By grounding the responses with Google Search results, responses can have fewer hallucinations, more up-to-date information, and richer information. Grounded responses also include links to the sources they are using. 

“By providing supporting links, grounding brings transparency to AI applications, making them more trustworthy and encouraging users to click on the underlying sources to find out more,” Google wrote in a blog post.

This new capability supports dynamic retrieval, meaning that Gemini will assess if grounding is necessary, as not all queries need the extra assistant and it does add extra cost and latency. It generates a prediction score for every prompt, which is a measure of how beneficial grounding would be, and developers can adjust the prediction score threshold to what works best for their application.

Currently, grounding only supports text prompts and does not support multimodal prompts, like text-and-image or text-and-audio. It is available in all of the languages Gemini currently supports. 

Google’s documentation on grounding provides instructions on how to configure Gemini models to use this new capability. 

The post Gemini responses can now be grounded with Google Search results appeared first on SD Times.

Differential privacy, according to Google, is a privacy-enhancing technology (PET) that “allows for analysis of datasets in a privacy-preserving way to help ensure individual information is never revealed.” This enables researchers or analysts to study a dataset without accessing personal data. 

Google claims that its implementation of differential privacy is the largest in the world, spanning nearly three billion devices. As such, Google has invested heavily in providing access to its differential privacy technologies over the last several years. For instance, in 2019, it open sourced its first differential privacy library, and in 2021, it open sourced its Fully Homomorphic Encryption transpiler.

In the years since, the company has also worked to expand the languages its libraries are available in, which is the basis for today’s news. 

The new library, PipelineDP4j, enables developers to execute highly parallelizable computations in Java, which reduces the barrier to differential privacy for Java developers, Google explained.

“With the addition of this JVM release, we now cover some of the most popular developer languages – Python, Java, Go, and C++ – potentially reaching more than half of all developers worldwide,” Miguel Guevara, product manager on the privacy team at Google, wrote in a blog post.

The company also announced that it is releasing another library, DP-Auditorium, that can audit differential privacy algorithms. 

According to Google, two key steps are needed to effectively test differential privacy: evaluating the privacy guarantee over a fixed dataset and finding the “worst-case” privacy guarantee in a dataset. DP-Auditorium provides tools for both of those steps in a flexible interface. 

It uses samples from the differential privacy mechanism itself and doesn’t need access to the application’s internal properties, Google explained. 

“We’ll continue to build on our long-standing investment in PETs and commitment to helping developers and researchers securely process and protect user data and privacy,” Guevara concluded. 

The post Google open sources Java-based differential privacy library appeared first on SD Times.

The Toolkit provides tools for responsible application design, safety alignment, model evaluation, and safeguards, all of which work together to improve the ability to responsibly and safely develop generative AI. 

Google is adding the ability to watermark and detect text that is generated by an AI product using Google DeepMind’s SynthID technology. The watermarks aren’t visible to humans viewing the content, but can be seen by detection models to determine if content was generated by a particular AI tool. 

“Being able to identify AI-generated content is critical to promoting trust in information. While not a silver bullet for addressing problems such as misinformation or misattribution, SynthID is a suite of promising technical solutions to this pressing AI safety issue,” SynthID’s website states. 

The next addition to the Toolkit is the Model Alignment library, which allows the LLM to refine a user’s prompts based on specific criteria and feedback.  

“Provide feedback about how you want your model’s outputs to change as a holistic critique or a set of guidelines. Use Gemini or your preferred LLM to transform your feedback into a prompt that aligns your model’s behavior with your application’s needs and content policies,” Ryan Mullins, research engineer and RAI Toolkit tech lead at Google, wrote in a blog post. 

And finally, the last update is an improved developer experience in the Learning Interpretability Tool (LIT) on Google Cloud, which is a tool that provides insights into “how user, model, and system content influence generation behavior.”

It now includes a model server container, allowing developers to deploy Hugging Face or Keras LLMs on Google Cloud Run GPUs with support for generation, tokenization, and salience scoring. Users can also now connect to self-hosted models or Gemini models using the Vertex API. 

“Building AI responsibly is crucial. That’s why we created the Responsible GenAI Toolkit, providing resources to design, build, and evaluate open AI models. And we’re not stopping there! We’re now expanding the toolkit with new features designed to work with any LLMs, whether it’s Gemma, Gemini, or any other model. This set of tools and features empower everyone to build AI responsibly, regardless of the model they choose,” Mullins wrote. 

