AI News — Wednesday, November 18, 2026
Nvidia is reportedly abandoning the Chinese market following the o ouster of CEO Jensen Huang, signaling a major shift in global AI hardware supply chains.
OpenAI unveils its latest flagship models, GPT-6.2 Astra and Sol, promising significant advancements in AI capabilities and performance.
Nvidia's venture capital investments in AI startups have reached $65 billion, but the departure of Jensen Huang raises uncertainty about the company's future investment strategy.
OpenAI launches ChatGPT for Sales, a specialized AI tool designed to enhance sales processes and improve customer engagement for businesses.
A new paper details the development of TeleChat to Xing4, showcasing advancements in frontier agentic models optimized for domestic chips and frameworks.
Google highlights new features for Gemini, enabling users to integrate the AI assistant more deeply into their daily tasks and workflows.
OpenAI announces Astra for Law, a new specialized version of its GPT-6.2 model tailored to provide advanced legal intelligence and support for the legal sector.
Despite recent leadership changes, Nvidia assures investors that its revenue growth trajectory is expected to continue its upward trend in the near future.
A new technical report introduces Sage S2.1 Flash, detailing performance improvements and new capabilities for this AI model.
AscendGYM is introduced as a new benchmark for optimizing kernel performance on Ascend NPUs, focusing on cost-awareness for real-world AI workloads.
InclusionMed is presented as a new benchmark designed to evaluate AI models for their ability to provide inclusive and equitable healthcare solutions.
SWE-CoWork proposes a new benchmark to evaluate the performance of coding agents as digital employees within organizational software development workflows.
A developer shares their experience building a 'gate' to filter out deceptive pull requests generated by AI agents, which are now opening millions monthly.
ScaleVLA introduces a new approach to scaling Vision-Language-Action models, achieving an impressive 552 billion parameters for enhanced embodied AI capabilities.
This paper explores an alternative programming paradigm, 'Weight-Space Programming,' arguing that it can achieve powerful results without relying on transformer architectures.