Mon–Fri · 6 AM ET
← All Episodes
EP  • 00:07:26

USPS now offers official ZIP Code boundary polygons | Build or Be Replaced

Today: USPS now offers official ZIP Code boundary polygons | Amazon saves millions by replacing Sonnet 4.6 with Strands Decider | Google Gemini routes personal Drive files and emails to human annotators Episode date: 2026-11-10.

Download MP3 →

Transcript

JOSH: It's Tuesday, November 10. This is Build or Be Replaced, powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.
ERIK: Today's theme: the expensive model is the most expensive way to do the boring job.
JOSH: Stick around. Erik's got an AI pro tip at the end about putting a cheap classifier in front of your big model.
[pause]
JOSH: Headlines. USPS now publishes official ZIP code boundary polygons. Quarterly updates. Why should anyone care?
ERIK: Because ZIP boundaries used to be approximations somebody sold you. Now it's the source. PrimeDistro does direct-mail outreach, and targeting by delivery area is the whole game. Cleaner polygons means fewer postcards landing on the wrong side of a line.
JOSH: Next. CoreWeave's bond yield just topped 16 percent.
ERIK: That's not a bond, that's a warning label. When the company renting out the GPUs has to pay 16 percent to borrow, AI credit stress is real. Build like compute prices can move on you.
JOSH: And Google Gemini reportedly routes personal Drive files and emails to human annotators.
ERIK: Read the fine print on every assistant you connect to your data. If a human can see it, treat it as shared. I keep customer data off any lane I can't audit.
[pause]
JOSH: Let's start with the big one. Amazon replaced Claude Sonnet 4.6 in a ticket-labeling pipeline with something called Strands Decider. What happened?
ERIK: Amazon had an internal pipeline that labels tickets. Sonnet 4.6 was doing it. Good model, accurate. They swapped in Strands Decider, which is a small decision model built on Gemma 4, the E4B size. Same labeling accuracy, and token costs dropped about 98 percent. Millions of dollars a year.
JOSH: Wait, 98 percent? Really?
ERIK: Really. And it shouldn't surprise anyone. Labeling a ticket isn't reasoning. It's picking one of twelve buckets. You were paying a frontier model to push a button.
[beat]
JOSH: So does that mean the big models are overpriced?
ERIK: No. It means people use them wrong. A frontier model is for the hard part. Planning, ambiguity, a weird stack trace at two in the morning. Classification, routing, yes or no decisions, those belong on something small.
JOSH: Do you see that in your own stuff?
ERIK: Constantly. ScanBrief scored 98 items across 56 sources today. Scoring and ranking is exactly that kind of job. You don't need a genius to decide whether a Hacker News post about paper airplanes is relevant to a network automation engineer. It isn't. Sorry, paper airplanes.
JOSH: Paper airplanes made the list, though.
ERIK: Score of 65. Everything on that list scored 65, which tells you what kind of news day it is. Anyway, the point stands. Every call in a pipeline should ask: does this need judgment, or does it need a label?
JOSH: How do you decide in practice?
ERIK: Start with the expensive model while you learn the task. Log what it returns. Once you see the output is always one of a few values, replace it. Keep the big model as the escalation path for the cases the small one isn't sure about.
JOSH: That's basically what PrimeBus does, right? Different job, same idea.
ERIK: Same idea. PrimeBus has run 2,295 auto-merge attempts since June fifth. 1,507 merged. 788 were blocked by Gandalf. Zero escalated to me. That's a decision pipeline. Most of the work is cheap checks, and the expensive review only matters at the gate.
JOSH: And the 788 blocked, that's not failure.
ERIK: Not failure. That's the guardrails working. A bad change got stopped before it touched prod, and I didn't get a phone call. Blocked is the product.
[pause]
JOSH: Second story. Anthropic released Claude Fable 5.2 and Mythos 5.2. What's the headline?
ERIK: One shared model upgrade, about 30 percent faster on coding, reasoning, and vision. Pricing dropped to 8 dollars per million input tokens and 40 per million output. And there's an optional fast-generation mode that runs 2.5 times faster.
JOSH: Is the faster mode just a speed boost, or is there a catch?
ERIK: There's always a catch. Faster costs more, and it matters where you use it. If a human is waiting on the response, pay for fast. If a background job is chewing through a queue at midnight, don't.
JOSH: They also mention hardened cybersecurity and biology safeguards. Does that matter to a builder?
ERIK: It matters in one way. Safeguards can change refusal behavior. If you run agents that touch security tooling, test your prompts again after every model bump. I have workflows where a model update changes what it'll do, and you want to find that in the lab on Echo, not in prod on Neo.
JOSH: So don't just flip the switch.
ERIK: Never. New model goes into the dev lane first. Run the same test set. Compare cost, speed, and output. Then promote. I run 144 services on the production server right now. I'm not letting a model swap surprise any of them.
[beat]
JOSH: Does the price drop change anything for you?
ERIK: It changes the math on the hard lane. But I'd rather see it next to the Amazon story. Cheaper frontier tokens are nice. Not sending tokens at all is better. Both can be true. Use the best model where it earns its price, and use the small one everywhere else.
JOSH: Does a 30 percent speed bump actually matter?
ERIK: In agent loops, yes. An agent that takes forty steps compounds latency. Thirty percent off each step turns a coffee-break task into a quick one. And PrimeSentinel, my watchdog for stuck jobs, gets fewer false alarms when jobs finish on time.
JOSH: Stuck-job detection gets easier if the jobs are faster.
ERIK: Slow and stuck look alike at three in the morning. Anything that separates them helps.
[pause]
JOSH: Third thing. The Gemini story. Humans reading personal Drive files and emails. Is that worse than it sounds?
ERIK: It's exactly as bad as it sounds. Human review for quality is standard, but people don't read the terms. If you wire an assistant into your inbox, assume somebody might see a sample. For client work, that's a contract issue, not just a privacy one.
JOSH: So what do you do?
ERIK: Know which lane handles which data. Customer data goes only through paths I can audit. Public news summaries can go anywhere. That's it. A routing rule, not a philosophy.
[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: Find your most expensive repeated LLM call and ask one question: does it return one of a fixed set of answers? If yes, log two weeks of its outputs. Then build a small classifier or a rules layer that handles the confident cases, and only send the uncertain ones to the big model. Track how often you escalate. That number is your savings. Amazon got to 98 percent on one pipeline. You probably won't, but 50 is realistic, and that's money back every month. 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.