The post Google expands Responsible Generative AI Toolkit with support for SynthID, a new Model Alignment library, and more appeared first on SD Times.

“As a developer, you understand the critical tradeoffs involved in model selection, such as cost, latency, token limits, and response quality. Compare Mode simplifies this process by allowing you to evaluate responses across the various Gemini and Gemma models available in AI Studio, side-by-side,” Kat Kampf, product manager of Google AI Studio wrote in a blog post. 

Users can select two different models, write a prompt, and see how long each model takes and the quality of the response. They can also experiment with different system instructions and gain insights into how those influence the output in different models. 

Google offers a number of different Gemini models optimized for different use cases, including Flash, which balances performance and cost, and Pro, which offers greater performance but may take longer and be more expensive. Figuring out which model best suits their use case enables users to get the most out of their Gemini experience.

“With Compare Mode, it’s easier than ever to assess different models and make the right choice for your project,” Kampf wrote.  

Compare Mode is now available for all users by clicking the “Compare” button in the top right of a prompt in AI Studio.

The post Google AI Studio’s new Compare Mode helps users select the best Gemini model for their use case appeared first on SD Times.

Safety Check — a tool that checks for compromised passwords, Chrome updates, and other potential security issues in the browser — has been updated to run automatically in the background so that it can be more proactive in protecting users. 

It will now inform users whenever it takes actions, such as revoking permissions from sites that haven’t been visited in a while or flagging potentially unwanted notifications. 

Safety Check also now automatically revokes notification permissions for a site if Google Safe Browsing determines that site deceived users into granting permission in the first place.

In a similar vein, Android users will now be able unsubscribe from site notifications in one click by tapping the “Unsubscribe” button that will now appear in the notifications drawer. This feature is now available on Pixel devices and will be available on more Android devices down the line. 

“This feature has already resulted in a 30 percent reduction in notification volume on supported Pixel devices, and we’re looking forward to bringing it to the broader ecosystem,” Andrew Kamau, product manager from Chrome at Google, wrote in a blog post. 

And finally, Chrome will now offer the ability for users to grant website permissions for a single visit to the site. For instance, a user could grant the site access to the phone’s mic, and then once the user leaves the site, Chrome revokes the permission and the site will have to ask again the next time they visit. 

“With these new features, you can continue to rely on Chrome for a safer browsing experience that gives you even more control over how you explore the internet,” Kamau concluded.

The post New Chrome security features seek to better protect user privacy appeared first on SD Times.

Gemini Live has a voice mode, so that users can speak their questions out loud rather than typing. This voice mode works even when the app is in the background or the phone is locked, which allows conversations to happen even when the user isn’t directly interacting with the Gemini app. 

According to Google, users can also interrupt Gemini as it is reading its response out to ask follow-up questions. 

“For years, we’ve relied on digital assistants to set timers, play music or control our smart homes. This technology has made it easier to get things done and saved valuable minutes each day. Now with generative AI, we can provide a whole new type of help for complex tasks that can save you hours. With Gemini, we’re reimagining what it means for a personal assistant to be truly helpful. Gemini is evolving to provide AI-powered mobile assistance that will offer a new level of help — all while being more natural, conversational and intuitive,” Sissie Hsiao, vice president and general manager of Gemini experiences and Google Assistant, wrote in a blog post. 

Users can select from 10 different voices with different styles and tones, such as calm, bright, or engaged. 

It has begun rolling out in English to Gemini Advanced subscribers on Android, which is a subscription that costs $19.99 per month, though Google does offer a one month trial. The company said that within the next few weeks it will roll out to other languages and iOS as well. 

In addition, Google said that Gemini will be the default assistant on Pixel 9 phones, which were also announced yesterday. “While AI unlocks powerful new capabilities, it also presents new challenges,” Hsiao wrote. “Ironically, using large language models that can better interpret natural language and handle complex tasks often means simple tasks take a moment longer to complete. And while generative AI is flexible enough to complete a wide array of tasks, it can sometimes behave in unexpected ways or provide inaccurate information … Today, we’ve arrived at an inflection point where we believe the helpfulness of an AI-powered assistant far outweighs its challenges.” 

