERIK: Yep. Then review exceptions every day for two weeks. You want the mistakes. The mistakes show you the rails you forgot to build.

[pause]

JOSH: YellowKey. This one sounds like security people arguing over boot sequences while everyone else zones out.

ERIK: They shouldn't zone out. This one matters because BitLocker is one of those things teams assume is handled. Check the box, move on, sleep better. YellowKey is a reminder that disk encryption is only as strong as the path that unlocks the disk.

[beat]

JOSH: The public summary made it sound like somebody can crack a protected drive with files on a USB stick. Is that real?

ERIK: Real enough to change behavior. The details matter, though. It's not magic. It doesn't mean every stolen drive on earth is instantly open. Reports around the proof of concept say it relies on the target machine's TPM flow, and it doesn't work the same way if TPM plus PIN is enabled. That detail matters because lazy summaries make people either panic too much or ignore it completely.

JOSH: So what's the practical lesson?

ERIK: Convenience has a cost. TPM-only boot is convenient. User never types anything. Fleet rolls easy. But if the pre-boot trust chain gets abused, convenience becomes your weak point. If the laptop leaves the building, your threat model changes. Full stop.

[beat]

JOSH: What should infra teams do right now?

ERIK: First, find out where you're using TPM-only BitLocker. Not next quarter. Today. Second, decide which devices deserve TPM plus PIN. Exec laptops. field laptops. anything that travels. Third, test recovery procedures before you roll a policy and brick your help desk with lockouts. Security changes that don't account for operations turn into their own outage.

JOSH: That last part gets skipped a lot.

ERIK: Constantly. People love saying "turn on the secure thing." Cool. Did you test escrow? Did you test remote recovery? Did you test what happens when Susan in accounting forgets the PIN in an airport lounge. Security without runbooks is theater.

[beat]

JOSH: How does this connect to AI? On the surface it doesn't.

ERIK: It does, because we're about to hand more machines more authority. If your foundations are loose, adding agents on top is insane. Everybody wants the flashy part. Nobody wants to inventory pre-boot settings or map identity paths. But that's the work. In my setup, the event bus sees problems first. ScanBrief catches the external signal. PrimeBus distributes the event. Then a workflow can decide whether it's advisory, whether it's a ticket, whether it needs a human. The point is not "AI fixes security." The point is you need systems that react fast without guessing.

JOSH: So security news becomes an event, not a memo.

ERIK: Exactly. Memo culture is slow. Event culture moves.

[beat]

JOSH: Is this one of those stories where the fix is old-school discipline, not a new product?

ERIK: Mostly, yeah. Better boot policy. better segmentation. better handling for traveling endpoints. better runbooks. Dry answer. Correct answer. The industry keeps wanting a magic box. Half the time the fix is "stop taking convenience discounts on security."

[pause]

JOSH: Cisco. This one hit close to your world. Record quarter. AI demand up. Jobs cut anyway.

ERIK: Yep. And nobody in networking should pretend this is surprising. Cisco reported record quarterly revenue, up 12 percent year over year, then said it's cutting fewer than four thousand jobs while shifting investment toward AI, security, silicon, optics. That's not a contradiction. That's the map.

[beat]

JOSH: Explain that.

ERIK: Companies are buying different value now. They still need networking. Probably more than ever. But they don't want armies of people doing repeatable manual work around that networking. If a task can be turned into a system, the system wins. Humans move up to architecture, policy, exception handling, and weird troubleshooting. Or they get cut out. That's the blunt version.

JOSH: That's rough.

ERIK: Rough, yeah. Also true.

[beat]

JOSH: What kind of work is most exposed?

ERIK: Anything repetitive with clear inputs and expected outputs. Ticket triage. config audits. standard change generation. dashboard summaries. first-pass incident correlation. A lot of NOC work is sitting right in that lane. Same for a bunch of help desk flows. The teams that survive are the ones building the automations, not waiting to be assigned the old steps.

JOSH: That's basically the show's title.

ERIK: It is not a metaphor.

[beat]

JOSH: Does AI actually help that much in infra, or are people forcing it?

ERIK: Helps a lot when you treat it like part of a system. Hurts when you treat it like a lucky rabbit's foot. In my home lab, everything automated means 229 cron jobs are already doing background work before I touch the keyboard. ScanBrief scores 396 items daily across 56 sources and lands at 6:30 AM. PrimeTrader watches 109 tickers in three batches an hour during market hours. That's the mindset. You build the rails first. Then AI slots into the spots where judgment or language is the bottleneck.

JOSH: Not "ask a chatbot to run the company."

ERIK: Right. That's comedy. Useful comedy, sometimes, but still comedy.

[beat]

JOSH: What does Cisco's move say about the market for engineers?

ERIK: It says being good at one tool is not enough anymore. If all you know is the manual path, you're exposed. If you know Terraform, NSO, Python, APIs, event systems, model routing, maybe some Kubernetes, now you're useful in the new stack. The demand is shifting from operators who click to builders who wire.

JOSH: That's a different career path.

ERIK: Very different. And people need to stop waiting for permission. Nobody handed me a perfect environment and said, go build your bus. I built PrimeBus because I was sick of isolated tools. Everything needed to talk. Events in. actions out. agents subscribing. humans only where they add value. That's where the work is going.

[beat]

JOSH: Does that make the human less important?

ERIK: No. It makes lazy human labor less important. Big difference. The human who designs the rails, defines the stop points, reviews the exceptions, and understands the blast radius? That person matters more, not less.

JOSH: Wait, really?

ERIK: Yeah. Somebody has to know when the machine is confidently wrong. That's high-value work.

[beat]

JOSH: And the people who don't move?

ERIK: They're gonna have a rough couple of years.

