Genie 4 | Build or Be Replaced

Today: Genie 4 | New 14‑inch MacBook Pro with M6 | Early progress on accelerating drug development with Claude Episode date: 2026-10-28.

Download MP3 →

Transcript

JOSH: It's Wednesday, October 28. This is Build or Be Replaced, powered by ScanBrief.dev. I'm Josh, here with Erik Anderson. ERIK: Today's theme is trust. Every layer you build on, from firmware to SSH to the model, is something you chose to trust. JOSH: Stick around. Erik's got an AI pro tip at the end about why the agent that writes the code should never be the one that approves it. [pause] JOSH: Headlines. Apple dropped a new 14-inch MacBook Pro and a new iMac, both with M6. Do you care? ERIK: I do, a little. Bill, my M3 Mac, does the iOS builds and the GPU work. M6 means the M3 gets older, and the Mac keeps being the cheapest way to run models locally. I'm not upgrading today. I'm watching the benchmarks. JOSH: Next one. Python's experimental JIT is getting removed. Wait, really? ERIK: Really. They shipped it as an experiment, it didn't pay off, and they're cutting it. I respect that. Most projects keep dead features around for a decade. Half my cron jobs are Python, and none of them will notice. JOSH: And there's a project putting the entire US Code into Git. ERIK: Line-level history on every statute. Author, date, prior text. I've been telling network teams to put configs in Git for fifteen years. Congress gets there first, apparently. Dry humor aside, it's a great use of version control. [pause] JOSH: Deep dive time. Tailscale published an investigation and remediation update on Tailscale SSH. Why should people care? ERIK: Because SSH is the front door to everything. I run Neo, Morpheus, Echo, Tank, all of it over SSH. When a tool that manages access to that door says "we investigated something," you read the whole post. Not the headline. The whole post. JOSH: Is this a panic situation? ERIK: I'm not going to tell you it is, because I haven't seen anything that says so. What I will say is that the update itself is the good part. They investigated, they wrote it up, they told you what to do. That's the behavior you want from a vendor. JOSH: So what does an engineer actually do with that? ERIK: Three things. Check which of your hosts rely on a managed SSH layer versus plain keys. Check your logs for anything odd around the dates in the post. And make sure you know how to turn that layer off in a hurry. The rollback matters more than the feature. [beat] JOSH: How does that connect to what you run? ERIK: My rule is nobody gets a permanent pass. Any change to SSH or auth configs on my fleet needs line-by-line approval. No exceptions, even for me. The agents can propose all day. They don't touch the door. JOSH: That sounds slow. ERIK: It's about thirty seconds of reading. A bad auth change costs you a weekend. [pause] JOSH: Staying on trust, there's also a code-execution flaw in UEFI firmware. The headline says it threatens most servers on the planet. ERIK: Headlines love that sentence. Here's the real version. Firmware sits below your OS, below your containers, below your EDR. If something runs there, nothing above it can honestly tell you it's clean. JOSH: So all those security layers don't help? ERIK: They help with everything above the firmware. They can't vouch for the firmware. That's the gap. And most of us just never think about it, because the box boots and the lights are green. JOSH: What do you do about it? ERIK: Same boring answer as always. Inventory. Know your hardware and know your firmware versions. Know which vendor ships the patch and how you roll it out. I've got 140 services running on the production server right now. Every one of them assumes the machine underneath is honest. That's a big assumption, and I'd rather write it down than pretend it isn't there. JOSH: Can't you automate the patching? ERIK: You can automate the inventory and the alerting, absolutely. PrimeSentinel watches my services and flags stuck jobs. A firmware check is the same idea one layer down. Firmware flashes themselves I'd still stage on one box first. A bad flash on a server is not a rollback situation. [beat] JOSH: That's the lab-first approach you keep talking about. ERIK: Always. Echo first, then prod. It's not a slogan. It's the reason nothing has gone sideways in a bad way. [pause] JOSH: Last deep dive. Anthropic posted early progress on accelerating drug development with Claude. Where do you land on that? ERIK: Cautiously excited. I use Claude all day to build and ship, so I know how good it is when the problem is well defined and the output can be checked. Drug development has a lot of both. Literature review, hypothesis generation, sorting through huge piles of structured data. JOSH: But it's medicine. Doesn't it get things wrong? ERIK: It does get things wrong. That's why the design matters. In medicine the model proposes and the lab disposes. A wet lab doesn't care how confident the model sounded. It either works or it doesn't. That's a built-in verifier, and honestly most software doesn't have one that clean. JOSH: How does that compare to what you do? ERIK: Same shape. PrimeBus has run 2,295 auto-merge attempts since June fifth. 1,507 merged. 788 got blocked by Gandalf, and zero escalated to me. JOSH: Seven hundred eighty-eight blocks. That sounds like a lot of failures. ERIK: It isn't failures. Those are the guardrails doing the job. The model proposes a change, a separate reviewer checks it against the actual code, and the weak ones don't ship. That's the whole architecture. Generation is cheap now. Verification is the product. JOSH: And that's the drug story too. ERIK: Right. Whoever builds the best checking layer wins. The model is the easy part. [beat] JOSH: Is there a version of this where it goes wrong? ERIK: Sure. When the checking layer is the same model grading its own homework. Then you've built a very confident echo chamber. Independent check, different eyes. In medicine that's the wet lab. In my pipeline that's Gandalf. [pause] JOSH: Quick scan of the rest. Ghostty 1.4 shipped. ERIK: Full Kitty graphics support, plus memory and speed gains. I live in terminals over SSH, so graphics that actually work remotely is a real win. JOSH: And ScanBrief had a big morning. ERIK: ScanBrief scored 92 items across 56 sources today. Genie 4, the Lean formalization of finite simple groups, a survival LLM that runs entirely offline. A lot of signal for people who just want to read one brief at six thirty. [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 agent that wrote the change approve the change. Use a second model, or a second session with a fresh context and a different prompt, and give it one job: find what's wrong. Don't ask it if the code looks good. Ask it to list the three most likely ways this breaks in production. Then make the first agent fix only what the reviewer can point to in the actual diff. Reviewers who can't cite a line are guessing, and you should ignore them. That one habit is why 788 bad changes never reached my main branch. That's your tip. Use it. [pause] JOSH: If you're building toward financial independence through automation, Erik's first book walks through the whole path. Wait, that's yours to say. 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.