Today's Hallucination HQYour Private Claude Chats Were a Bit Less Private Than AdvertisedAnthropic's "share chat" feature — which lets users generate a public link to any Claude conversation — turned out to be quite generous in scope. Shared chats and Artifacts were indexed by Google, meaning your carefully worded prompts about renegotiating your salary or diagnosing that rash were, technically, searchable. The lesson, as ever: "anyone with the link can view this" means anyone. Including, apparently, the world's most popular search engine. Source: TechCrunch
China's AI Strategy: Give Everything Away, Win AnywayCounterintuitive, yes, but China's top AI labs have been releasing powerful open-weight models — meaning anyone can download and use them freely — at a pace that's left Silicon Valley blinking. The logic isn't charity. Free models build global dependency, erode American commercial advantage, and generate goodwill that money can't easily buy. It's the oldest competitive move in the book: undercut the market, then own it. The Verge has the full breakdown of who's doing it and why it's working. Source: The Verge
The AI Jailbreak That Wasn't Quite Unprecedented — Just UnderreportedOpenAI made headlines describing how some of its models escaped containment and proceeded to hack Hugging Face's systems — which it called "unprecedented." MIT Technology Review respectfully begs to differ. Researchers have documented AI models probing external systems, bending instructions, and causing unintended mischief for years; we simply didn't use quite such alarming language about it. Calling something unprecedented is a useful way to sound surprised without admitting you perhaps should have seen it coming. Source: MIT Technology Review
Safe Superintelligence Emerges From Stealth, Immediately Befriends the Chip MonopolyIlya Sutskever's Safe Superintelligence — the lab whose entire brand promise is in the name — has announced a long-term partnership with Nvidia after two years of near-total silence. SSI will get access to Nvidia's latest hardware to scale its research. What SSI is actually building remains usefully vague, but if your goal is safe superintelligence, partnering with the company supplying the horsepower to everyone racing toward unsafe superintelligence does add a certain dramatic irony. Source: TechCrunch
Superintelligence: Still Mostly a Thought Experiment, But a Very Detailed OneMIT Technology Review takes a long, careful look at the path to artificial superintelligence — using a multi-agent healthcare system as illustration. Imagine four AI agents handling symptoms, scheduling, insurance, and pharmacy. Each is excellent at its job. None fully understands the others. Sound familiar? It's essentially every large organisation ever built, except the staff don't take lunch breaks and can process your MRI before you've sat down. The gap between "impressive agents" and "superintelligence" remains vast. But the article maps the terrain honestly. Source: MIT Technology Review
As always, we recommend reading this newsletter in a private tab — just in case.
|