[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 model router before you build a giant prompt library. Most teams do that backward. They spend weeks polishing prompts, then send every task to the same expensive model. Bad move. Start by labeling your work. Extraction. classification. summarization. planning. code review. long-context analysis. Then define a default model per class, a fallback model, and a hard rule for when a human has to approve output.

[beat]

ERIK: Keep the first version stupid simple. One YAML file is enough. Task type, primary model, fallback model, max cost, timeout, approval flag. Then log every call. Latency, cost, success, retry, who touched it. After a week, you'll know which jobs are overpowered and which ones are underpowered.

[beat]

ERIK: That's how PrimeRouter should be built in any environment. Not as "pick my favorite model." As traffic control. Cheap fast brain for cheap fast work. Bigger brain for high-risk synthesis. Local model where privacy and speed make sense. If one provider gets weird, fail over without drama. 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.

JOSH: It's Wednesday, May 13. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.

ERIK: AI's moving into the plumbing now. That's when it stops being a toy and starts replacing jobs.

JOSH: Stick around — Erik's got an AI pro tip at the end about using small models to route work before you wake up.

[pause]

JOSH: First headline. Google pushed Gemini deeper into Android yesterday. Widgets, app actions, handoff between devices. Useful, or just more keynote theater?

[beat]

ERIK: Useful if it survives contact with real life. The model isn't the product anymore. The handoff is.

[beat]

JOSH: Second one. A team open-sourced Needle, a 26 million parameter tool-calling model distilled from Gemini. Tiny thing. Big claim.

[beat]

ERIK: That's the right direction. You do not need a frontier model deciding whether to call Slack or hit an API.

[beat]

JOSH: Third headline. CERT dropped six dnsmasq CVEs this week, and patched dnsmasq 2.92rel2 is already out. How nervous should infra people be?

[beat]

ERIK: Pretty nervous if they forgot where dnsmasq is hiding. Old network glue is still production glue.

[pause]

JOSH: Start with Google. What actually happened?

[beat]

ERIK: Google used the Android show on May 12 to make Gemini more like system behavior and less like a separate chatbot. They showed agentic app actions, custom widgets built from prompts, better autofill, and cross-device handoff in Android 17. That's the important part. Not the demo voice. The placement.

[beat]

JOSH: Placement meaning it lives in the operating system now.

[beat]

ERIK: Exactly. Everybody spent two years arguing about which model writes the prettiest paragraph. Cute contest. The winning move is being one layer lower. If the assistant sits inside the OS, sees the app state, knows the file path, knows what device you're on, and can carry intent across screens, that's harder to compete with than another chat window.

[beat]

JOSH: Why does that matter to normal people and not just nerds like you?

[beat]

ERIK: Because normal people don't care what the model is called. They care whether the thing removed friction at 8:07 in the morning. Can it grab the PDF from the phone, open the right app on the laptop, and keep going without a login scavenger hunt. That's the bar. Nobody wakes up hoping for "multimodal orchestration." They want fewer taps.

[beat]

JOSH: Sounds obvious. Still feels rare.

[beat]

ERIK: Because the hard part isn't intelligence. It's state. It's identity. It's retries. It's permissions. It's what happens when the phone has the file but the laptop lost the token and now the "magic" turns into three prompts and a support article.

[beat]

JOSH: Wait, really? That's the main risk?

[beat]

ERIK: That's always the main risk. Success is easy to demo. Failure is what you bought. Every vendor shows the clean path. None of them show what happens when app number two hasn't updated yet, the user has two Google accounts, and the file is in the wrong workspace. That's where products get judged.

[beat]

JOSH: You're making this sound less like AI and more like boring systems work.

[beat]

ERIK: That's because it is boring systems work. The wiring is the product now. ScanBrief is the same lesson in a different shirt. It scores 396 items daily across 56 sources and lands at 6:30 AM. Nobody cares how the sources got ranked, deduped, summarized, and posted. They care that it shows up clean every morning. If one source stalls, the system has to recover without me playing detective before coffee.

[beat]

JOSH: So when Google says the assistant will move across devices, you hear an orchestration problem.

[beat]

ERIK: Yep. Same with the PAS website-pipeline. A lead comes in, assets move, copy gets drafted, changes get reviewed, pages get published. The customer sees a finished site. They do not see the ugly middle where files shift between services and a model has to make a decision without wrecking the output. AI helps. It does not excuse bad plumbing.

[beat]

JOSH: And if they get it right?

[beat]

ERIK: Then the expectation changes fast. Once users taste one clean handoff, they stop tolerating clunky software. That's when a lot of SaaS products find out they were charging money for tab switching.

[beat]

JOSH: That's a rude way to put it.

[beat]

ERIK: Still true.

[beat]

JOSH: Any part of Google's move you don't buy yet?

[beat]

ERIK: The phrase "seamless" means nothing to me anymore. I want to see the failure behavior. Does it retry. Does it fall back. Does it explain the miss in plain English. Can I override it without a fight. If those answers are bad, then it's another polished demo stapled to brittle glue.

[beat]

JOSH: So the takeaway isn't "Gemini got smarter."

[beat]

ERIK: No. The takeaway is Google is trying to make AI feel ambient. If they pull that off, builders who still ship AI as a separate tab are going to look old fast.

[pause]

JOSH: The Needle story feels smaller, but maybe more useful. Why did that jump out at you?

[beat]

ERIK: Because it's real engineering. Needle is a 26 million parameter model built for one job: tool calling. Not poetry. Not personality. Press the right button at the right time. That is a huge deal because most builders are lighting money on fire with giant models doing receptionist work.

[beat]

JOSH: Put that in human language.

[beat]

ERIK: If the task is "read this event, decide if it needs a webhook, call the function, pass the fields correctly," you do not need a premium model writing a TED talk to itself first. You need a fast bouncer at the door. Yes tool. No tool. Which tool. Next.

