Your Menu Was AI-Written, And Your Tongue Knew

Today on Briefly AI — OpenAI says its newest model has quietly tipped us into the age of AGI. ChatGPT, Grok and Claude all went down at once on Thursday, and nobody's explaining why. And it turns out diners can actually taste when a menu's been written by a chatbot.

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. A machine reading you the headlines about machines — which sounds like a gimmick, right up until you notice it hasn't missed a morning yet.

Okay — let's get into it.

Someone managed to take down ChatGPT, Grok and Claude at almost exactly the same moment this week — which means either the entire AI industry is plugged into the same slightly dodgy socket, or something stranger's going on. We'll get to that. First, a company that would very much like you to know it thinks it's just built something historic.

Right, so — OpenAI has released GPT-6 Astra, and the language coming out of the company is, let's say, elevated. As The Verge reported, OpenAI is calling it "a generational leap in capability" across cybersecurity, professional work, software engineering, science, and computer use — that last one meaning the model can actually operate a computer, clicking round a screen and using software the way you or I would, rather than just answering questions in a text box. Wired went further, reporting that OpenAI's own leadership thinks Astra might mark the start of what they're calling the AGI era — AGI being the industry's term for AI that can do more or less any cognitive task a human can, rather than being good at one narrow thing.

Here's the bit that actually matters more than the branding. Astra is the first OpenAI model to trigger what the company calls its "critical cybersecurity capability threshold" — a line in its own safety framework that, once crossed, requires extra safeguards because the model is now judged capable enough to meaningfully assist in offensive hacking, not just defensive research. OpenAI says it's built in protections to stop that capability being misused. That's worth sitting with for a second, because it's OpenAI marking its own homework and telling us the answer is fine — the same tension that comes up every time a lab ships something newly dangerous and newly useful in the same release.

So, plain state of play: Astra is out, OpenAI is calling it a leap forward in coding, science and general capability, and it's the first model the company itself has flagged as crossing into serious cybersecurity risk territory. Whether "AGI era" turns out to be marketing or genuinely the start of something, nobody outside OpenAI can verify yet — but the cybersecurity threshold is a concrete, checkable claim, and one worth watching for independent testing.

Meanwhile, in a rather less triumphant story about the same industry — on Thursday, at around eleven in the morning Eastern time, ChatGPT started throwing error messages. OpenAI's status page reported, in its words, "elevated errors across ChatGPT and Codex" — Codex being OpenAI's coding assistant, which a genuinely enormous number of developers now rely on to actually get work done. And it wasn't just OpenAI. Around the same window, Grok — that's Elon Musk's xAI chatbot — and Claude, Anthropic's assistant, both started having issues too. Three separate companies, three separate models, built on three theoretically separate bits of infrastructure, all wobbling at once.

The Verge confirmed all three are back online. But Wired's report is the interesting one, because its entire premise is that nobody will say what happened. Not OpenAI, not Anthropic, not xAI. No shared cloud provider named, no root cause published, no "here's what broke and here's what we're doing about it." Just three outages, at the same time, and silence.

Why should you care, beyond the mild inconvenience? Because Codex, Claude and Grok aren't toys anymore — plenty of people now have actual paid work routed through them, coding agents mid-task, customer queries mid-flight. When the tools go down together and the companies running them won't explain why, that's not just an outage, it's a reminder of how little visibility any of us actually have into the infrastructure underneath these things we've started depending on. As of now, all three are working again. Why they broke at the same moment remains, genuinely, a mystery — one of the few times this week that word's actually earned.

Now, this next one's a bit more fun, and rather more human. TechCrunch has a piece out about restaurant owners using generative AI to redo their menus — the photos, the descriptions, sometimes the whole layout — as a quick, cheap alternative to hiring a photographer or a copywriter. Sensible enough shortcut, on paper.

Except, according to the reporting, diners can tell. Not because the food photos are obviously fake or broken — AI image tools have got well past the six-fingered-hand stage — but because everything produced this way tends to converge on the same look and the same language. The glossy, suspiciously perfect plating. The dish descriptions that lean on the same handful of words — chef-driven, locally-sourced, umami-rich — regardless of what restaurant, what cuisine, or what city you're in. TechCrunch's framing is that people don't need to consciously spot "that's AI" to feel something's off; they register it more viscerally, the way you'd sense a photo's been filtered within an inch of its life without being able to say exactly why.

It's a small story, but it's a sharp one, because it's adoption meeting the limits of the tool in a way most AI coverage skips straight past. The technology works. The output is technically fine. And it still manages to make a restaurant feel less trustworthy, not more appealing — which is precisely the opposite of what the menu was for in the first place.

That's the trap with a lot of these AI shortcuts: cheap and fast were never really the problem businesses needed solving. Feeling like somewhere real, run by someone who actually cares what's on your plate — that's harder to fake than a photo.

That's Briefly AI for today. Short by design, sourced properly, and back again tomorrow — same time, same arrangement.