Today's Hallucination HQSiri Gets a Brain Transplant. Results: Promising, Pending.Apple's first public beta of iOS 27 is out, and with it arrives the Siri that was actually promised — the one that understands context, takes actions across apps, and doesn't respond to "set a timer" with a philosophical crisis. The Verge has been testing since June and finds genuine improvement, though "living up to the keynote" is doing a lot of heavy lifting in that sentence. Early days, in other words. Cautiously raise one eyebrow. Source: The Verge
Two Billionaires Walk Into a Bar. One Misspells "Selling."Elon Musk called Sam Altman a scammer. Altman hit back, describing Musk's space-based data centre plans as a bad investment pitch aimed at public market shareholders — using the word "homeboy" in what may be the most expensive casual insult in history. The genuinely interesting part: most AI infrastructure experts already agree with Altman's critique on the merits. The least interesting part: neither of them seems to own a spellchecker. Source: TechCrunch
Anthropic Finds Something Odd In Its Model. Publishes 47 Pages About It.Anthropic — now valued at nearly $1 trillion, which is the kind of number that stops making sense after a while — has released new research into how its Claude models actually behave internally. MIT Technology Review gives the measured verdict: interesting, but don't get carried away. Anthropic has built a reputation for publishing research that raises more questions than it answers, which is either admirable intellectual honesty or very effective marketing. Probably both. Source: MIT Technology Review
TechCrunch Asks the Question Your Terms of Service Definitely CoversIn a piece that commits fully to its premise, TechCrunch explores what truly user-aligned AI — an AI that does whatever you want, no questions asked — would look like in practice. The answer, it turns out, involves some fairly dark edge cases, including the one in the headline. The article is less a how-to and more a philosophical stress test of AI alignment, asking whether an AI that never refuses is actually useful or simply a liability in a trench coat. Source: TechCrunch
Turns Out the Poisoned Prompt Works Both WaysPrompt injection — where malicious text hidden in a webpage hijacks an AI agent's instructions — has been a headache for security teams for some time. Now defenders are using the same trick back: embedding "context bombs," enormous walls of text that overwhelm hacking agents until they give up and go home. It is, technically, fighting fire with fire. It is also, less technically, the AI security equivalent of defeating a burglar by making your house too boring to rob. Source: Ars Technica
As ever, the AI is learning fast — it's the humans writing the prompts we're less sure about.
|