[beat]

JOSH: And this little model can do that?

[beat]

ERIK: In narrow lanes, yes. That's the point. The team claims it runs around 6000 tokens per second prefill and 1200 decode on consumer gear. Even if your real-world setup is lower, who cares. For routing and structured calls, tiny and fast beats huge and dramatic.

[beat]

JOSH: Is this where builders mess up with AI stacks?

[beat]

ERIK: Constantly. They buy one expensive model and ask it to do everything. Parse JSON. Decide retries. Read simple status pages. Route by confidence. Open tickets. Summarize logs. Then they wonder why the bill looks stupid and latency feels mushy.

[beat]

JOSH: What do you do instead?

[beat]

ERIK: Split the work by value. PrimeBus is built around that idea. It runs the auto-review and auto-merge pipeline on an event bus, and I keep the intelligence where it matters. Right now it's at 214 auto-fixes with a 78% success rate across a mesh of 3 Claude instances. Those numbers matter, but the better story is the guardrails. The cheap decisions happen early. The expensive judgment gets escalated.

[beat]

JOSH: So a small model handles the front door.

[beat]

ERIK: Exactly. It can classify an event, pull the right tool, maybe enrich the payload, maybe reject junk. Then a stronger model only wakes up for the hard part. That's how you keep cost sane and throughput high.

[beat]

JOSH: What's an example?

[beat]

ERIK: Test failure comes in. Tiny router model tags it. Build issue, lint issue, flaky test, dependency issue, whatever. If it's obvious, call the right collector, gather the files, prep the patch context. Then Claude gets a clean bundle instead of raw chaos. Same idea in PrimeDash too. Dashboards don't need Shakespeare. They need fast status, clean alerts, and the right action button.

[beat]

JOSH: This feels like the opposite of the usual AI sales pitch.

[beat]

ERIK: Good. The usual pitch is bad. Builders need to stop asking one model to be a philosopher, sysadmin, copywriter, and event router at the same time. That's lazy architecture.

[beat]

JOSH: Is there a downside to the small-model approach?

[beat]

ERIK: Sure. Narrow models are brittle outside their lane. If the schema drifts, if the tool descriptions are sloppy, if the task is fuzzy, they can get weird fast. That's not a reason to avoid them. That's a reason to define the lane. Tool names matter. Input shape matters. Return values matter. If your system prompt reads like a ransom note, don't blame the small model.

[beat]

JOSH: That's wild. So this is less about replacing the big model and more about stopping it from doing dumb chores.

[beat]

ERIK: Exactly. Premium models should spend time where judgment changes the outcome. Not deciding whether an event with a red status should call the red-status function. If your Ferrari is delivering mail, that's a management problem.

[beat]

JOSH: You sound personally offended by wasted tokens.

[beat]

ERIK: I am. Waste shows up in latency, cost, and fragility. All three are annoying.

[pause]

JOSH: The dnsmasq story is a different mood. Six CVEs. Patched release out. Why should people care if they don't even know where dnsmasq is running?

[beat]

ERIK: That's exactly why they should care. Dnsmasq sits in a lot of quiet places. Routers. Labs. Embedded setups. Container-ish edges. Home networks. Small office networks. A bunch of people are using it because it worked once and then nobody touched it for three years.

[beat]

JOSH: What did CERT actually flag?

[beat]

ERIK: CERT vulnerability note 471747 covers six issues disclosed on May 11. The set includes memory safety bugs, validation problems, DNS redirect risk, crash paths, and a local privilege escalation angle. The patched version is 2.92rel2. Translation: if dnsmasq is in your estate, this isn't a "put it on the list for next quarter" situation.

[beat]

JOSH: Why does old infra keep biting people like this?

[beat]

ERIK: Because old infra earns trust by being quiet. Quiet turns into invisibility. Invisibility turns into neglect. Then one day the component you forgot about is sitting on a very real blast radius.

[beat]

JOSH: What do you do with that kind of thing in your own setup?

[beat]

ERIK: Two things. First, inventory. If you don't know where the service is, you're already behind. Second, recovery posture. PrimeRestorer exists because patching is only half the story. Before you touch anything important, you need known-good restore paths for databases, system state, and offsite copies. Security people talk patch first. Operators should think patch plus rollback plus proof.

[beat]

JOSH: So this is not just "apply update and move on."

[beat]

ERIK: No. Especially not in infra. If you're running a home lab or a small business stack, patching a name service or DHCP component can expose all the weird assumptions you've been living with. Static leases. Old scripts. Forgotten containers. Random appliance behavior. You want visibility before the maintenance window, not during it.

[beat]

JOSH: Does AI help there, or does it just make the mess louder?

[beat]

ERIK: It helps if you use it like a mechanic, not a mascot. Feed it configs. Ask it to compare versions. Have it map references to dnsmasq across repos and host notes. Have it draft the checklist. Don't let it freestyle a network change into production because it sounded confident.

[beat]

JOSH: Dry understatement there.

[beat]

ERIK: Confidence is cheap. Rollback plans are better.

[beat]

JOSH: So today's security lesson is boring on purpose.

[beat]

ERIK: The boring stuff is the stuff that keeps the lights on. People chase shiny AI demos and forget the resolver underneath half their environment is one weird packet away from a bad day.

[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?

[beat]

ERIK: Build a router before you build an agent. That's the tip. Put a small cheap model or a hard rule layer in front of the expensive model and make it answer three things only: does this need a model, which tool should fire, and what context is required. Keep that prompt tiny. Keep the outputs structured. Then log misses and promote only the failures to a stronger model.

[beat]

ERIK: Most people start with one giant prompt and then act surprised when the system is slow, expensive, and confused. Don't do that. Use a front-door classifier, then a worker, then a reviewer if the action matters. That's how you make agentic workflows feel fast without trusting one brain to do every job in the building.

