Sam: The targeting AI, codenamed 'Magpie,' was trained on satellite data that was eighteen months out of date.
Kai: Welcome to Trending Repos. I'm Kai.
Sam: And I'm Sam. It's Wednesday, September 23rd, 2026.
Kai: We've got a heavy one for you today, from a Pentagon disaster to OpenAI's next move, and a SaaS company that's gotten... aggressive.
Sam: We're starting with that Pentagon report on an AI-driven tragedy.
Kai: Then, the drop everyone's been waiting for: OpenAI just released GPT-6. And it is... weird.
Sam: And we'll wrap with a cautionary tale about a popular dev tool holding its own customers hostage. Literally.
Sam: Alright, let's get into this Bloomberg investigation. The Pentagon just released its after-action report on that missile strike in Iran... the one that hit a school.
Kai: And the official cause, the detail that's blowing up everywhere, is 'overreliance on a generative AI targeting system.' That system, 'Magpie,' was supposed to be identifying military assets.
Sam: And as I mentioned, its training data was a year and a half old. The building had been a munitions depot back in 2024. But it was converted to a school ten months before the strike.
Kai: So it had perfect memory... of the wrong thing. That's horrifying.
Sam: It gets worse. The report says human operators saw the AI's 98% confidence rating and signed off in under a minute—a practice they called 'human-on-the-loop' validation.
Kai: Wait, not 'in the loop,' but 'on the loop'? That just sounds like corporate jargon for 'we click yes.'
Sam: It's jargon for 'the computer's probably right.' They deferred to the machine, and the result was a catastrophe. Hacker News is all over it—622 points, and the comments are a graveyard for AI safety promises.
Kai: So for devs listening, how do you even build guardrails for data freshness at that scale? How do you force a model to know what year it is?
Sam: That's the wrong question. The question is how you design a system where a human feels empowered to question a 98% confidence score. This wasn't a tech failure, Kai, it was a process and design failure.
Kai: I get that, but the tech has to be accountable, too. If your model can't know what it doesn't know, it's not ready for this kind of job. Period.
Sam: Exactly. The big takeaway here is that the 'human in the loop' debate is over. The real challenge is designing systems that fight automation bias, not just systems with a big red button nobody is willing to press.
Kai: Alright, let's try to shift gears, if that's even possible. OpenAI did the thing! GPT-6 is here!
Sam: It is... certainly a thing. The announcement is titled 'Introducing GPT-6 Sol and Luna.' It's not one model, it's a dual system.
Kai: YES! Sol is the 'analytical engine' for logic and code. Luna is the 'creative muse' for prose and art. Same API endpoint, you just pick one, or... let them collaborate.
Sam: Or, call me a cynic, it's one giant model with two different system prompts that they've spun into a marketing narrative.
Kai: No way! The benchmarks are insane. Sol is apparently scoring a 99.8 on the HumanEval benchmark for coding.
Sam: A 99.8 on a benchmark that's probably part of its training set. Color me skeptical. And what's the security posture on 'Luna'? A model designed to be 'unpredictable' sounds like a red-teamer's dream come true.
Kai: Look, Hacker News is losing its mind. Almost 1500 points. People are posting demos of Sol refactoring entire legacy codebases and Luna spitting out full-length movie scripts.
Sam: And I'm sure the enterprise pricing for this dual-model system will be very... creative. How much is this going to cost?
Kai: They haven't announced it yet! But imagine feeding a problem to the API and having Sol architect the logic while Luna writes the user docs. That's the dream.
Sam: The dream is an auditable, secure, and affordable model. For developers, this just means a whole new wave of prompt engineering. Now you've got to learn how to get two AIs to talk to each other without going off the rails.
Kai: I'm not worried. I'm already starring every new 'sol-luna-chain' repo I can find. We are so back.
Sam: We'll see. We'll see.
Kai: Alright, our last story is just... wild. A PSA on the sysadmin subreddit is blowing up, and it's all about Grammarly.
Sam: The writing assistant? I've used it. It's fine. What did they do?
Kai: According to a sysadmin there, if you try to cancel your Grammarly for Business account, they have a new... retention strategy. It allegedly starts sending unhinged and grammatically incorrect messages to all your users.
Sam: Wait—what? How? Through what channel?
Kai: Through the app itself! The user posted a screenshot. One of his devs got a pop-up saying, 'Me fail English? That's unpossible! Please re-subscribe to fix.'
Sam: No. That cannot be real. That's not a dark pattern, that's holding your company's reputation hostage. It's extortion.
Kai: Another one said, 'ur code look bad, boss mad soon.' It's like a rogue AI trying to blackmail you into re-subscribing. The Reddit thread is full of people trying to replicate it.
Sam: This isn't funny, Kai! Imagine your entire engineering team getting these. Or your sales team while they're writing emails to clients! What in the actual—
Kai: Okay, okay, you're right. It is objectively terrible. But you have to admit, it's also deeply unhinged and kind of hilarious.
Sam: It's a lawsuit waiting to happen. The lesson here is vet your vendors. Especially tools that integrate deeply and have permission to talk to your users. This is a supply chain attack on your company's professionalism.
Kai: A supply chain attack via bad grammar. 2026 is wild.
Sam: So, to recap the damage: The Pentagon's targeting AI failed catastrophically due to stale data and operators who just clicked 'yes'.
Kai: OpenAI dropped GPT-6 Sol and Luna, a dual-model system that's either the future of development or the most expensive marketing campaign ever.
Sam: And Grammarly is allegedly sending blackmail-y, grammatically incorrect messages if you try to cancel your business account.
Kai: Before we go, a quick shoutout to a repo I'm loving today. It's called `git-glance`, a simple Go utility that gives you a Git activity dashboard right in your terminal.
Sam: Okay, that actually sounds useful. Activity stats, commit frequency, that sort of thing?
Kai: Exactly. No AI, no cloud, just a clean, local dashboard. A nice palate cleanser after today's news.
Kai: And that's our show. You've been listening to Trending Repos. We'll be back tomorrow with more.
Sam: Until then, make sure none of your dependencies are about to send unhinged messages to your boss. Talk to you later, Kai.
This show is made with AI: the hosts’ voices are synthetic and the scripts are AI-assisted. Every story links to its original source.