AI News — Saturday, September 19, 2026
Researchers introduce DeepSeek-V4.1-Flash, a new method significantly improving KV cache compression for large language models, enhancing efficiency and performance.
A new startup focused on developing and launching other companies has secured $100 million in funding, with a strategic emphasis on advancing physical AI technologies.
A new paper details SoL-Pi, a system that uses recursive auto-research loops to create more efficient and capable AI agents, pushing the boundaries of autonomous research.
Hex announces a new integration with OpenAI's GPT-6 Astra, enabling users to transform complex data analyses into intuitive visual reports with advanced AI assistance.
OpenAI has released a new framework outlining its approach to reporting and addressing instances of AI model misalignment, emphasizing safety and ethical development.
This research proposes ActionPiece, a novel approach to action tokenization designed to improve the performance and understanding of autoregressive vision-language-action models.
A new empirical study investigates various harness designs to optimize the effectiveness and reliability of AI coding agents, providing insights for practical deployment.
Google's AI & Economy team has expanded with the addition of new experts, reinforcing the company's commitment to understanding and shaping the economic impact of AI.
Google shares fresh insights from its AI & Economy ATLAS initiative, offering a deeper understanding of AI's evolving influence on global economic landscapes.
EvolveTrade introduces a novel method for refining policies in LLM-based trading agents through experience-driven learning, enabling them to self-evolve and adapt.
Zing-0.5 is presented as a step towards creating playable virtual worlds that allow for real-time joint action and text-based control, enhancing interactive AI environments.
OpenAI provides guidance on effectively linking AI implementation to tangible business value, offering strategies for organizations to maximize their AI investments.
Researchers present VC-Attention, a technique that employs value smoothing and softmax casting to enable efficient low-bit attention mechanisms in AI models.
This paper explores the phenomenon of length inflation in on-policy distillation, specifically when End-Of-Sequence tokens exhibit disagreement, impacting model efficiency.
The press tour for AI personality Tilly Norwood is reportedly facing challenges, highlighting the complexities and public perception issues surrounding advanced AI interfaces.