[beat]

ERIK: 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.

[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.

The brainstorming skill is designed for open-ended creative work where requirements need exploration — here the spec is fully defined, including exact structure, stories, format rules, and stats to use. Proceeding directly to the script.

JOSH: It's Tuesday, May 12. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.

ERIK: Criminal hackers used AI to find a real zero-day this week. Not theoretical — confirmed. We're going to talk about what that actually means.

JOSH: Stick around — Erik's got an AI pro tip at the end about locking down your automated build pipelines.

[pause]

JOSH: Quick headlines. GitLab cut staff and killed their CREDIT values. They're calling it adapting to the "agentic era." What's the read?

ERIK: When a company fires people and buries the culture doc in the same press release, that's a reset, not an evolution. Watch the product roadmap over the next 90 days. That's where the real story is.

[beat]

JOSH: Hacker News is debating whether Python matters anymore now that AI writes most of the code.

ERIK: Python's still the language AI writes in. You still have to read it, debug it, understand what it's doing at 2 AM when something breaks. The job shifted — more code review, less code typing. Fluency still matters.

[beat]

JOSH: And there's a new paper out on interaction models — real-time multi-modal AI collaboration with human-in-the-loop baked into the architecture from the start.

ERIK: That's the right framing. Most people treat human-in-the-loop as a fallback for when AI fails. It's not. It's a first-class architectural decision. You route to a human when confidence drops below threshold — proactively, not as a recovery path.

[pause]

JOSH: Okay, first deep dive. TanStack — 42 NPM packages compromised, 84 malicious versions published in 6 minutes. How does something like that even happen?

ERIK: Three techniques stacked on top of each other. GitHub Actions cache poisoning, a pattern called "Pwn Request" where a fork-based pull request can trigger a workflow with base-repo permissions it shouldn't have, and runtime OIDC token extraction. Those are the short-lived credentials GitHub issues to authenticate Actions workflows mid-run. Stack those three together and you get package publish rights without ever touching a stored secret.

JOSH: 6 minutes for 84 versions. That wasn't a human doing that manually.

ERIK: Someone built an automated publish pipeline specifically for this attack. They knew which packages to hit, they went fast on purpose — before maintainers could revoke anything. That's a planned operation with a prepared script, not a grab-bag exploit.

[beat]

JOSH: The report says no credentials were stolen though. Does that matter?

ERIK: It's the wrong question. What matters is code execution. Every project that ran npm install and pulled one of those 84 versions potentially executed attacker code — in a build pipeline, in a local dev environment, in a containerized prod deploy. The question isn't what they took. It's what ran.

JOSH: So the damage is invisible until it isn't.

ERIK: Right. And by the time you know, it's already happened.

[beat]

JOSH: How do you protect against this if you're running automation like you are?

ERIK: First layer: pin your dependencies. Exact versions, not ranges. "1.0.0", not "~1.0.0". Then go further and verify hashes. Every package on NPM and PyPI has a published integrity hash for every version. You fetch it, you compare, you fail fast if anything doesn't match. That catches supply chain tampering, not just version drift.

JOSH: Is that expensive to add to an existing pipeline?

ERIK: Twenty-five lines of Python. One build step. If your pipeline is already automated, it's an afternoon.

[beat]

ERIK: The second layer is scope reduction in CI. The OIDC exploit works because most GitHub Actions pipelines are granted way more permission than they actually need. If your runner has package publish rights, cloud credentials, API keys — all in the same workflow — that's a single point of compromise waiting to happen. Minimum permissions, scoped to exactly what each job requires.

JOSH: You're running Gandalf on every push. Does it flag that kind of thing?

ERIK: Gandalf looks for overly permissive configs, broad secret access, patterns that have caused incidents before. Since April 17th the auto-merger has run 1,215 attempts — 467 merged, 736 blocked. Those 736 aren't failures. That's the guardrail catching what shouldn't ship. Four code changes ran through that overnight and were in production before I had coffee this morning.

[beat]

JOSH: What's actually getting caught in those 736?

ERIK: Bad import paths, dependency version issues, missing error handling, patterns Gandalf recognizes from past incidents across the codebase. Some of them would've been fine in prod. Some of them would've been a 2 AM page. The gate doesn't try to predict which — it reviews everything, and anything it's not confident about doesn't merge. Twelve cases have been escalated to me since April 17th. Those are the ones that needed a human decision.

JOSH: Twelve out of 1,215. That's a pretty quiet morning routine.

ERIK: That's the whole point of building it that way.

[pause]

JOSH: Second story. Google confirmed that criminal hackers used AI to find an actual novel software vulnerability. Not leveraging an existing CVE — original vulnerability research, AI-assisted. What changes?

ERIK: Nothing changed suddenly. The tools security researchers use to find bugs — static analysis, fuzzing, pattern recognition across large codebases — those are all things frontier models are now genuinely competent at. The gap between "has the knowledge" and "can do the work" closed. And that applies to attackers exactly the same way it applies to defenders.

JOSH: How much faster does AI make the research cycle?

ERIK: A senior security researcher finding a novel memory corruption bug in a complex codebase might spend weeks. You're reading assembly, writing custom fuzzers, building a mental model of the control flow. With an LLM in the loop, you describe the vulnerability class, hand it the codebase, iterate on hypotheses in hours. The model doesn't replace the expertise — it compresses the iteration time dramatically.

[beat]

JOSH: You're running 89 services in production right now. Does this change your threat model?

ERIK: It changes the timeline assumption. My old mental model was: stay ahead of human attackers moving at human speed. That's done. An attacker's scanner can move as fast as my monitoring pipeline. PrimeBus processed 2,519 events across 9 projects today — my defenders are automated. So are their scanners now. The reaction-time gap is gone.

