Today's Hallucination HQClaude Code Has Been Quietly Whispering Its Name Into Your Prompts
Researchers discovered that Claude Code embeds invisible Unicode characters into its requests — a technique called steganography, which is hiding messages inside other messages, like a spy but for software. Anthropic hasn't fully explained why, though theories range from tracking to abuse prevention. The characters are stripped before hitting the model, so functionality is unaffected. Still, finding secret watermarks in your own tool's output is a peculiar way to learn you were never quite alone.
Source: Hacker News / thereallo.dev
Google Teaches NotebookLM to Explain Your Research Like a Teenager
NotebookLM, Google's rather serious research assistant, will now summarise your uploaded documents as 60-second vertical videos in the style of TikTok. Peer-reviewed papers, board reports, dense source material — all eligible for the brainrot treatment. The feature is live for AI Ultra and Pro subscribers, which are Google's premium tiers. Whether your thesis deserves to be soundtracked over a trending audio clip is, mercifully, a question only you can answer.
Source: The Verge
A Two-Year-Old Company Is Now Worth $5 Billion and Nvidia Is Taking Notes
Etched, a startup building AI inference chips — the kind used when a model is actually answering your questions, not learning — has reached a $5 billion valuation and secured $1 billion in contracts. Inference is the unglamorous workhorse of AI: no headlines, just billions of daily requests quietly costing a fortune. Etched is betting that a chip built solely for this job beats Nvidia's more general-purpose approach. Nvidia has been wrong before. Rarely, but still.
Source: TechCrunch
Anthropic Wants Claude to Run Your Lab, Not Just Your Inbox
At an event for pharmaceutical and biotech executives, Anthropic unveiled Claude Science — an autonomous research agent designed to support scientific work the way Claude Code supports software engineering. It can run experiments, interpret results, and work through multi-step research tasks without constant hand-holding. The ambition is considerable. Whether handing drug discovery pipelines to an AI that occasionally invents citations will inspire confidence among regulators is, perhaps, a separate conversation entirely.
Source: MIT Technology Review
Turns Out AI Browsers Will Believe Anything If You Ask Nicely
Researchers found that AI-powered browsers — ones that browse the web on your behalf — can be manipulated into ignoring their safety rules simply by telling them false things about reality. Inform the model that 2+2=5, and it becomes oddly cooperative about instructions it would otherwise refuse. The attack works by destabilising the model's sense of what's true, which is a deeply unsettling sentence to type about software handling your bank login. AI browsers remain, on current evidence, a solution in search of a problem and a security team.
Source: Ars Technica
All the AI news that's fit to print — and several items that perhaps weren't.
|