8am☕Coffee

Together AI introduces ATLAS, an adaptive speculator that learns from live workloads to accelerate inference. The system targets up to 400% speedups by predicting and optimizing execution paths in real time to reduce latency and boost throughput in dynamic production traffic.

Nvidia researchers propose a pre-training approach that has language models ‘think’ during training to improve reasoning. By integrating intermediate reasoning steps before fine-tuning, the method aims to strengthen problem-solving and generalization on complex tasks.

Microsoft Azure unveils what it calls the world’s first NVIDIA GB300 NVL72 supercomputing cluster for OpenAI. The deployment aligns Azure’s AI infrastructure with next‑generation NVIDIA systems to scale training and inference for frontier models.

Google rolls out a new Gemini capability that operates on the web in a user-like manner. The system navigates sites and performs actions to complete tasks, extending beyond static search responses to handle multi‑step workflows.

Cisco presents an AI data center router intended to relieve major infrastructure bottlenecks. The platform targets higher bandwidth, lower latency, and scalable reliability for networks moving intensive training and inference traffic.