JOSH: So both sides are running automation and the humans are just setting the objectives.

ERIK: Right. And that means the quality of your automation matters more than ever. Sloppy automation on the attack side still finds bugs. Sloppy automation on the defense side gives you false confidence, which is worse than nothing.

[beat]

JOSH: Does this change what Gandalf should be checking for?

ERIK: It has to expand. Right now Gandalf is focused on code quality and known-bad patterns. The next layer is configuration hygiene — exposed debug endpoints, overly permissive CORS headers, unvalidated inputs at API boundaries. Those are exactly the patterns an AI scanner excels at finding because they're consistent and recognizable across codebases. If my reviewer isn't catching them, an attacker's tooling might find them first.

JOSH: You're building your defender to think like an attacker's scanner.

ERIK: That's exactly where this is going.

[pause]

JOSH: Quick third one — Claude is now fully available on AWS. Full API, managed agents, code execution, the whole platform integrated into AWS billing and IAM. Big deal?

ERIK: It's a procurement story more than a technology story. Enterprise teams that are AWS-native had to manage Anthropic credentials outside their existing cloud infrastructure — separate billing, separate IAM, separate audit trails, separate compliance reviews. That friction is now gone. It's in your VPC, your existing dashboard, your established approval process.

JOSH: You're not using this. You've got your own setup.

ERIK: PrimeRouter handles my LLM routing — multi-provider failover, priority-tier queuing, fleet telemetry. I'm not going through any cloud intermediary. But for teams that have to satisfy compliance requirements or work inside existing cloud procurement contracts, this is the path of least resistance to a frontier model. The model didn't change. The buying motion did. At this stage of the market, distribution matters more than features.

[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: Wire a hash verification step into any agentic pipeline that installs packages. Before anything executes, call PyPI's or NPM's package API — every published version has an integrity hash. Fetch it, compare it against what's installed, fail the step if they don't match. It's about 25 lines of Python and you can drop it into any existing workflow in an afternoon. ScanBrief scored 35 items across 54 sources this morning and a supply chain attack was one of the top signals. This is the week to add that check. That's your tip. Use it.

[pause]

JOSH: We also drop daily market picks and automation tips on YouTube — search Build or Be Replaced.

[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.

[beat]

ERIK: Build or be replaced.

JOSH: If you want these signals in your inbox every morning, scanbrief.dev. See you tomorrow.

The brainstorming skill is loaded, but the user's prompt is itself the complete design spec — format, voice, word targets, story sources, live stats, and all section requirements are fully defined. The only creative choice left is story selection and framing, which I'll make now and execute directly.

**Story plan:**

- **Deep Dives (3):** "Going back to writing code by hand" (AI architecture gap), Maryland $2B grid upgrade (AI infrastructure costs), Local AI needs to be the norm + M4 local models (connected)

- **Headlines (3):** CVE-2024-YIKES supply chain attack, AWS return regret, Obsidian PHANTOMPULSE RAT

- **Bobaverse stats to weave:** 3,004 automation events, 1,147 auto-merger attempts / 446 merged / 690 blocked, 46 items / 54 sources, 89 services, 65 projects

---

JOSH: It's Monday, May 11. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.

ERIK: AI writes the features. You own the architecture. Miss that distinction and you're rewriting everything in six months.

JOSH: Stick around — Erik's got an AI pro tip at the end about the three lines you write before you hand anything to a model.

[pause]

JOSH: Three quick headlines. First — a supply chain attack hit roughly four million developers. Walk me through it.

ERIK: JavaScript package, poisoned. Payload traveled through a Rust compression library into a Python tool. Credentials stolen at scale. It got patched by a crypto miner who wanted clean compute and the malware was in the way.

JOSH: The patch came from a cryptominer. Accidentally.

ERIK: Unintentional remediation. That's open-source security in 2026.

[beat]

JOSH: Someone went back to AWS after leaving and immediately remembered why they left.

ERIK: Platform trust has a memory. They spent years accumulating reasons to leave. Left. Came back. Every reason was still there. At some point the exit cost is lower than the stay cost. They hit that point twice.

[beat]

JOSH: And an Obsidian plugin got weaponized into a remote access trojan. What's different about this one?

ERIK: The command-and-control server ran on the Ethereum blockchain. You can pull the plugin, block the domain, burn the infrastructure — you can't take down a blockchain. The attacker's control channel is permanent. That's a meaningful upgrade to the standard RAT playbook.

[pause]

JOSH: Biggest story of the day. A developer just posted that they're throwing out an entire Kubernetes TUI codebase they built with Claude — called k10s — and rewriting it from scratch by hand. What happened?

ERIK: Vibe-coding worked exactly as advertised. Features shipped fast. The model solved every problem in front of it. The problem is that every solution was local. The model doesn't carry the full system in its head. It patches. It doesn't refactor. So after a few months you've got a codebase that functions and has no coherent shape underneath it.

JOSH: Is this a prompting problem? Better instructions fix it?

ERIK: Partially. But there's a ceiling. The model is genuinely good at implementation. It's not good at knowing when a new feature should force a rethink of the design. That judgment requires standing outside the feature and looking at the whole system. You can prompt toward it. You can't replace it.

[beat]

JOSH: You're building with Claude constantly. Do you hit this?

ERIK: All the time. PrimeBus processed 3,004 automation events today across 65 projects. Every one of those projects has code a Bob wrote, I reviewed, or both. The ones that hold up are the ones where I defined the interfaces before the model touched anything. The ones that get messy are the ones where I said "figure it out."

JOSH: So it's about what you hand the model, not the model itself.

ERIK: You're the architect. The model is a skilled contractor who's never seen your codebase. You wouldn't hand that contractor a napkin sketch and say "build whatever makes sense." You'd give them a spec. Doesn't have to be long. Three lines per module: what it does, what goes in, what comes out. Hand it that — you get something coherent. Without it, you get something that works and nothing more.

