Build or Replaced
Today: Why I Am a Programmer | Overgrown, when plants invade your abandoned tabs | IBM CEO says open weight models are now good and cheap on-prem Episode date: 2026-10-26.
Download MP3 →JOSH: It's Monday, October 26. This is Build or Be Replaced, powered by ScanBrief.dev. I'm Josh, here with Erik Anderson. ERIK: The frontier labs are cutting prices and the open models are catching up. The moat is shrinking while you watch. JOSH: Stick around. Erik's got an AI pro tip at the end about stopping your agents from drowning in their own context. [pause] JOSH: Episode 134. ScanBrief scored 75 items across 56 sources this morning. What stood out? ERIK: Plenty. Three quick ones first, then we go deep on the model economics. [beat] JOSH: First. A ten-year-old CNN beat frontier LLMs at diagnosing skin cancer. Ten years old. Really? ERIK: Really. A small model trained for one narrow job beat the biggest general models on that job. Specialized beats general when the task is narrow and the data is clean. Remember that when someone tells you the biggest model always wins. JOSH: Next. Dreamhost has an active incident and is forcing password resets. ERIK: If you host there, rotate credentials today and check your logs. And if you reuse passwords, that's the real incident. Put a password manager on it. JOSH: Last one. Someone removed Next.js from their blog and hit 100 on every Lighthouse category. ERIK: Of course they did. A blog is static text. You don't need a framework to serve words. I'd rather ship plain HTML and spend the saved hours on something that earns money. [pause] JOSH: Okay, the big one. IBM's CEO says open-weight models are now good and cheap enough to run on-prem. What's the claim? ERIK: About ninety percent of frontier performance at ten to twenty-five percent of the token cost. And you can run it in a few racks. JOSH: Is that true, or is it a CEO talking? ERIK: It's a CEO talking, and IBM sells the racks. Discount it a little. But it matches what I see. Hermes runs on my Mac M3. It's a local agentic backend running qwen3-coder at around seventy tokens a second. It costs me nothing per token. JOSH: Nothing? What can it actually do? ERIK: The boring work. Summaries, first-pass triage, log reading, drafting fixes. It's not writing the hard code. But boring work is most of the volume. JOSH: So where does the frontier model still win? ERIK: Hard reasoning and long multi-step planning. That's where I still pay. Everything else goes to the cheapest model that can do the job. Ninety percent of the quality for a quarter of the cost is a great deal when the task is easy. [beat] JOSH: Does this mean enterprises abandon the big labs? ERIK: No. It means they stop paying frontier prices for commodity work. The routing decision is the skill now. Which model gets which task. JOSH: And that's what you've built. ERIK: It's how my whole fleet runs. We have 143 services running on the production server right now. If every one of them called the most expensive model for everything, I'd be broke by Friday. Instead each call gets sent to the cheapest model that passes. [pause] JOSH: Which ties into the next story. OpenAI and Anthropic cut prices by up to eighty percent within a few months. Why? ERIK: Three reasons. Inference got cheaper. Open models are applying pressure from below, which is the IBM story. And they want your workload locked in before you build around someone else. JOSH: Is that good for builders? ERIK: Great for builders. Bad if you hardcode one vendor. Prices dropped eighty percent this time. They can move the other way next time. JOSH: So how do you protect yourself? ERIK: Put an abstraction between your code and the model. Your services call one gateway. The gateway decides who answers. When a price drops, you change one config. When a vendor has an outage, the gateway fails over. Nothing else changes. JOSH: That sounds like a lot of work. ERIK: It's a weekend. And it pays back the first time a provider goes down at two in the morning and your pipeline keeps running. [beat] JOSH: Does cheaper tokens change how you build? ERIK: Completely. When tokens were expensive, I rationed them. Now I'll let an agent read ten files instead of one. I'll run a second model as a reviewer. That's what PrimeBus does. Tests fail, it generates fixes, and Gandalf reviews the result before anything merges. JOSH: Give me numbers on that. ERIK: Since June fifth, the auto-merger has run 2,295 attempts. 1,507 merged. 788 got blocked by Gandalf. Zero escalated to me. JOSH: Seven hundred and eighty-eight blocked. That sounds like a lot of failures. ERIK: It's the opposite. Those are the guardrails working. A bad fix got caught before it touched production. Cheap tokens let me afford that second look on every change. That's what cheaper models buy you. Not just savings. More checking. [pause] JOSH: Last deep dive. There's a piece called The State of Context Engineering. What is that? ERIK: It's the discipline of deciding what goes into the model's window and what stays out. Prompts are the easy part. The hard part is what the model sees when it makes a decision. JOSH: Why does that matter more now? ERIK: Because agents run long. A single chat is easy. An agent that works for an hour collects tool output, old attempts, stale file contents, all of it. The window fills up with junk, and quality drops before you hit the limit. JOSH: Have you seen that happen? ERIK: Constantly. A fix agent gets a huge log dumped on it and starts chasing the wrong error. The fix isn't a smarter model. It's feeding it the failing test, the relevant file, and the last error. Nothing else. JOSH: So less is more? ERIK: Less, but the right less. ScanBrief works the same way. It pulls from 56 sources, scores everything, and only the top items reach the summary step. If I fed the model everything, I'd get mush. Filter first, then think. JOSH: Does a bigger context window solve it? ERIK: It helps the ceiling, not the quality. A bigger window just lets you make a bigger mess. Treat context like a 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: Give every agent task a context budget. Before you run it, write down the three things the model needs: the goal, the one file or dataset that matters, and the last error or result. Everything else gets summarized or left out. ERIK: Then add a rule. If a tool returns more than a few hundred lines, trim it or summarize it before the agent sees it. Most agent failures I debug come from the model reading too much, not too little. ERIK: Try it on your worst-performing agent today. Cut its input in half and watch what happens. 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. [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.