Transcript
JOSH: It's Thursday, November 12. This is Build or Be Replaced, powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.
ERIK: Today's theme: the machines found the bug, the machines proved the code, and somebody still has to patch the server.
JOSH: Stick around. Erik's got an AI pro tip at the end about knowing exactly what you're running before the next CVE lands.
[pause]
JOSH: Episode 147. ScanBrief scored 73 items across 56 sources this morning. Where do we start?
ERIK: Headlines. Then the real stuff.
[pause]
JOSH: First headline. Tim Peters has died.
ERIK: He's the Python core developer who wrote the Zen of Python. Every Python engineer has quoted it in a code review without knowing it was him. "Readability counts" has settled more arguments than any linter.
[beat]
JOSH: Meta is losing its longtime head of infrastructure.
ERIK: Santosh Janardhan joined in 2009 and is stepping down. Sixteen years running infrastructure at that company is a long time. When the person who knows where all the bodies are buried leaves, the runbooks had better be written down.
JOSH: And a British chip startup just raised a lot of money.
ERIK: Fractile, $720 million. That makes them one of the top-funded players in purpose-built AI chips. The money is going to hardware because everybody is tired of renting it. Which leads to the third deep dive.
[pause]
JOSH: Okay, deep dives. Apache HTTP Server. Critical remote code execution. How bad?
ERIK: Bad. It's a race-condition bug in mod_http2, the HTTP/2 module. CVE-2026-119618. An unauthenticated remote attacker can corrupt heap memory and get arbitrary code execution. No login, no foothold needed.
JOSH: And it affects a lot of servers?
ERIK: Roughly two-thirds of cloud deployments, according to the write-up. The call is immediate patching or mitigation. Not "next sprint."
[beat]
JOSH: Wait, who found it?
ERIK: Researchers from Anthropic and Google Cloud. An AI lab and a cloud provider both looking at the same old piece of internet plumbing. I'll say it plainly: this is the pattern now. Machines read decades-old C code faster than humans ever did, and they find what's been sitting there.
JOSH: So what do you actually do on a morning like this?
ERIK: First question is "do I run it, and where?" Not "how do I patch." If you can't answer the first one in five minutes, you're already behind. I have 145 services running on the production server right now. If a CVE dropped for one of them, I want to know which ones in under a minute.
JOSH: How do you know?
ERIK: PrimeDash shows every running service in the lab on one screen. PrimeSentinel watches them and flags anything stuck. That's not a security tool. But if you've got a list of what's alive, you've done half the work.
JOSH: And the other half?
ERIK: Mitigation. The text says patch or mitigate. For a web server, mitigating usually means turning the vulnerable module off until the patch is in. Disabling a module is a config change. A config change is something you can roll back. I'd rather have a rollback path at 2 AM than a heroic upgrade.
[beat]
JOSH: Does it bother you that an AI company found this?
ERIK: No. It makes me nervous about the other direction. The same capability that finds the bug can be aimed at your server. Patch windows used to be measured in weeks. Assume they're shorter now.
[pause]
JOSH: Next. DeepMind's AlphaProof Code. What's the headline number?
ERIK: Three million lines of proofs, written in Lean, by AI agents. The target was Signal's core encryption. They verified it.
JOSH: Three million lines? That's a lot of proof.
ERIK: That's the point. A human team would never write it. Nobody has the patience, and nobody would pay for it. Formal verification has always worked. It just cost too much. Now an agent writes the proof and a checker confirms it. The agent can be wrong. The checker can't be talked into anything.
JOSH: Explain that. Why does that matter?
ERIK: Because it's the cleanest version of something I already believe. Don't trust the model. Trust the thing that checks the model. Lean either accepts the proof or it doesn't. No vibes.
JOSH: How does that connect to what you build?
ERIK: PrimeBus works on the same idea, just much dumber. A fix gets generated, then tests run, then review runs, then a merge gate decides. The model proposes. A deterministic check disposes. Mine are pytest and a review gate. Theirs is a proof assistant. Same shape.
[beat]
JOSH: So are you going to start proving your own code?
ERIK: No. I'm not writing Lean for a cron job. But look at where this goes. Encryption libraries and payment code get proofs first. Then it gets cheap enough to reach the boring middle. That's the thing to watch.
JOSH: Does it change anything for a regular builder today?
ERIK: Yes. Next time you ask an agent to write something critical, ask what checks it. If the answer is "me, reading it," you have a weak link.
[pause]
JOSH: Last deep dive. CoreWeave. Money problems?
ERIK: Money structure. Their third-quarter earnings say they're asking customers to pay up front. They're borrowing against specific contracts. And they're looking at shorter-term deals, which could mean higher prices.
JOSH: Why would shorter contracts mean higher prices?
ERIK: Short deals give the provider less certainty. Less certainty gets priced in. That's the source's read, and it tracks. Prepayment tells you the provider needs cash now to build capacity. That's not a company with endless runway. That's a company funding a race.
JOSH: What does that mean for someone who isn't buying GPUs by the rack?
ERIK: It flows downhill. If compute gets pricier and contracts get shorter, the API price you pay doesn't stay flat forever. I route every LLM call through PrimeRouter on Neo. Every call goes through one gateway, with caps and fallbacks. If one provider's price moves, I change one config, not forty scripts.
JOSH: So that's your hedge?
ERIK: One gateway. Everything else is a plug. If your app has a vendor's SDK hardcoded in thirty places, you don't have a stack. You have a hostage situation.
[beat]
JOSH: And the Fractile money?
ERIK: Same story from the other side. Lots of capital chasing alternatives to the rented-GPU model. I'd treat any single compute vendor as temporary.
[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: Build an exposure inventory, and make your agent maintain it. Today's Apache bug is the reason.
Here's how. Open a session with Claude and ask it to write one script. The script walks your machines and records every listening port, the process behind it, and the exact software version. It writes the result to a single JSON file in a git repo. Run it nightly from cron.
Now when a CVE drops, you don't search your memory or grep a dozen boxes. You ask the agent one question: which entries in this file match mod_http2? You get an answer in seconds, and the git history shows you what changed since last week.
Two details matter. Give the agent the file, not the whole network. And make the script read-only. It reports, it never fixes.
That's your tip. Use it.
[pause]
JOSH: Track your freedom score and net worth with the Freedom Blueprint app — free download, link in the show notes.
[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.