Moondream has released the Photon system – a solution that enables AI to process video streams with zero latency on any class of device.
AI: Events
Fault Tolerance in Large Language Models: How DeepSeek Learns to Handle Failures
Technical context • Infrastructure
SGLang developers have introduced a partial fault tolerance mechanism for MoE models – now, the failure of a single node doesn't bring down the entire system.
OpenAI has released GPT-5.4 – a model with a one-million-token context window, built-in computer control capabilities, and a reduced error rate. But what does this have to do with our power grid?
AI: Events
When 31% of the Cache Just Vanishes: The Story of a Silent Bug Deep Within GPU Code
Technical context • Development
AI21 engineers spent weeks hunting down mystical glitches during model training, only to find the culprit in two characters of code at the GPU level.
AI: Events
A Thousand GPUs, One Cluster, and an Award for Best Cloud Solution: How SK Telecom Built «Haein»
Infrastructure
SK Telecom received the prestigious GLOMO award at MWC26 for its «Haein» GPU cluster – an infrastructure that combines over 1,000 NVIDIA B200 accelerators into a single system.
AI: Events
When an Agent Doesn't Know the Answer: How Retrieval Models Are Learning to Find the Unreachable
Products
Mixedbread has released Search v3 – a retrieval model that significantly narrows the gap between what an agent actually finds and what is theoretically discoverable within the data.
AI: Events
NVIDIA Contributes Key GPU Management Driver to Open Source Community for Cloud Infrastructure
Infrastructure
NVIDIA has handed over a vital GPU management software component to the Kubernetes community. This tool will now be developed by the entire industry, reducing the ecosystem's reliance on a single vendor.
PyTorch 2.11 has been released. This update to the popular neural network training framework brings notable improvements for distributed systems and Apple Silicon.
OpenAI has developed the IH-Challenge approach, which helps language models correctly prioritize instructions from different sources.