[beat]

JOSH: The author rewrote from scratch. Would you do that?

ERIK: For a personal tool — yeah. For something in prod — you refactor incrementally, you don't throw away working infrastructure. But if you're starting fresh with the lessons from the first pass? That rewrite is probably cleaner and faster than the original build was. The mistakes cost you once. They don't have to cost you twice.

[pause]

JOSH: Maryland's utility regulator is asking FERC to block two billion dollars in grid upgrade costs. The argument: Maryland ratepayers are paying for infrastructure that powers AI data centers in other states. How does that happen?

ERIK: The grid is regional. When a data center in Virginia needs a massive capacity upgrade, PJM — the grid operator — spreads those costs across the whole region. Maryland didn't build those data centers. They're still paying for the transmission infrastructure that feeds them. Projected hit is $1.6 billion extra for Maryland ratepayers over the next decade.

JOSH: That's an insane externality.

ERIK: It's the hidden cost of scale. Everyone talks about GPU hours and API pricing. Nobody talks about the substations, the copper in the ground, the water cooling. The physical infrastructure bill is just starting to show up in visible places.

[beat]

JOSH: What does that mean for someone running your setup?

ERIK: I run 89 services on a home lab on residential power. My marginal compute cost is essentially electricity. That math works because I'm not trying to run a hyperscaler. But the models I'm calling into — those are running in exactly the kind of data centers Maryland is being charged for. That cost is baked into API pricing. As grid costs climb, API prices follow. It's not abstract.

JOSH: So local inference starts making economic sense.

ERIK: Eventually — not yet for frontier models. But for the work that doesn't need frontier, the local case gets stronger every quarter. The auto-merger has run 1,147 attempts since April 17 — 446 merged, 690 blocked by Gandalf before they ever touched prod. Those blocks aren't failures, that's the guardrail working. But every one of those attempts consumed tokens. At scale, your API bill is your power bill.

[pause]

JOSH: Two connected stories today — one argues local AI should be the default, not cloud APIs. Another shows someone getting real work done on an M4 with 24 gigs running local models. Where's local inference actually at?

ERIK: The hardware case is real. Apple Silicon can run models in the 7-to-13 billion parameter range without breaking a sweat. Ollama, llama.cpp — tooling is solid. The honest trade-off is that a local 13B model isn't Claude Sonnet. It's also not trying to be.

JOSH: What tasks actually go local?

ERIK: Anything transformation-heavy where quality variance is low. Dedup, classification, relevance scoring. ScanBrief scored 46 items this morning across 54 sources. The relevance pass on that feed is a candidate for local. The final summary generation — I want frontier quality there. The preprocessing layer? Local wins on speed and cost. Rough estimate — you could route a third to forty percent of a typical agentic pipeline to local without any user-visible quality drop.

JOSH: There's also a hardware attestation piece today about tech companies using attestation to control which models run on their devices.

ERIK: That's the long game concern. Apple already does this in places. If attestation becomes a gate on which models a device will run, that's a market structure decision dressed up as a security feature. You're paying for hardware and not owning what it runs. The reason local inference matters isn't just cost and privacy — it's that you actually control the stack. Open hardware matters for the same reason open software matters.

JOSH: And that's what ties the Maryland story and the local AI story together.

ERIK: Every layer of lock-in — infrastructure, platform, hardware — has a cost that eventually becomes visible. The Maryland regulator made one of those costs visible. Same thing is coming for the others.

[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: Before you hand any new module to an AI agent, write three lines first. What it does. What goes in. What comes out. That's the spec. Takes two minutes. Without it, the model makes its own choices about scope and boundaries — and those choices accrue until you have coherent modules inside an incoherent system. With it, you get implementation that fits the design you actually intended. Applies to prompts, to functions, to services, to agents. Same rule at every level. Three lines before you delegate. That's your tip. Use it.

[pause]

ERIK: That tip is straight out of The Autonomous Engineer — my book on building systems that run themselves. Grab it on Amazon.

[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.

Writing the episode 20 script now.

---

JOSH: It's Friday, May 8. Episode 20. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.

ERIK: Agents, layoffs, and code that shipped while you were sleeping.

JOSH: Stick around — Erik's got an AI pro tip at the end about why your prompts are not your error handlers.

[pause]

JOSH: ScanBrief pulled 47 items this morning across 54 sources. Three things to flag quickly. Canvas is offline — ShinyHunters got in, student data exposed. How bad?

ERIK: Bad. Names, emails, private messages all out. The group is threatening to leak unless schools negotiate. Instructure patched, but the platform is still rocky. Education infrastructure runs on thin security budgets and legacy systems. Soft targets. This won't be the last one.

JOSH: Cloudflare cutting 20% of its workforce. That's not a small trim.

ERIK: That's hundreds of engineers at a company running critical internet infrastructure. The official framing will be restructuring. The real story is what changed in the cost-benefit math of maintaining a large engineering org.

[beat]

JOSH: And Dirtyfrag — a universal Linux privilege escalation. How worried should people be?

ERIK: Very. Universal means every major distribution is in scope. User goes to root. If you're running exposed Linux boxes, you're patching this today. Not this week.

[pause]

JOSH: Let's go deeper on that agents piece. Hacker News today — "Agents need control flow, not more prompts." You're running five agents in your fleet right now. Is this right?

ERIK: Completely right. This is the most important thing people are missing about building with AI right now. Prompts don't make agents reliable. Architecture does. Those are different problems.

JOSH: What does "control flow" actually mean in practice?

