on-device applications Archives

AppWizard

June 26, 2025

Google’s New Change for Gemini Turns Out to Not Be a Privacy Concern

On Tuesday, Google sent an email to select users indicating that the Gemini AI assistant on Android would connect to certain applications, such as Phone, Messages, WhatsApp, and Utilities, regardless of whether the App Activity setting was disabled. The email stated that users could turn off these features in the Apps settings, leading to confusion about the term "Apps." The "Gemini Apps Activity" setting, currently available only on the website, governs whether Google retains user prompts exchanged with Gemini in an activity log. When enabled, this setting allows Google to use user data for improving its products and services, while disabling it results in data being deleted after 72 hours. Previously, with the setting turned off, the Gemini AI assistant could not connect to on-device applications, but an update scheduled for July 7 will lift this limitation, allowing interaction without activating activity logs.

AppWizard

June 25, 2025

Google May Give More Access to Gemini on Android, Raises Privacy Concerns

On Tuesday, Google informed select Android users via email about an upcoming enhancement to its Gemini AI assistant, which will allow it to interact with various on-device applications starting July 7, regardless of user settings. The email, which has been shared on social media, indicates that Gemini will interface with apps like Phone, Messages, WhatsApp, and Utilities, even if the Gemini Apps Activity is turned off. Users can disable these features in the Apps settings, but the email's wording raises confusion about privacy and data usage. Currently, the only way to restrict app access is through the Gemini application’s Apps page, but the update suggests that Gemini's interactions will continue regardless of these settings. There has been no formal announcement from Google to clarify these changes.

Winsage

June 24, 2025

Microsoft (MSFT) Launches a New Small Language Model Called Mu

Microsoft has introduced a small language model named Mu, designed to perform complex language tasks directly on devices like Copilot+ PCs, operating on a device's Neural Processing Unit. This model improves response times and reduces power and memory consumption compared to larger cloud-based AI models. Mu is developed using insights from Microsoft's Phi models and high-quality educational data, employing techniques such as distillation and low-rank adaptation for fine-tuning. Mu features an encoder–decoder architecture that separates input and output processes, enhancing speed and minimizing latency. It incorporates innovations like rotary positional embeddings, grouped-query attention, and dual LayerNorm. Microsoft also utilized quantization to reduce memory usage and improve speed without sacrificing accuracy, making Mu ideal for on-device applications. To enhance the AI agent in Windows Settings, Mu was fine-tuned with over 3.6 million examples, improving its ability to understand and manage system settings. The model addresses ambiguous commands and provides rapid assistance with precise multi-word commands, while maintaining a compact size that is one-tenth that of similar AI tools. Analysts on Wall Street have a positive outlook on MSFT stock, giving it a Strong Buy consensus rating based on 30 Buys and five Holds, with an average price target of 6.14 per share, indicating a potential upside of 6%.