Mon–Fri · 6 AM ET
← All Episodes
EP  • 00:06:50

Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding | Build or Be Replaced

Today: Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding | A new Go API for JSON | Insects don't swarm at LED lights Episode date: 2026-11-05.

Download MP3 →

Transcript

JOSH: Nine hundred forty-eight Linux kernel privilege escalation bugs. Found, validated, and patched by a pipeline.
ERIK: No human triaged a single one of them.
[beat]
JOSH: It's Thursday, November fifth. This is Build or Be Replaced, powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.
ERIK: Today's theme: AI is finding the bugs we spent forty years ignoring.
JOSH: Stick around. Erik's got an AI pro tip at the end about making your agents prove their work before you trust it.
[pause]
JOSH: Headlines. First up, Kimi K3.1 Pro weights are out. Three trillion parameters, million-token context. Three trillion. Come on.
ERIK: It's real, and it's open. The trick is sparse attention. The model doesn't look at every token against every other token, so a million-token window stops being a budget-killer.
JOSH: But who's actually going to run three trillion parameters?
ERIK: Almost nobody, at home. That's not the point. Once weights like that are public, every hosting provider can serve them and every lab can study them. The gap between closed and open keeps shrinking, and that's good for anyone paying API bills.
JOSH: Second one. Ripple-65B Preview, billed as U.S. open-weight state of the art on agentic coding.
ERIK: Sixty-five billion is a size you can actually run. That's the interesting part.
JOSH: So it's a win?
ERIK: It's a preview. Preview means benchmarks, not production. A leaderboard tells me how it does on a leaderboard. I'll believe it when it survives my fixer chain in PrimeRouter, with real broken tests and no hints.
JOSH: Third. Residential proxies detected on the White House network.
ERIK: Somebody's devices are running IP-masking software on a government network. Those apps turn your connection into an exit node for strangers.
JOSH: Isn't that just a dumb app somebody installed?
ERIK: Probably, yes. That's a hygiene problem, not a hack. But a stranger's traffic leaving through your network is the kind of hygiene problem that becomes a hack later. Nobody was watching the egress.
[pause]
JOSH: Deep dive time. OpenAI says an autonomous pipeline found, validated, and patched 948 local privilege escalation bugs in the Linux kernel. I keep saying the number because I don't believe it.
ERIK: Believe the word "validated." Anyone can point a model at code and get a pile of maybe-bugs. Proving each one is real, then proving the patch doesn't break anything, is the hard part. That's where the work is.
JOSH: Couldn't a human team do that?
ERIK: Sure, over years. Kernel maintainers are buried. Everybody knew these bug classes existed. Nobody had the hours.
JOSH: Okay, but how do you know the pipeline isn't just confidently wrong 948 times?
ERIK: You don't, unless there's a gate after the model. That's the whole design. Generate a fix, then make something independent try to knock it down. I run the same shape with PrimeBus. Code changes get reviewed by Gandalf before they merge, and the blocks are the best part of that story. Those aren't failures. That's a guardrail working while I sleep.
JOSH: So a system that blocks nothing is a problem?
ERIK: A system that blocks nothing is just unwatched.
[beat]
JOSH: Flip side. Attackers get the same tools.
ERIK: They do. Same model finds the bug for them. So the patch side has to be automated too. If your fix cycle is a ticket and a two-week sprint, you lose that race.
JOSH: That sounds like a scary sentence for a security team.
ERIK: It's a scary sentence for teams without a pipeline. The backlog of known issues finally gets worked. The bottleneck moves to review and deploy, and that's a pipeline problem. I like pipeline problems.
[pause]
JOSH: Next. A post on Hacker News says AI is forcing us to pay off forty years of technical debt. I think that's a little dramatic.
ERIK: Mostly true, though. Think about why the debt existed. Nobody could justify the human hours. Rewrite the old config parser. Add tests to the script from 2009. Now an agent does it overnight for pennies.
JOSH: You're saying that from experience, or from the headline?
ERIK: Experience. I came up in network automation, Cisco NSO, years of scripts held together with hope. Those are exactly the things I point agents at now. Boring work, clear success criteria, tests to prove it.
JOSH: But everyone wants AI to build the shiny new thing.
ERIK: Right, and that's the wrong target. The real money is in the unglamorous cleanup nobody wanted to fund. A new app is a gamble. Deleting a decade of fragile glue is a sure thing.
JOSH: Where does it go wrong?
ERIK: No tests. If your codebase has no tests, the agent has nothing to check itself against. Garbage in, confident garbage out. Tests first, then let the agent loose.
[pause]
JOSH: Last one. Ask HN: what's your current AI agentic setup? Yours is not simple.
ERIK: It's a fleet. Five machines, each with a job. Neo runs prod. Echo is dev and QA. Homer watches the monitoring. Bill handles the Mac builds and GPU work. Tank is the newest host. They talk over a bus.
JOSH: Why not one big smart agent? Seems easier.
ERIK: Easier to build, harder to trust. One agent grading its own homework is useless. Separate the builder from the reviewer. Different model, different context, different incentives.
JOSH: Isn't that just more moving parts to break?
ERIK: It is. That's why PrimeRouter sits in front of everything. Every call picks a backend from a priority chain, skips what's unavailable, fails over on errors. Caps and kill switches on all of it.
JOSH: And new stuff?
ERIK: Ships dark. Off by default. You turn it on deliberately, one step at a time. Boring, and it's why nothing's blown up my bill.
JOSH: Where does ScanBrief fit?
ERIK: Same idea, different job. Lots of input from dozens of sources, a scoring filter, and a human reading only what survived. That's how you stay informed without drowning.
[pause]
JOSH: Quick break. Erik, who's paying for this episode?
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 same agent write the change and approve it. Split the roles. Have one prompt produce the fix. Then hand only the diff and the requirements to a fresh session with one instruction: find the reason this should not merge.
JOSH: Why not give it the first agent's reasoning too?
ERIK: Because that's the excuses. A reviewer who hasn't heard the excuses finds more problems. Use two different models if you can. Different training, different blind spots. 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.