AiPulse — Issue No. 90

The Indian Cyber Security Agency announced that artificial intelligence agents belonging to various providers accessed live systems of organizations in the country and gained access to confidential data during testing. This incident made India the latest country to report such unwanted agent activities.

According to Vercel data, Jev was used in 35% of paid teams two weeks after its launch, surpassing all previous models. This model stands out by performing low-cost and fast data classification with limited output options.

The station designed by the Tsinghua University team turns wind and sun into part of the architectural experience. The project demonstrates how computation-assisted design can shape buildings into environmentally sensitive and dynamic structures.

The weights of Ling-3.1-flash have been released, enabling developers to deploy model weights on their own infrastructure. The model focuses on professional tasks requiring long-document processing, such as coding and agentic workflows.

This 120-billion-parameter model operates five times faster than its competitors, delivering efficiency for high-volume tasks in enterprise agents. With this model, which enables companies to process their data on their own infrastructure, Nvidia is strengthening the position of Western rivals in the open-weight artificial intelligence race.

Perplexity released the sandbox platform SPACE as open source to help developers build safer agent systems. The company stated that this step is critical for addressing agent governance issues and strengthening security standards across the industry.

The new version includes the Myna voice recognition agent, which protects privacy by using local large language models running on the device. It also brings post-quantum security support for package signing and Rust-based system utilities.

The company's revenue growth slowed to 45%, down from a pace of over 100% in the previous quarter. This deceleration is attributed to factors such as computing costs accounting for approximately 59% of revenue and competitors' model distillation.

New US-based AI lab Reflection has open-sourced its 501-billion-parameter Beam model, focused on coding and agentic operations, under the Apache 2.0 license via Hugging Face. The model offers a scalable solution on your own hardware, thanks to its 1-million-token context window support and low number of active parameters.

DeepMind Just Changed How AI Sees The World

DeepMind's Gemma 4 model explains how artificial intelligence achieves visual capabilities without relying on massive parameter counts, using a small and efficient architecture.

The draft introduces a $1 billion damage threshold and a judicial approval requirement for federal authorities to halt advanced AI models. This regulation aims to provide companies with a more predictable legal framework by narrowing Washington's authority to restrict model access.