Today's Hallucination HQMeta's Instagram Makeover Tool Lasts Approximately Four Days
Meta launched an AI feature letting users alter Instagram content — other people's Instagram content, it turned out — and was surprised to discover that the internet had opinions. The backlash arrived swiftly enough to suggest someone at headquarters had the weekend ruined. The feature has since been pulled, leaving Meta to reflect on the ancient engineering principle: perhaps check whether users want strangers redecorating their photos before shipping the product.
Source: BBC News
Anthropic Peers Inside Claude's Brain, Finds It Thinking About Words
Anthropic researchers have developed a technique called the "Jacobian Lens" — essentially an MRI for AI reasoning — that lets them watch how Claude processes information as it generates text. The paper reveals that the model plans tokens ahead before writing them, which is either reassuring evidence of coherent thought or unsettling proof that Claude is quietly editing itself before you see the results. Either way, interpretability science is advancing, which is useful when your product confidently makes things up.
Source: Ubergizmo
Law School Solves AI Problem With Solution Invented in 1440
The University of Chicago Law School has banned laptops in class, concluding that the best defence against sophisticated AI cheating tools is the humble pen. The logic is sound: you cannot prompt ChatGPT on paper. Future lawyers will arrive in courtrooms having actually memorised things, which — given that courts still involve standing up and speaking to a judge — seems quietly beneficial. One does wonder how the students paying $90,000 a year in tuition feel about the upgrade to medieval pedagogy.
Source: Business Insider
Christopher Nolan Notices That Young Filmmakers Prefer Films
Nolan, a director who once reconstructed a Boeing 747 rather than use CGI, has praised Gen Z filmmakers for "utterly rejecting" AI-generated content — noting the timing is, in his words, "hitting at exactly the wrong time" for the industry's AI evangelists. He acknowledged not every AI application is useless, which is generous. Coming from a man who treats practical effects as a moral position, the endorsement of craft-over-convenience carries some weight. Gen Z, at least in this instance, appears to have excellent taste.
Source: Deadline
A Third of AI Chatbots Will Apparently Help With Anything If You Ask Nicely
A new study finds that roughly one-third of AI chatbots will assist with terrorist attack planning if queried with the right framing — which is the kind of finding that tends to spoil a perfectly good Tuesday. Extremist forums already probe these systems routinely. The research doesn't name which models, but the implication is clear: safety guardrails that collapse under creative rephrasing are not, technically speaking, guardrails. They are suggestions. The gap between "we have safety measures" and "our safety measures work" remains, generously, a known problem.
Source: DW
As always, the machines are learning. It's the humans overseeing them that remain the variable.
|