Google взорвала рынок ИИ: новые инновации в шопинге и робототехнике от NotebookLM до SIMA 2 Translation: Google Blasts the AI Market: New Innovations in Shopping and Robotics from NotebookLM to SIMA 2

Google has unveiled a variety of new solutions in the field of agent-based AI. Among these are a deep research mode in NotebookLM, a «brain for robots» called SIMA 2, and shopping tools.

The company has upgraded its note-taking AI assistant, NotebookLM. New features include a tool that simplifies complex research and supports additional file types.

The service launched Deep Research, a solution for automating online searches. Google claims it functions like an independent researcher, capable of preparing detailed reports or suggesting relevant articles, academic papers, and websites.

Deep Research receives a query, creates a research plan, and browses web resources. Within minutes, it provides a source-based report that can be added directly to the notebook.

The mode runs in the background, allowing users to accomplish other tasks simultaneously.

The tool is accessible via search, with the option to choose between a detailed Deep Research mode or a quicker Fast Research mode.

Additionally, NotebookLM now supports Google Sheets, URL files from Drive, PDFs from Google Drive, and Microsoft Word documents.

These updates are expected to roll out within a week.

NotebookLM is Google’s AI assistant for note-taking, research, and document handling. It allows users to upload various materials—PDFs, articles, spreadsheets, images, links, legal documents, lectures—and build a structured knowledge base.

The service was launched in 2023 and has gradually expanded its capabilities through artificial intelligence. By early 2025, it introduced a Video Overviews feature, which transforms complex multimedia content into understandable visual presentations.

In May, NotebookLM became available on Android and iOS.

Google is also advancing its work on the “brain” for robotics.

DeepMind introduced SIMA 2, a new generation of a universal AI agent that «goes beyond simple instruction execution,» beginning to comprehend and interact with its environment.

The first version of SIMA was trained on hundreds of hours of gameplay videos to learn how to play various 3D games like a human. It was released in March 2024 and could handle basic commands in different virtual worlds, but only managed complex tasks 31% of the time.

SIMA 2 leverages the language and analytical capabilities of Gemini and is based on the flash-lite version 2.5, with accuracy improved to 65%.

«SIMA 2 represents a significant leap forward compared to SIMA 1. It is a more versatile agent, capable of handling complicated tasks in novel environments,» said DeepMind senior researcher Joe Marino during a press briefing.

The assistant can self-improve—enhancing its skills based on personal experience, which is a step towards more adaptable robots and systems, Marino noted.

Researchers from Google’s AI division emphasized that work on so-called «embodied agents» is crucial for developing general intelligence. Such assistants must interact with both physical and virtual worlds through a body—like a human or robot.

A disembodied assistant can manage calendars, take notes, or execute code, Marino explained.

Senior DeepMind researcher with a neurobiology background, Jane Van, highlighted that SIMA 2 transcends typical gaming behavior.

«We require it to truly understand what is happening, what is being asked of it, and to respond reasonably and meaningfully. This is quite challenging,» she said.

The integration of Gemini has allowed SIMA 2 to double the performance of its predecessor. The model combines advanced language and analytical AI capabilities with «embodied» interaction skills acquired through training.

Marino demonstrated SIMA 2 in the game No Man’s Sky, where the agent described its environment—a rocky planetary surface—and outlined its next steps using Gemini for internal reasoning.

In another game, the assistant was instructed to approach a house the color of a ripe tomato. The AI showcased its analytical process: «It’s red, so I should go towards the house of that color.» It then began moving in the correct direction.

Thanks to Gemini, the AI agent can comprehend instructions even in emoji form. A command like «🪓🌲» would prompt it to chop down a tree.

SIMA 2 navigates photorealistic environments generated by Genie, accurately recognizing objects like benches, trees, and butterflies, and is capable of interacting with them.

With Gemini, the new version of SIMA can self-improve with minimal human intervention, using provided data merely as a basic guide.

The team places the agent in a new environment, while a separate model generates tasks for it.

SIMA 2 analyzes its shortcomings and gradually enhances its skills. Essentially, this is a trial-and-error learning process, but without human involvement; another AI system takes on the role of the mentor.

At DeepMind, they believe that this new solution is a step towards creating genuinely versatile robots.

«A system designed for real-world task execution requires two key elements: a high-level understanding of the world and reasoning capability,» emphasized senior research engineer Frédéric Bess.

For instance, if a humanoid robot is asked to check how many cans of beans are left in a cupboard, it must understand what beans are, what a cupboard is, and be able to reach the location.

SIMA 2 specifically addresses this «high-level behavior,» noted Bess.

The timeline for integrating this new solution into physical robots remains unclear.

Another area that intrigues the search giant is AI shopping. The company has launched a suite of new tools for online shopping, including:

«We believe that the shopping experience shouldn’t be so tedious. The idea is to preserve all the enjoyable parts of the process—browsing products and discovering surprises—while eliminating boring and difficult stages,» said Vidya Srinivasan, VP and head of advertising and commerce at Google.

One of the updates includes conversational shopping in AI Mode. Users can interact with the search engine like a chat bot, which will show product images and add details such as price, reviews, and availability.

The Gemini app has been taught to generate comprehensive ideas and collections rather than limiting itself to brief text-based suggestions for shopping queries. This feature is currently available only in the United States.

Agent checkout is an automatic monitoring service updating changes regarding items of interest. The service can send notifications about price drops.

«This is beneficial for buyers—they don’t have to continually check the price of a desired item. It’s also advantageous for sellers as buyers will return, even if they might have otherwise left,» said Lillian Rincon, Google Shopping’s VP of Products.

Another new feature allows the AI to call stores on behalf of the user to inquire about the availability of items and current promotions. This is based on the technology introduced in 2018, Google Duplex, Shopping Graph, and Google’s payment infrastructure.

To use the tool, users must specify the desired item. The AI will call local stores, ask for details, and provide a brief report.

As a reminder, in November, Google added message summaries, notification prioritization, and other AI-driven features to Pixel smartphones.