Transcript
JOSH: It's Tuesday, November 3rd. This is Build or Be Replaced, powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.
ERIK: Today's theme: the machines are editing themselves, and the humans are still slow on patches.
JOSH: Stick around. Erik's got an AI pro tip at the end about making your agents prove their work before you believe them.
[pause]
JOSH: Headlines. First one. Someone built a way to talk to your computer without making a sound?
ERIK: Subvocal. It's a Show HN project, an RF-based silent-speech interface. It picks up internal speech from inside the mouth. No lip movement, no microphone. Earlier versions used EMG sensors, and nobody wants electrodes on their face at a coffee shop.
JOSH: Is that real or a demo?
ERIK: Demo. But I'd use it. I talk to Claude all day. Typing is the bottleneck now.
[beat]
JOSH: Next. NIST put out a cybersecurity guide for solopreneurs.
ERIK: Good. Most one-person shops have no security policy at all. Reading a free government checklist beats nothing, and it's more than most small companies do. Do the basics. Backups, MFA, patching.
JOSH: Speaking of backups, there's a post called "If you love a website, back it up."
ERIK: Right. Sites disappear. Companies shut down, hosting lapses, a founder gets bored. If something matters to you, mirror it. Takes ten minutes and a cron job.
[pause]
JOSH: Okay, deep dive time. The one everyone's talking about. Astra, and a tiny Qwen model.
ERIK: This one's a big deal. Astra is described as a GPT-6-class model. Researchers at MIT CSAIL showed it reading and editing a neural network's weights directly. Not training. Not fine-tuning. Programming the weights.
JOSH: Wait, really? It just opens the model up and changes numbers?
ERIK: Pretty much. They had Qwen 3.5, the 0.8 billion parameter version, stuck in a reasoning loop. Astra diagnosed it and edited about 0.016 percent of the parameters. MATH-500 accuracy went from 7.9 percent to 48.1 percent. IFEval roughly tripled.
JOSH: That's wild. From 7.9 to 48?
ERIK: Yeah. And no training loop. No big GPU bill, no data set. A surgeon instead of a boot camp.
[beat]
JOSH: Is that actually better than fine-tuning?
ERIK: Too early to say. It's one demo on a small model. But the direction matters. Right now, when a model misbehaves, you retrain or you prompt around it. This says a model can debug another model the way an engineer debugs code.
JOSH: How does that compare to what you do?
ERIK: It's the same pattern I run with PrimeBus. A test fails, an agent reads the failure, finds the cause, makes a small targeted change. Small diff, specific cause. Astra is doing that at the weight level instead of the code level.
JOSH: And what's the catch?
ERIK: Verification. Editing is easy. Knowing the edit didn't break something else is hard. Look at the numbers I can actually cite from the pipeline. PrimeBus auto-merger has run 2,295 attempts since June 5th. 1,507 merged. 788 blocked by Gandalf. Zero escalated to Erik.
JOSH: Those 788 blocked, that's a lot.
ERIK: That's the story. Those aren't failures. That's the guardrail working. Gandalf reviews every push and says no when the diff doesn't hold up. A model that rewrites weights needs the same thing. A reviewer that isn't the author.
[pause]
JOSH: Does this change what builders should do today?
ERIK: Not today. But watch it. If models can patch models, the gap between a big lab and a guy with a home lab shrinks. A small model with the right fix beats a big one that's looping. I run local models for cheap jobs, and a fix like this would make that tier a lot more useful.
JOSH: Okay. Different kind of story. Ubuntu.
ERIK: Ubuntu still hasn't released a patch for a Linux zero-day RCE. Almost three weeks. There's a post titled "WTF Ubuntu?" and that tells you the mood.
JOSH: Three weeks. Is that long?
ERIK: For remote code execution, yes. Painfully long. Everyone with a public-facing box has been exposed for weeks, waiting on a vendor.
[beat]
JOSH: What do you do when you're waiting on a patch like that?
ERIK: You don't wait. You reduce what's exposed. Firewall rules, turn off the affected service, put it behind a VPN or a reverse proxy. I've got 142 services running on the production server right now. If one of them is the vulnerable path, I need to know in minutes, not days.
JOSH: How do you know?
ERIK: PrimeDash shows every running service in the lab on one page. PrimeSentinel watches for stuck or dead ones. That's the inventory. You can't mitigate what you can't list. Most people with a zero-day open can't answer "what's listening on my network" in under an hour.
JOSH: So the lesson isn't "Ubuntu bad."
ERIK: No. Every distro has a slow week. The lesson is your exposure shouldn't depend on one vendor's schedule. Know your attack surface, have a kill switch for each service, and have a mitigation you can apply without a patch.
[pause]
JOSH: Last deep dive. Momentic, out of Y Combinator. Agents that catch bugs by using your app.
ERIK: Launch HN, W24 batch. The pitch is AI agents that use your app like a user would and report what breaks. Not unit tests. Actual clicking.
JOSH: Isn't that just Selenium with a chatbot?
ERIK: Partly. Selenium scripts are brittle. Button moves, selector breaks, test fails, nothing is actually wrong. An agent that looks at the screen and understands intent doesn't care where the button is. That's the real pitch.
JOSH: Do you buy it?
ERIK: I buy the idea. I've written enough Selenium to know the pain. But I'd want two things. Deterministic checks alongside the agent, and a clear record of what it did. An agent that says "looks fine" isn't evidence. A recorded run with steps is.
JOSH: Who's it for?
ERIK: Small teams that ship fast and have no QA. That's most people building with AI right now. You can generate a feature in an hour. Testing it still takes a day. That's the gap tools like this are chasing.
[beat]
JOSH: There's also a piece called "Resisting Cognitive Surrender." Relevant here?
ERIK: Very. If the agent writes the code, tests the code, and approves the code, you've handed over your judgment. That's the trap. Use the agent for the work. Keep the decision.
[pause]
JOSH: Quick pulse check before we wrap. What did ScanBrief find this morning?
ERIK: 94 items scored across 56 sources. Average relevance 52.9. Quiet day for volume, but the Astra story alone is worth the read.
JOSH: And this is episode 140.
ERIK: 140. Still showing up.
[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 receipt. Whenever you hand an agent a task, add one line to the prompt: "End with a list of every file you changed, every command you ran, and the output of the test that proves it worked." Then read the receipt, not the summary.
JOSH: Why does that matter?
ERIK: Summaries are where models drift into confident fiction. A receipt is checkable. You can run the same command and compare. If the agent says tests pass and the receipt shows no test output, you caught it in ten seconds. Pair it with a second reviewer that didn't write the code. That's what Gandalf is for me. Author and reviewer should never be the same model in the same session.
JOSH: Simple enough.
ERIK: Costs you one sentence in the prompt. 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.