What events in the field of AI took place during the week
The field of artificial intelligence continues to show high rates of development: the largest technology companies are updating models and tools, governments are clarifying regulatory rules, and developers are receiving more and more specialized solutions for working with code and data. Izvestia has compiled a selection of key events in the field of AI over the past week.
Google has released Gemma Scope 2 for LLM analysis
Google has introduced Gemma Scope 2, a new open set of interpretability tools for the Gemma 3 family of models. The solution is designed to analyze the behavior of language models and their internal representations and is aimed primarily at developers using Gemma in industrial operation.
The toolkit allows you to examine the work of individual layers of the model, analyze the mechanisms of attention and activation, and identify complex effects, including hallucinations and attempts to circumvent restrictions. Gemma Scope 2 is based on sparse autoencoders and transcoders trained on all layers of Gemma 3 models.
Microsoft will rewrite all the code for working with neural networks by 2030
Microsoft intends to transfer a significant part of its software products from C and C++ languages to Rust by 2030. Galen Hunt, the company's lead engineer, said that artificial intelligence agents capable of performing large-scale code changes will play a major role in the implementation of the project.
Hunt noted that the goal of the initiative is to eliminate unsafe sections of code from the point of view of working with memory and increase the stability of the largest systems. For this purpose, the company has created a specialized infrastructure that combines algorithmic graph models of the source code and AI agents controlled by these algorithms. Microsoft expects to achieve a processing rate of up to 1 million lines of code per month through the efforts of one engineer.
The new approach is expected to accelerate the modernization of software products and increase their reliability by switching to the Rust language, which minimizes the risks of errors when working with memory.
Mistral has released OCR 3 for document recognition
Mistral has released the third version of its OCR 3 document recognition system, demonstrating a significant increase in quality compared to the previous generation. According to the developers, the accuracy of recognizing shapes, tables, scans, and handwritten text has increased by 74%.
The model is capable of extracting text and images from complex PDF documents and converting them to structured Markdown and HTML formats. This simplifies subsequent data analysis and integration with generative AI systems.
High-precision OCR is the basic technology for document management automation, archive digitization, and technical documentation analysis, providing companies with access to previously unstructured data.
OpenAI has released GPT-5.2-Codex for programmers
OpenAI has introduced GPT-5.2-Codex, a specialized version of the GPT—5.2 model optimized for agent-based programming in the Codex environment. The new model focuses on complex and long-term tasks, including refactoring, migrating large codebases, and working in a Windows environment.
The company noted that GPT-5.2-Codex has received enhanced cybersecurity capabilities. The growth of such capabilities requires a cautious approach to deployment due to dual-use risks, so access to the model will expand gradually.
Currently, GPT-5.2-Codex is available to paid ChatGPT users in all Codex access options. In the coming weeks, OpenAI plans to open secure access to the model via the API, as well as launch a pilot trusted access program for specialists and organizations working in the field of defensive cybersecurity.
All important news is on the Izvestia channel in the MAX messenger.
Переведено сервисом «Яндекс Переводчик»
