What events in the field of AI took place during the week

The field of artificial intelligence continues to gain momentum: large technology companies are expanding their model lines, focusing on personalization and specialized industry solutions, and Russian research centers are offering applied AI tools for ecology and software development. Izvestia has prepared a selection of key events in the field of AI over the past week.

Google has updated the MedGemma medical line and introduced MedASR

Google has announced the release of MedGemma 1.5, an updated version of an open medical generative model focused on image interpretation and working with clinical data. The update affected the MedGemma 1.5 4B model, which received expanded support for high-dimensional medical images, including CT, MRI and digital histopathology, as well as improved X-ray series analysis and localization of anatomical objects.

According to Google Research, the accuracy of disease classification on CT increased by 3%, on MRI — by 14%, and the quality of interpretation of histopathological images approached specialized models. At the same time, the textual capabilities of the model were improved: the results on the MedQA and EHRQA medical question and answer kits increased by 5% and 22%, respectively. MedGemma 1.5 is available for research and commercial use through Hugging Face and Vertex AI and can work offline.

At the same time, Google introduced MedASR, an open automatic speech recognition model optimized for medical dictation. In comparison with universal ASR solutions, the number of errors in decoding medical dictations has more than halved. The company notes that the combination of MedASR and MedGemma allows you to build full-fledged voice scripts for clinical tasks, from filling out documentation to preliminary data analysis.

A new level: how neural networks will change cybercrime in 2026

Artificial intelligence opens up wide opportunities for hackers to attack

Gemini has received a beta feature of personalized responses based on user data.

Google has launched a new beta Personal Intelligence feature in the Gemini app, which allows the AI assistant to generate proactive and personalized responses based on photos, emails, search history, and video viewing. We are talking about deeper work with the Google ecosystem — Gmail, Google Photos, Search — without having to manually specify information sources.

Josh Woodward, vice president of Google Labs and head of Gemini, explained that the key feature of the mode is the ability of the model to reason based on heterogeneous data and link them together. In particular, Gemini can match correspondence, images, and video content to offer recommendations tailored to the user's personal context, from travel planning to shopping and entertainment.

Google emphasizes that the feature is disabled by default and is activated only at the user's initiative. At the same time, data from Gmail and Google Photos are not used to train the model and are used solely to generate specific responses. The company also notes the existence of protective restrictions for sensitive topics, including medical information.

Go to waste: AI will help clean the Arctic seas of debris

How the new technology will help save the region from plastic pollution

Russian scientists have developed AI to detect debris in the Arctic

Scientists at the Moscow Institute of Physics and Technology (MIPT), together with the Institute of Oceanology of the Russian Academy of Sciences, have developed an innovative system based on artificial intelligence (AI) technology for automatic detection of debris in the Arctic seas. The algorithm analyzes images from cameras mounted on the sides of ships, and is more accurate than its foreign counterparts.

According to Mikhail Krinitsky, head of the Machine learning Laboratory in Geosciences at MIPT, the system was trained on an array of more than 500,000 photographs taken in the Barents and Kara Seas. AI is able to distinguish between marine debris, birds, glare on the water and drops on the lens. The development is especially relevant for the Arctic, where anthropogenic waste practically does not decompose due to low temperatures, and a significant part of the pollution comes from sea currents from other regions of the world.

Experts note that such solutions can become the basis for automated environmental monitoring and assessment of anthropogenic load in hard-to-reach areas.

In Russia, programmers accelerated the training of AI models by 40%

Scientists from the Scientific and Educational Center of the Federal Tax Service of Russia, together with specialists from the Bauman Moscow State Technical University, have developed an automated scenario for configuring language models that allows AI to be adapted to specific measurable tasks without manually selecting parameters. According to the developers, the approach reduces the number of complex checks by about 1.6 times and speeds up setup by 40% without loss of quality.

Acceleration maneuver: the Russian Federation accelerated the training of AI models by 40%

The new system will eliminate the need to test hundreds of parameter combinations manually.

Igor Masich, Doctor of Technical Sciences, a leading researcher at the REC of the Federal Tax Service of Russia and the Bauman Moscow State Technical University, told Izvestia that the system automatically generates a set of optimal configurations for different scenarios, from maximum speed to accuracy priority. According to him, this allows the solution to be used in industrial AI services, as well as in government and corporate projects where reliability and efficiency requirements are particularly high.

The developers emphasize that the proposed approach is focused on practical application and can simplify the implementation of language models in applied IT systems.

All important news is on the Izvestia channel in the MAX messenger.