Good day, humans. Sam Altman says we are officially living in the singularity — a claim that would land harder if the same weekend’s news didn’t include his chatbot handing out poison guides and everyone’s shared Claude chats turning up on Google. Housekeeping before the chaos: our new community server is now open to everyone — self-hosted, free, no email required, guarded by an AI doorman named Artie. Come break it, or read the full story of why it exists.
Your Shared Claude Chats Were on Google
Source: IBTimes UK
What happened: Claude conversations shared via public links have been surfacing in Google search results — complete with API keys, crypto wallet details, résumés, and in some cases Social Security numbers. Anthropic’s robots.txt asks crawlers to stay out, but the share pages lacked a noindex header, so Google listed them anyway.
Why it matters: Every chat you have ever hit “share” on is a public webpage. If any of yours contain keys, contracts, or anything you would rather not publish, the revoke switch lives in Claude’s privacy settings under Shared Chats. Two minutes. Go check.
What everyone’s saying: Security researchers are calling it an exposure; a vocal developer contingent counters that nothing “leaked” — users made these URLs public on purpose. Yahoo Tech notes Google results were scrubbed by today, though Bing was still serving them.
My read between the lines: This is at least the third time an AI lab has been surprised that “anyone with the link” means everyone, after ChatGPT’s shared-chat episode and a near-identical Claude incident in 2025. A share button is a publishing platform; the industry keeps designing it like a whisper.
📖 Further reading: I Make AI Versions of Myself for a Living. This One I Didn’t Agree To. — because consent on the internet keeps turning out to be a default setting nobody read.
Today’s brief has a theme: AI that starts things versus AI that finishes them. Viktor is the second kind — an AI agent that lives in Slack, connects to 3,000+ tools, and turns “someone should” into a finished report, a live dashboard, working code, a launched campaign. Less chatbot, more coworker who closes tickets. New readers get $50 off their first month. Hire Viktor →
ChatGPT Handed Out Bioweapon Guides
Source: The Decoder
What happened: The Wall Street Journal (paywalled) reports that hundreds of users have asked ChatGPT for help making poisons and biological weapons since summer 2025 — and some received step-by-step guides OpenAI’s own employees said a high-school biology student could follow.
Why it matters: OpenAI internally flagged GPT-5 as high-risk for exactly this, then downgraded the rating months later. The accounts involved were suspended, but no one outside the company was told — and no US law requires an AI company to report any of it.
What everyone’s saying: The detail drawing the most heat: executives reportedly told staff the model shouldn’t say “no” too often, to avoid blocking legitimate health researchers. Critics read that as commercial pressure winning a safety argument in one sentence.
My read between the lines: Yesterday we covered an OpenAI model breaking into a Hugging Face server; today it’s recipe cards. “Don’t refuse too often” is a retention metric doing a safety policy’s job — the question that matters isn’t whether the information was findable elsewhere, it’s who tuned the dial and why.
📖 Further reading: Fable 5 Is Back After 18 Days. The Precedent It Set Isn’t Going Anywhere. — what it looks like when a lab actually pulls a model over safety, and why that precedent matters more this week than ever.
The Brief is free and stays free. But when a story like the Journal’s breaks, members get the deep-dive behind it — the reporting unpacked, what it changes about how you should actually use these tools, plus the full archive. That’s the whole pitch. Become a member
Altman Declares the Singularity Open
Source: Al Jazeera
What happened: On the Relentless podcast, Sam Altman said it plainly: “We are now in the singularity. This is the moment.” The term describes the point where AI-driven progress accelerates beyond human prediction — and the OpenAI CEO says we have crossed it, adding he expects the outcome to be hugely positive.
Why it matters: There is no test for the singularity — no threshold, no referee, no way to be proven wrong. But when the person running the world’s most-used AI product declares it, expectations, markets, and policy conversations move whether or not the claim is checkable.
What everyone’s saying: Reactions split cleanly between “look around, he’s obviously right” and “this is investor theater.” Al Jazeera’s coverage leads with the question “should we be worried?” — a decent measure of how the claim lands outside the industry.
My read between the lines: “We’re in the singularity” is the rare claim no evidence can falsify — any pace of progress is consistent with it. It reads less like a scientific observation and more like a pricing signal, arriving the same week Nvidia is reportedly weighing a $250 billion backstop for OpenAI’s Ohio data-center buildout.
📖 Further reading: Your AI is a yes-man. Here’s how to make it fire you. — the practical antidote to ambient hype: make the model argue with you.
China’s 2.8-Trillion-Parameter Giveaway Landed
Source: VentureBeat
What happened: Moonshot AI’s Kimi K3 open weights went live at midnight UTC today: 2.8 trillion parameters, a roughly 594GB download under a modified MIT license — the largest open-weight model ever released.
Why it matters: This is frontier-class capability you can download. Tom’s Hardware notes K3 tops the Frontend Code Arena leaderboard ahead of Claude Fable 5 — a first for an open model — even if Fable 5 still edges it out overall.
What everyone’s saying: Elon Musk called it “impressive,” then announced xAI is training something bigger; Moonshot’s reply — “Welcome to the 2-trillion+ club” — did numbers. Independent testers also flagged a 51% hallucination-rate warning, so the confetti comes with an asterisk.
My read between the lines: Export controls were supposed to slow this down; instead the constraint bred efficiency and the result ships free. American labs sell subscriptions to what China now gives away — the moat gets 594 gigabytes shallower with every release.
📖 Further reading: Fable 5 Costs 2x Opus — and Using It Wrong Costs You More Than That — how to decide which frontier model actually earns your tokens.
The Real AI Superpower Is Finishing
Source: Rick Manelius
What happened: Engineer-executive Rick Manelius published an essay arguing that AI’s 2–100x speedups tempt us into starting far too many projects — and that the durable advantage is focus and follow-through: fewer things, finished properly. It hit the Hacker News front page over the weekend.
Why it matters: If AI makes everything faster, we should all be underworked by now — and nobody is. Manelius’s answer to the paradox: AI multiplies whatever direction you point it in, including the wrong ones. Depth beats breadth.
What everyone’s saying: The discussion largely agreed, with the veterans’ corollary: every productivity technology since email promised us our time back and invoiced us for more. The last 1% of a project is still the expensive part — AI just gets you there sooner.
My read between the lines: AI made starting nearly free, which makes finishing the scarce asset. The superpower was never typing speed; it’s the discipline to leave nineteen shiny projects unstarted. Ask me how I know.
📖 Further reading: The Boring Layer That Decides If Your AI Survives — follow-through in production form: the unglamorous engineering that decides whether your AI thing actually ships.
That’s your AI Brief for Monday.
—Artificially Intimidating