ERIK: It means your agent has a decision tree, not a conversation. When something fails, the code knows the next step. It doesn't ask the model what to do — the model isn't in charge of the logic. PrimeBus processed 1336 automation events today across 6 projects. None of that runs on prompts. The router reads the event type, checks the rules, dispatches. The model comes in when there's something to interpret — a diff to review, an error message to classify. But the branching is code. Always.

JOSH: Where do most people get this wrong?

ERIK: They write one big prompt and call it an agent. Then they're surprised when it fails unpredictably. A prompt chain isn't an agent. It's a suggestion chain. The moment something goes sideways, there's nothing to catch it. So they add more instructions. The prompt gets longer. The model starts ignoring the earlier parts because the context is too cluttered. Now you're debugging 4,000 tokens of natural language instead of a function. Nobody wins that fight.

[beat]

JOSH: So what belongs in the model versus what belongs in code?

ERIK: The model handles ambiguity. Anything where the right answer is fuzzy or requires judgment. Gandalf reviews code diffs for my auto-merger — that's a model job. Evaluating whether a fix is sound, whether it has downstream risk, whether the approach makes sense — that requires understanding context. Code can't do that. But "did the merge succeed?" is a boolean. Write the boolean in code.

JOSH: The auto-merger numbers are pretty striking. How's it tracking?

ERIK: 1121 attempts since mid-April. 435 merged, 675 blocked by Gandalf, 11 escalated to me. Eleven. Out of eleven hundred runs. That's not a 38% success rate — people look at that wrong. The 675 blocks are the guardrails doing exactly what they're supposed to do. Those are fixes that shouldn't have gone to prod. Gandalf caught them. That's the system working.

JOSH: And the 11 that came to you — what were those?

ERIK: Architectural edge cases. Things where a fix touched something with downstream implications Gandalf flagged but couldn't resolve itself. Those are the ones that actually need a human in the loop. Eleven. That's the whole point of building it this way — I'm not the bottleneck for the other eleven hundred.

[pause]

JOSH: Cloudflare. 20% cut. The show's called Build or Be Replaced — is this what that looks like?

ERIK: Pretty much. Cloudflare grew its engineering headcount over years assuming a certain ratio of humans to shipped output. AI changed that ratio. You need fewer people to build and maintain the same systems. Eventually that shows up in headcount decisions. This isn't unique to them — it's the math changing industry-wide.

JOSH: Is this AI replacing engineers or something else?

ERIK: It's both. Some of it is macroeconomic. Some of it is that tooling which required a dedicated team two years ago can now be built by a much smaller group. The engineers who survive are the ones building the systems, not the ones being replaced by them.

JOSH: How do you actually get to the other side of that? It sounds like a lot of upfront work.

ERIK: It is. But so is the alternative. Every hour I don't spend building automation is an hour I spend doing the thing manually. Right now I've got 93 services running on one prod server and 65 projects emitting telemetry into PrimeBus. I didn't hire a team to run that. I built systems that run it. You build toward that state. You don't flip a switch one day — you make one thing autonomous, then another, then another.

JOSH: What's the first thing someone should automate?

ERIK: Whatever you do every day that has a clear trigger and a predictable outcome. If you can describe it as "when X happens, I do Y" — that's a cron job or an event handler waiting to exist. Start there.

[pause]

JOSH: One more — "AI slop is killing online communities." Short take?

ERIK: ScanBrief scored 47 items this morning and I can already tell you which sources are polluted with generated filler. No specific claim. No point of view. No named entity. Just content-shaped noise that fills a feed. The scoring helps filter it — low-signal items drop out before the brief gets built. But the volume is real and it's getting worse.

JOSH: Is there a fix?

ERIK: Not a systemic one. You can score for specificity, penalize vague generalities, filter algorithmically. You can degrade the slop in your own pipeline. But as long as publishing more is cheaper than publishing better, you're fighting a current. The communities that survive this are the ones where real people actually show up and say something specific.

[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 using prompts to handle errors. This is the most expensive habit I see in AI workflows. Something breaks, so you add more instructions to the prompt — "make sure to handle the case where the API returns null." The prompt grows. The model gets inconsistent. Now you're debugging natural language instead of code, and you have no stack trace and no way to reproduce it reliably. Here's what you do instead: write the error handler in code. Write a retry with backoff. Write an event that fires on failure and routes to an agent that can actually deal with it. The model is not your try-catch. Code is your try-catch. If you're building agents, build them like software. Not like chatbots. That's your tip. Use it.

[pause]

JOSH: Binge all five episodes this weekend plus our YouTube shorts — links at buildorbereplaced.dev.

[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.

JOSH: It's Thursday, May 7th. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.

ERIK: Today we're talking about the gap between looking like you know what you're doing and actually knowing.

JOSH: Stick around — Erik's got an AI pro tip at the end about picking the right model for the right job.

[pause]

JOSH: First up — the US Library of Congress just added SQLite to its recommended storage formats list. Is this actually significant?

ERIK: Very. The Library of Congress lists formats they trust to still be readable in 50 years. They're not recommending it because it's trendy. Single file, no server required, fully documented, open spec. If you've been going back and forth between JSON, CSV, and SQLite for long-term data storage, that debate is over. Federal endorsement.

[beat]

JOSH: Valve open-sourced the hardware design files for the Steam Controller. CAD files, Creative Commons. What's the play?

ERIK: They dropped STP and STL files — the native CAD format and the 3D print format. Non-commercial, attribution required. That scope is enough for the community to build replacement parts, better grips, modded accessories. Valve gets a longer product life without spending a dollar on it. Smart move wrapped in generosity.

[beat]

JOSH: And Google just relaunched reCAPTCHA as something called Fraud Defense. What changed?

ERIK: reCAPTCHA was built to separate humans from bots. That problem is mostly solved. Fraud Defense is built for a web where AI agents are transacting alongside humans — buying tickets, filling forms, making API calls. It classifies which of three actors is operating: human, bot, or AI agent. The fact that Google shipped this is an admission that the agentic web isn't coming. It's already here.

