Build or Replaced
Today: The Exponential Time Hypothesis Is False [pdf] | Three new discoveries in mathematics, computer science, and physics | Anthropic S-1 Episode date: 2026-10-27.
Download MP3 →JOSH: It's Tuesday, October 27. This is Build or Be Replaced, powered by ScanBrief.dev. I'm Josh, here with Erik Anderson. ERIK: Today's theme: adding more people to a late project makes it later, and now we're adding agents. JOSH: Stick around. Erik's got an AI pro tip at the end about making one model check another model's homework. [pause] JOSH: Headlines. First, a preprint on arXiv claims the Exponential Time Hypothesis is false. Is that as big as it sounds? ERIK: If it holds up, yes. ETH says problems like SAT can't be solved in sub-exponential time. A lot of "this is provably hard" results sit on top of it. Big if, though. Extraordinary claims get picked apart by a hundred grad students before lunch. JOSH: So wait and see. ERIK: Wait and see. Don't rewrite your crypto stack on a preprint. [beat] JOSH: Next. Someone is doing SIMD-accelerated computer vision on a two-dollar microcontroller. ERIK: I love this one. Two dollars. No GPU, no cloud, no API bill. Everyone's pointing a frontier model at problems a tiny chip can solve. Before you call an API, ask whether the answer fits on hardware that costs less than a coffee. JOSH: And there's an essay called "Websites Should Be Printable." ERIK: Correct, and boring in the best way. Clean HTML, sane CSS, a print stylesheet. Half the web breaks when you hit print. It's an easy win nobody bothers with. [pause] JOSH: Okay, deep dive. The one that caught my eye: "The Mythical Man-Month in the Age of AI." Fred Brooks, 1975. Does it still hold? ERIK: It holds better than anything else from that era. Brooks said adding people to a late project makes it later. The reason is communication overhead. Every new person is another channel you have to keep in sync. JOSH: But agents aren't people. They don't need onboarding. ERIK: They do, just differently. A new agent has zero context. It doesn't know your repo, your conventions, or why that weird retry exists. You pay for onboarding every single session, and you pay in tokens instead of salary. [beat] JOSH: So the overhead is still there. ERIK: It moves. It stops being meetings and becomes context. I run a fleet of twelve agents. Neo, Homer, Bill, Echo, Gandalf, a mix of Claude and GPT. If I just throw them all at one problem, they step on each other. Two agents edit the same file, one reverts the other, and you've burned an hour of compute arguing with yourself. JOSH: How do you stop that? ERIK: Lanes. Every agent has a job, and the jobs don't overlap. Echo is dev and QA. Neo owns prod. Gandalf reviews. Nobody edits prod directly. The structure is the whole product. Brooks would call it surgical team organization. I call it not letting the robots freelance. [pause] JOSH: Does Brooks' other big point survive? "No silver bullet"? ERIK: Even more so. Brooks split complexity into essential and accidental. Accidental is boilerplate, syntax, glue code. AI crushes that. Essential complexity is figuring out what the system should actually do, and that hasn't moved. A model can write a perfect function for the wrong requirement in about four seconds. JOSH: That's a little terrifying. ERIK: It's fast wrongness. Which is why the bottleneck shifted. Writing code used to be the slow part. Now deciding and verifying is the slow part. [beat] JOSH: So what does verifying look like at your scale? ERIK: Here's a real number. The PrimeBus auto-merger has run 2,295 attempts since June 5th. 1,507 merged. 788 got blocked by Gandalf. Zero escalated to me. JOSH: Seven hundred eighty-eight blocked. That sounds like a lot of failure. ERIK: It's the opposite. Those 788 are the guardrails working. Code that would've shipped and broken something, stopped by a reviewer that doesn't get tired or feel polite. Generation is cheap now. Rejection is where the value is. JOSH: So the Mythical Man-Month answer is... ERIK: More workers isn't the fix. A better gate is. The agent count didn't make me faster. The review layer let me trust the output without reading every diff. [pause] JOSH: Second deep dive. Anthropic S-1. That's the IPO filing, right? ERIK: That's what the title says. I haven't read the whole thing, so I'm not going to pretend I have. But a filing like that matters regardless of what's in it. An S-1 means the numbers become public. Revenue, costs, compute commitments, risk factors. Right now most of this industry runs on vibes and leaked slides. JOSH: Why should a builder care? ERIK: Because you're building on top of these companies. Pricing, rate limits, model deprecations, all of it comes from their economics. When those economics get audited, you learn how stable your foundation is. [beat] JOSH: Does that change how you build? ERIK: It reinforces something I already do. PrimeRouter sits between my services and the models, and it can swap providers. I don't hardwire one vendor into everything. If a price changes or a service goes down, the chain falls through to the next backend. Treat every model as a replaceable part. JOSH: Even the one you like best? ERIK: Especially that one. Loyalty is a nice thing for a dog. In infrastructure it's a single point of failure. [pause] JOSH: One more. Sebastian Raschka put out an eleven-hour course on building a reasoning model from scratch. ERIK: Take it. Eight videos, PyTorch, starting from a pretrained transformer, covering inference scaling, reinforcement learning, and distillation. Free education like that used to cost a degree. JOSH: I thought you went and got the degree. ERIK: At 43. Closing a loop, not because I needed it. But the people who learn how these models actually reason will build better systems than the people who treat them as a magic box. You don't have to train one yourself. You do need to know why it fails when it fails. JOSH: Is that the move for someone who's feeling replaceable? ERIK: Yeah. Understand the machine well enough to run it. Nobody replaces the person who can explain why the pipeline broke. [pause] JOSH: Quick pulse check. What's ScanBrief seeing today? ERIK: ScanBrief scored 98 items across 56 sources this morning, average relevance around 52. Plenty of strong stuff, but most of it is noise. Filtering is the product. [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: Never let the model that wrote the code be the one that approves it. Use a different model, or at minimum a fresh session with no memory of writing it, and give it one job: find what's wrong. Hand it the diff and the requirement, not the author's explanation. When the reviewer reads the author's reasoning, it inherits the author's blind spots. JOSH: And that's how you get 788 blocks. ERIK: That's exactly how. A reviewer with fresh eyes and a skeptical prompt will catch what the author can't see. Set it up once and it runs forever. 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. [pause] ERIK: Build or be replaced. JOSH: If you want these signals in your inbox every morning, scanbrief.dev. See you tomorrow.