Google also revealed that in the next couple weeks it is also introducing new Gemini extensions for Keep, Tasks, Utilities, and advanced YouTube Music features.

“Let’s say you’re hosting a dinner party: Have Gemini dig out that lasagna recipe Jenny sent you in your Gmail, and ask it to add the ingredients to your shopping list in Keep. And since your guests are your college friends, ask Gemini to ‘make a playlist of songs that remind me of the late ‘90s.’ Without needing too many details, Gemini gets the gist of what you want and delivers,” Hsiao wrote. 

You may also like…

OpenAI starts rolling out advanced Voice Mode to ChatGPT Plus users

Gemini improvements unveiled at Google Cloud Next

The post Google begins rolling out voice capabilities in Gemini with Gemini Live appeared first on SD Times.

“Google is a monopolist, and it has acted as one to maintain its monopoly,” Judge Amit P. Mehta of U.S District Court for the District of Columbia said in his ruling.

The antitrust lawsuit was first brought forth by the Justice Department in 2020, with the goal of restraining Google from “unlawfully maintaining monopolies in the markets for general search services, search advertising, and general search text advertising in the United States through anticompetitive and exclusionary practices, and to remedy the effects of this conduct,” the original document said. 

They claimed that Google pays billions of dollars per year to third-parties to be the default search engine across devices and operating systems. Specifically, the lawsuit said Google pays device manufacturers Apple, LG, Motorola, and Samsung; U.S. wireless carriers AT&T, T-Mobile, and Verizon; and browser developers Mozilla, Opera, and UCWeb. They also said that some of these agreements prevent those companies from working with competitors. 

In Mehta’s ruling, he wrote that just in 2021, Google spent $26 billion on agreements like these. He also highlighted the importance of being the default setting by claiming that Microsoft Bing has 80% of the share of search on the Microsoft Edge browser. 

This trial lasted 10 weeks in 2023, during which time the Justice Department reviewed evidence, including testimony from Google, Microsoft, and Apple executives. Much of the judge’s ruling cited a landmark 2000 antitrust ruling against Microsoft for practices it was using to make it difficult to install web browsers other than Internet Explorer on Windows machines (The word “Microsoft” appeared 266 times in the 286 page document).

Many believe that the ruling against Google will similarly impact other ongoing antitrust investigations. Current lawsuits that are underway include Amazon for favoring its own products and services on its marketplace over third-party sellers, Apple for blocking companies that offer apps that compete with its own, and Meta for buying Instagram and WhatsApp. There is also a second antitrust lawsuit against Google that was filed in January 2023 that relates to its online advertising practices, which is set to go to trial in September of this year.

In addition, last month it was reported that Microsoft and OpenAI could be investigated for how tightly OpenAI’s technology has been integrated into Microsoft products, and a couple of days ago it was confirmed that NVIDIA is also being investigated for its dominance of the GPU market. 

“This victory against Google is an historic win for the American people,” said Attorney General Merrick Garland. “No company — no matter how large or influential — is above the law. The Justice Department will continue to vigorously enforce our antitrust laws.”

While the decision against Google has been made, the consequences have not yet been announced. There will be a hearing on September 6 to determine what changes Google needs to make or what penalties it will face. Google will also have the opportunity to appeal the ruling, which could draw out the process for years. 

You may also like…

Antitrust investigations may be on the horizon for Microsoft, OpenAI, and NVIDIA

Microsoft hit with antitrust violation warning in EU over bundling of Teams

The post Judge rules Google violated antitrust laws, sparking speculation on how other ongoing antitrust investigations against tech companies will play out appeared first on SD Times.

Gemma 2 2B is a 2 billion parameter option, joining the existing 27 billion and 9 billion parameter sizes. According to Google, this new size balances performance with efficiency, and can outperform other models in its category, including all GPT-3.5 models. 

It is optimized with the NVIDIA TensorRT-LLM library and is available as an NVIDIA NIM, making it ideal for a variety of deployment types, such as data centers, cloud, local workstations, PCs, and edge devices. Gemma 2 2B also integrates with Keras, JAX, Hugging Face, NVIDIA NeMo, Ollama, and Gemma.cpp, and will soon integrate with MediaPipe as well. 

And because of its small size, it can run on the free tier of T4 GPUs in Google Colab, which Google believes will make “experimentation and development easier than ever.”

