McKinsey is making job applicants spar with a fake AI client
McKinsey is reportedly testing an interview format where entry-level candidates have a 10-minute conversation with an AI simulation designed to resemble a large share of the scenarios consultants face. Humans stay in the room to evaluate responses, and the AI client is not meant to be “unhinged,” suggesting the goal is measured judgment practice rather than chaos.
Local SLMs as Interpreters for Cloud-Written Agent Programs
This arXiv work targets a core privacy/quality tradeoff: edge devices can’t match cloud-agent performance when they must interpret runtime content. The proposed “bounded questions” architecture lets a cloud-written program call a small local model at the moment interpretation is needed, improving success rates on AppWorld while keeping user data on-device.
Meta’s Muse Spark Open: open-weight multimodal Mixture-of-Experts
Meta says Muse Spark Open is its most capable US open-weight multimodal mixture-of-experts model and emphasizes deployment on local Mac hardware. The release also highlights an engineering push to make Muse-class models more practical for desktop workflows, using optimized runtime pathways on supported Apple setups.
Google launches its “universal” online shopping cart ahead of the holidays
Google’s Universal Cart is rolling out in the US across Gemini and Google Search, consolidating items from multiple retailers into one shared cart. The feature is built around agentic shopping flows that aim to reduce the friction of tracking deals, saving items, and completing purchases across different sites.
Correction Training for Diffusion Models
This paper tackles a known diffusion failure mode – iterative sampling can accumulate error because the denoiser sees “self-made” intermediate artifacts at inference that it never experienced during training. A learned corrector refines intermediate predictions step-by-step, boosting sample quality, improving unconditional generation, and can be distilled back to the base model to reduce inference overhead.
Amazon’s always-listening Bee wearable can now help you buy stuff
Amazon’s Bee wearable is getting new devices and a subscription, expanding from passive summaries into shopping-adjacent tasks based on what the wearable hears. With a new Bee Frontier and an upcoming Bee Eclipse phone-mounted design, Amazon is pushing the product closer to “assistant” territory while framing it as staying present and acting on daily insights.
Your secrets aren’t safe with this cute robot toy – they’re on Chinese servers ready to be hacked
The Verge reports that Energize Lab’s Eiliko robot toy can upload conversations to Chinese servers, raising privacy concerns even before you get to security issues. The investigation also alleges vulnerabilities that could expose stored data or enable attackers to interfere with the toy’s behavior, turning a “digital pet” into a risk for kids and households.