Meta released Muse Wonder and Muse Wonder Full, their most capable AI models yet, targeting long-running agent work. Muse Wonder leads in memory benchmarks but trails Claude Opus 5.5 in program benchmarks, offering active memory and focus behaviors for developers.
Fused MoE optimizes large language model operations, reducing latency and costs. NVIDIA Hopper architecture utilizes WGMMA for peak tensor processing throughput.
Artificial neural networks can reorganize themselves to learn tasks like humans. MIT scientists argue that networks should be able to self-organize, like the brain, to provide insights into biological learning for therapies and brain-computer interfaces.
MIT Lincoln Laboratory hosted the NSCEB to discuss strengthening national biodefense capabilities. NSCEB's report calls for $15 billion federal investment to maintain U.S. leadership in biotechnology.
Orca Security's Discovery AI simplifies cloud security queries with natural language, reducing costs and latency significantly. The tool converts plain text into valid Sonar queries, bridging the gap between natural language and structured queries for over 900 cloud models.
SpaceXAI's Grok 4.7 is now on Amazon Bedrock, offering advanced AI models for complex tasks. Grok 4.7 excels in verifying its work and managing longer contexts, showing significant performance improvements over its predecessor.
AI is challenging accountability in public feedback, says MIT professor Deb Roy. MIT Center for Constructive Communication emphasizes trust in civic dialogues using AI.
Amazon SageMaker AI enables deploying decision models with llama.cpp for serving quantized models on CPU or GPU infrastructure. AWS DLC offers specialized containers for generative AI inference, including OpenAI-compatible chat completions and SemIf for structured decisions, enhancing deployment options and efficiency.
NVIDIA partners with industry leaders at OCP to advance open standards for next-gen AI factories, driving innovation and scalability. New contributions and partnerships aim to establish a common foundation for AI infrastructure, enhancing efficiency and reliability.
Amazon Bedrock offers global cross-Region inference for OpenAI GPT-6 models, including Sol, Luna, and Astra, providing broad access to capacity. Learn how to configure, set permissions, and monitor inference from your source Region, with reasoning and coding capabilities varying among the models.
Encoding categorical variables for tree-based regression can be simple (ordinal) or complex (one-hot, smoothed target). Surprisingly, simple ordinal encoding often works as well as more intricate methods like one-hot encoding.
Deploy large language model (LLM) across two GPU nodes on Amazon EKS with multi-node Ray LLM DLC. Multi-node setup improves performance for disaggregated inference workloads like RAG and document summarization.
An experiment with Gradient Boost regression using Blind Trees showed surprising success with weak learners. The Blind Trees technique improved predictions significantly with fast construction.
MIT report explores AI's impact on academia, highlighting rapid advancements in mathematics and the changing landscape of research and graduate training. Universities must adapt to a future where AI outperforms humans in many intellectual tasks, reshaping how departments evaluate and reward valuable contributions.
MIT associate professor Christina Delimitrou is using machine learning to optimize data centers, reducing energy waste and improving performance. By redesigning cloud computing systems and utilizing AI, she aims to enhance efficiency and minimize the need for new data centers.