Meta AI will stop suggesting you investigate your own kids
Meta's assistant was caught turning a family's public posts into a dossier — and suggesting the questions itself. Also today: Altman says OpenAI would slow down if rivals do, and a 210M text-to-image model publishes its training recipe, measurements included.
Meta is changing the prompts its AI suggests after a viral video showed Meta AI nudging a mother to dig up personal details about her own young daughters. Travel creator Kalie Robins filmed the assistant proposing "Who's the child passenger?" about a girl in her own video, then "What are Kalie's daughter's ages?" — and answering with an age timeline assembled from family birthday posts and a grandmother's newborn announcement, plus her daughters' birth weight, favorite beaches and hiking trails, and a multi-paragraph attempt at "pinpointing Kalie Robins' location." Meta confirmed the search feature draws on a user's own content and public web information, said responses are based on information the user already has access to, and called the prompting a miss: the feature "never should have prompted the individual with questions like that," and it has been fixed. The uncomfortable part is not one bad prompt — suggestions are the product, and this one proactively pointed a parent toward doxxing her own children. Robins' accounts were public, but she had a few hundred followers, not an influencer's audience; the assistant assembled the dossier in minutes because years of mundane posts are the corpus. This is the same assistant Meta is pushing deeper into personal context — Meta ships a Mac app for Meta AI, hooks it into Instagram, Facebook, and Google Workspace — so "the user already had access" is a thin answer when the product's job is to propose what to ask next.
Sam Altman told OpenAI staff the company is open to slowing frontier development — if rivals slow too. At a companywide meeting this week, Altman said OpenAI might match its pace with a handful of peer labs while acknowledging not every company would join, according to Bloomberg; OpenAI declined to comment. Bloomberg also reported OpenAI has already halted certain internal training runs and scaled back parts of model development over safety concerns, and chief scientist Jakub Pachocki has called for voluntary slowdowns "until shared safety bars are established." We covered the antitrust angle this morning — OpenAI asked Congress if an industry-wide AI slowdown is even legal. The new fact is posture: the CEO of the field's most valuable startup putting the brakes on the table, with paused training runs as the first evidence the talk costs real compute.
A solo researcher trained a 210M-parameter text-to-image diffusion transformer from scratch — one GPU, 3.5 days, 4.2 million images — and published the measurements. Ivan Mikhnenkov's tinydit write-up is worth reading for three findings. Learned null attention slots act as the model's scratchpad, absorbing about 90% of cross-attention mass at mid-noise while the usual EOS sink drops to 4%. The flow-matching loss fell only 0.805 to 0.754 while held-out FID improved 33.7 to 27.0 — a health signal, not a quality signal, so don't read a loss curve as progress. And a training-time timestep shift buys more than doubling the sampling step count. The quiet lesson is curation: five aspect-ratio buckets instead of center-crops (a near-square filter would have discarded 94% of COCO), long and short captions sampled at 50/40 with 10% empty for guidance, synthetic data capped at a quarter of the mix. Weights, a browser demo and the repo are public; the failure grid — readable text, close faces, crowds, counts above three — doubles as a to-do list for the planned RL phase.
Would you trust an assistant embedded in your own photo archive to decide what to suggest about your family? Tell us in the comments.
Sources: tinydit (GitHub)
