88Z AI / NEWS
The news, the context, and what changes.
NVIDIA has introduced a Python interface for cuDLA, enabling developers to perform native inference using the Deep Learning Accelerator (DLA) on the Jetson AGX Orin platform. This solution addresses the challenge of running multiple neural networks within constrained compute and power budgets at the edge, improving system efficiency by offloading work from the GPU.
CICC and Ant Group signed a strategic cooperation agreement on November 13, combining their strengths in investment research and artificial intelligence to focus on financial large model innovation, AI-enabled wealth management, and full-lifecycle customer services.
Runway has released Praxis-1, an open-source robotics model with weights and code under the Apache 2.0 license. Leveraging video training knowledge for robot control, it tops Nvidia's RoboLab-120 benchmark in the World Action Model category, though it cannot yet plan complex action sequences independently.
Microsoft confirms its flagship software services have fully recovered from a global outage that disrupted governments and businesses. Affected products including Outlook, Teams, and Copilot are operational, with backlogged emails processed.
Ahead of a potentially record-breaking IPO, Anthropic CEO Dario Amodei acknowledged significant uncertainty about the coming year. With compute commitments nearing $901 billion and unclear revenue prospects, the listing aims to provide a financial buffer against these mounting costs.
OpenAI has dissolved its standalone robotics team for the second time, not due to data shortages, but because general-purpose models like GPT-6 Astra can reportedly control robots directly. Researchers are being reassigned to other teams to accelerate progress.
Google's Gemini 4 Carbon and Anthropic's Fable 5.2 have surpassed OpenAI's GPT-6.1 Sol in the MathArena AI math benchmark, ending its long-standing lead. Despite their superior performance, both new models come with significantly higher costs, with Gemini 4 Carbon priced at 15 times that of GPT-6.1 Sol. Tests indicate that Gemini 4 Carbon ignores time limits and rarely gives up on tasks.
Former Anthropic researcher Jacob Coxon states that the company's algorithmic improvements are accelerating, allowing it to retrain the Mythos model using only a fraction of the original compute. Citing this efficiency leap as his reason for leaving, he calls for a pause in frontier model development to let competitors catch up and commoditize human-level AI.
NVIDIA has introduced new double-precision floating-point (FP64) algorithms in its cuBLAS library, designed to significantly reduce computation time for high-precision dense calculations in fields ranging from molecular chemistry to materials science. This update targets a core bottleneck in scientific simulations, offering researchers more efficient computational tools to accelerate complex scientific discoveries.
The Wall Street Journal suggests Jensen Huang’s departure from Nvidia stems primarily from high-risk business tactics, such as offering massive financial guarantees for chip purchases. These moves sparked internal conflict when then-CFO Colette Kress attempted to curb the exposure, ultimately leading to his exit.
NVIDIA introduces cuPhoton library to accelerate black hole interferometry data processing using GPUs, addressing imaging challenges for distant celestial objects.
Google announces that developers can now claim Google Cloud credits directly through Google AI Studio, aiming to lower barriers to AI app development and accelerate prototyping.