[pause]

JOSH: Alright, there was a piece on Hacker News today I want to dig into. Title: "Appearing productive in the workplace." The argument is that generative AI lets non-experts produce output that's indistinguishable from expert work. That feels like something you have strong opinions about.

ERIK: The piece splits this into two cases. A novice mimicking a senior in the same field — that's the obvious one. You find out eventually when something breaks and they can't debug it. The more dangerous case is someone producing work in a domain they have no background in at all. A project manager generating network architecture diagrams. An analyst writing security policies. Nobody in the room knows what good looks like.

JOSH: And the output looks completely authoritative.

ERIK: Polished, confident, grammatically correct. The model doesn't signal uncertainty the way a junior engineer does. It just delivers. And if the reviewer doesn't have the domain knowledge to catch what's missing, the decision gets made on something that looks like expert output but has no grounding.

[beat]

JOSH: You use AI to move faster than anyone else. How is what you're doing different from this?

ERIK: I know when the output is wrong. That part doesn't come with the tool. It comes from years of breaking things in production, debugging at 2am, watching systems fail in ways you didn't predict. The model can generate a PrimeBus event routing config that looks perfect. I can tell in thirty seconds if the retry logic is off. Someone who's never built a message bus can't make that call.

JOSH: So the AI accelerates execution but doesn't replace the judgment.

ERIK: The judgment is what you're paying for. Always was.

[beat]

JOSH: How do you protect your own systems from this? You're running AI that reviews and merges code automatically.

ERIK: Guardrails in the pipeline. I run an automated code reviewer — Gandalf — between every change and production. Since mid-April the auto-merger has run 997 times. 384 changes made it to production. 602 got blocked.

JOSH: Wait — 602 blocked. That's not a problem?

ERIK: That's the system working. A 38.5% merge rate means the guardrails are doing their job. If I was seeing 90% merges, I'd be auditing the reviewer. Eleven cases escalated to me directly — those were the genuinely ambiguous ones where Gandalf couldn't make the call. Every block either caught a real issue or enforced an explicit rule.

JOSH: What does it actually catch?

ERIK: Asyncio task handling, SQL race conditions, retry logic that could burn API credits in a loop. Every time Gandalf blocks something I haven't seen before, I write a rule and drop it in the fix-guides directory for that project. Next service I build starts with those lessons already baked in. You can't skip that part and call it automation.

[beat]

JOSH: That's the thing the article is missing — most companies don't have a Gandalf.

ERIK: Most companies have humans doing code review after they've already looked at a polished output. ScanBrief scored 37 items this morning across 54 sources. If the relevance scoring was misconfigured, every brief downstream would be wrong — and it would arrive in your inbox looking perfect. The format doesn't tell you if the signal is bad.

JOSH: So the lesson is: build the review layer before you give AI the surface area.

ERIK: Build the cage first. Then let it run.

[pause]

JOSH: There's a related piece today — "Vibe coding and agentic engineering are getting closer than I'd like." What's the actual convergence?

ERIK: Vibe coding is describe-and-generate. You tell the model what you want, it writes code, you run it. Agentic engineering is the model taking autonomous action — executing commands, modifying files, making decisions in a loop without a human reviewing each step. They've been separate workflows. The new LLM toolkit in that piece starts merging them. Same interface, now with autonomous execution baked in.

JOSH: Why is that a problem if the individual pieces work fine?

ERIK: Because the person who's been vibe coding doesn't know what constraints an autonomous agent needs. They've been treating the model as a code generator. Now it's a code executor. That's a different thing entirely. The surface area for unintended behavior goes up fast.

[beat]

JOSH: Give me a concrete example of what goes wrong.

ERIK: Early on I had an automation touching files in a directory I hadn't explicitly excluded from its scope. Nothing blew up, but it could have. That was the lesson. Now every agent has an explicit whitelist — what it can read, what it can write, what it can delete. Anything outside that list, it stops and escalates. I've got 93 services running in production right now. A runaway agent with undefined scope is a bad day.

JOSH: So you learned this before the tools were this capable.

ERIK: PrimeBus processed 2,285 automation events today across nine projects. Overnight, one code change went through Gandalf and merged while I was asleep. That's the system working. But it only works because I defined the scope before I deployed anything. Echo is my dev environment — everything new runs there first. When it breaks on Echo I write a rule. When it passes Echo it goes to prod with the guardrails already tested.

JOSH: How does someone starting with these new tools avoid the trap?

ERIK: Answer four questions before you deploy anything autonomously. What can it read? What can it write? What can it delete? What does it do when it's not sure? If you can't answer all four, you're not ready. The tools moving fast doesn't mean your governance has to lag. Write the rules first.

[beat]

JOSH: The author said "closer than I'd like." You sound less worried than they are.

ERIK: The capability isn't what worries me. I want more of it. What worries me is people deploying capability without the architecture to hold it. That's not an AI problem — that's just engineering. You test in the lab before you go to production. That rule didn't change because the thing doing the work is a language model.

JOSH: Same boring engineering answer.

ERIK: It's always the same boring answer. Nobody wants to hear it until something breaks.

[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: Model selection. Most people default to the biggest, most capable model for everything. That's expensive and slow on the wrong tasks. You need tiers. Fast, cheap model for high-frequency, low-stakes work — routing, classification, quick summaries. Heavy model for architecture questions, ambiguous errors, anything where being wrong costs real time or money. In PrimeRouter I route by priority tier. Routine calls hit Sonnet. Anything involving retry logic, failure analysis, or a decision I'd want to audit — that escalates to Opus. I define the tiers before I write a line of code. That's what keeps token costs flat even as the system grows. Define your tiers first. 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.