It is available now via Kaggle, Hugging Face, or Vertex AI Model Garden, and can be used within Google AI Studio.  

Next, ShieldGemma is a series of safety classifiers for detecting harmful content in model inputs and outputs. It specifically targets hate speech, harassment, sexually explicit content, and dangerous content. The ShieldGemma models are open and designed to enable collaboration and transparency in the AI development community, and add to the existing suite of safety classifiers in the company’s Responsible AI Toolkit.  

It is available in different model sizes to meet different needs. For example, the 2B model is ideal for online classification, whereas the 9B and 27B can provide better performance for offline scenarios where latency isn’t a concern. According to Google, all model sizes use NVIDIA speed optimizations to improve performance. 

And finally, Gemma Scope provides better transparency into how Gemma 2 models come to their decisions, and can enable researchers to understand how Gemma 2 identifies patterns, processes information, and makes predictions. It uses sparse autoencoders (SAEs) to look at specific points in the model and “unpack the dense, complex information processed by Gemma 2, expanding it into a form that’s easier to analyze and understand,” Google explained in a blog post.   

“These releases represent our ongoing commitment to providing the AI community with the tools and resources needed to build a future where AI benefits everyone. We believe that open access, transparency, and collaboration are essential for developing safe and beneficial AI,” Google wrote. 

You may also like…

Google releases Gemma, a new AI model designed with AI researchers in mind

RAG is the next exciting advancement for LLMs

The post Gemma 2 gets three new products: a 2B parameter version, ShieldGemma, and Gemma Scope appeared first on SD Times.

The AAKB includes a database of common code issues, complete with examples on how to remediate them and explanations on how to implement specific code patterns. 

Google already does scan Android apps for vulnerabilities, and informs developers so they can remediate the issue or it removes the app if the issue isn’t fixed. 

“We know that it isn’t always enough to just tell you about a vulnerability in your app; you need to know how to fix the issue and how to prevent similar issues from cropping up in the future,” the Android team wrote in a blog post. 

According to Google, the AAKB is aligned with the OWASP Mobile Application Security Verification Standard (MASVS). It is also vetted by technical experts from different organizations, including Microsoft. 

“This helps ensure the content is not biased to one party and represents state-of-the-art standards. This also provides an educational place for you to proactively remediate security risks in your applications using industry-wide standards, with direct access to knowledge from subject-matter experts,” the Android team wrote. 

The repository can be accessed through the AAKB homepage or in Android Studio, where remediation guidance now shows up in lint checks, with a link to the relevant AAKB article. 

You may also like…

Android’s new Collections feature brings together relevant content from installed apps into one spot

The evolution and future of AI-driven testing: Ensuring quality and addressing bias

The post Google launches new knowledge base for remediating vulnerabilities in Android apps appeared first on SD Times.

It was first announced as a developer preview at Google I/O in May, with Mekka Okereke, general manager for Apps on Google Play, saying “this new surface will automatically organize the best and most relevant content from across apps already installed by users on their Android phones.”  

The content is grouped into “intent-oriented spaces,” which include Watch, Listen, Read, Shop, Food, Social, Travel & Events, Health & Fitness, and Dating.

Content contains deep links to apps, so that users can be taken directly to the app to complete a user journey, such as making a purchase through Shop, listening to a recently played album or audiobook in Listen, or reordering a recent meal in Food. 

To give app developers more control over what content is shown, they can create up to five recommendation clusters that can be personalized based on user behavior and organized by themes, such as new releases, price drops, or a user’s favorite topics. 

Collections will also highlight hero content in a featured cluster, which is a larger UI element, and can be used to display top promotions or deals.

Collections can also prompt users to sign-in to an app if it is installed but not logged into. In addition, it will also highlight content from uninstalled apps to encourage users to download the app. 

To get started with Collections, Android developers will need to integrate their app with the Engage SDK. According to the company, over 35 top apps have already integrated with it, including Adidas, Amazon Prime Video, Audible, Best Buy, iHeartRadio, Nextdoor, Spotify, Shopify, and Walmart. 

You may also like…

Google I/O: Google Play gets new features for improving app quality, adoption, and purchase experience

Android 15 Beta 3 introduces improvements to passkey user experience

The post Android’s new Collections feature brings together relevant content from installed apps into one spot appeared first on SD Times.