on-device processing

Winsage
May 29, 2025
Windows collects telemetry data, which includes hardware specifications, installed software, and system events, but does not reveal personally identifiable information. This data is encrypted and sent via HTTPS during significant events like system crashes. Users can manage optional data collection settings during initial setup or in the Privacy section of the Settings app, with enterprise versions offering more control. Mandatory data collection in Windows is more extensive than in macOS, while Linux has minimal data collection due to its decentralized structure. Users can disable some data collection, but some information will always be gathered unless advanced measures are taken.
AppWizard
May 17, 2025
Google announced on May 14, 2025, the expansion of its ML Kit offerings with new GenAI APIs powered by the Gemini Nano model, aimed at providing Android developers with access to advanced AI features while ensuring user privacy through on-device processing. The ML Kit GenAI APIs include four key capabilities: summarization, proofreading, rewriting, and image description. These APIs process data locally, keeping user data on the device, ensuring consistent functionality without internet connectivity, and avoiding recurring server costs. This development enhances the accessibility of sophisticated AI capabilities for Android developers and aligns with consumer privacy concerns. The architecture allows for future expansion of capabilities as Gemini models evolve.
AppWizard
February 27, 2025
Google's SafetyCore app was launched in October 2023 as part of system updates for Android devices, designed to classify content like spam and explicit material through on-device image analysis. It operates locally using machine learning, avoiding cloud uploads, and is promoted as a privacy-enhancing feature. However, it requires 2GB of storage and operates in the background without prior user disclosure, leading to criticism regarding its transparency and potential privacy implications. Users discovered its existence through online discussions, with some labeling it as "spyware." Although Google claims that SafetyCore only activates upon app requests and does not report findings back to the company, the closed-source nature of the app raises concerns about independent audits and user trust. Advocates for privacy emphasize the need for transparency in technology, especially regarding features that access personal data. Users can disable SafetyCore through the Settings menu, but many remain unaware of this option. The situation highlights a broader issue of balancing innovation with user consent in the tech industry.
Winsage
December 26, 2024
Copilot+ PCs are the first personal computers to run Small Language Models (SLM) directly on-device, allowing for quicker interactions without relying on the cloud. Microsoft has introduced the AI Dev Gallery, which offers over 25 samples for developers to integrate on-device AI features into applications on Windows 10 and 11. The gallery requires building the project in Visual Studio, needing at least 20GB of storage and a multi-core CPU. A GPU with 8GB VRAM is recommended for heavier models but not mandatory for lighter applications. The app has two operational modes: Sample and Models. Testing models for image generation typically requires around 5GB of bandwidth, while a smaller image upscaling model under 100MB was successfully tested, completing the process in under 30 seconds with peak RAM usage of 1GB. The resulting image resolution was 9272x4900, but clarity issues were noted, especially with text. The application lacks features for previewing images in larger formats or downloading outputs directly. A model named Detect Human Pose was able to identify positions within images, including desktop screenshots. Substantial storage and robust CPUs are necessary for effective model accommodation, and the practicality of downloading large models for niche use cases is questioned.
Winsage
October 30, 2024
In 2024, many Windows laptops are equipped with dedicated Neural Processing Units (NPUs) to enhance computing capabilities by efficiently handling AI workloads. Laptops featuring Intel Core Ultra processors, Qualcomm Snapdragon processors, or AMD 7040 series or later utilize NPUs. Windows Studio Effects allows users to enable features like background blur and eye contact during video calls, utilizing the NPU for energy-efficient processing. Amuse 2.1 is a generative AI application that performs text-to-image generation on-device using an NPU, although it currently lacks support for Intel's NPUs. Adobe Premiere Pro incorporates AI-driven features leveraging Ryzen AI hardware, and GIMP offers text-to-image generation through the OpenVINO plugin for Intel Core Ultra CPUs. Users with Snapdragon-powered laptops have access to features like live captions and image enhancements, with more AI capabilities expected in Windows 11 by 2025.
AppWizard
October 2, 2024
Google has opened access to its Gemini Nano model for all Android app developers, allowing them to integrate advanced AI functionalities into their applications. The AI Edge SDK, previously in closed preview, is now available, particularly for developers working with the Google Pixel 9. This toolkit facilitates the integration of features such as text rephrasing, smart replies, proofreading, and summarization. Gemini Nano operates on devices, enabling tasks like text summarization and image description without a network connection, ensuring user data privacy. Developers can explore text-to-text prompts and customize how Gemini Nano processes information for their apps.
AppWizard
August 27, 2024
Google is enhancing the Files app to include AI-driven document summaries, allowing users to navigate digital files more efficiently. Recent findings from an APK teardown indicate that the feature is in active development, with references to “AI-generated summaries” discovered in the beta version 1.4678.667290778.0. These summaries will describe file contents succinctly and will be generated directly on the user’s device to maintain data privacy. Users can disable the feature in the app’s settings, but should verify critical information as the AI may not always be accurate. There is speculation that the Gemini Nano technology may be used for on-device processing, potentially limiting the feature's availability to Pixel devices and a few others that support it.
Search