Pulse

Amazon SageMaker AI is enhancing robot reinforcement learning by integrating NVIDIA Isaac Lab. This allows for accelerated training of robot policies, such as for the Unitree H1 humanoid, using either SageMaker HyperPod for resilient, large-scale distributed training or SageMaker Training Jobs for ephemeral, on-demand compute. The platform aims to compress months of real-world training into hours by leveraging GPU-accelerated simulation and managed infrastructure, reducing the burden of compute cluster management for AI and robotics teams. AI

Researchers have developed a novel approach to accelerate machine learning on Field-Programmable Gate Arrays (FPGAs) using Kolmogorov-Arnold Networks (KANs). This method aims to achieve ultrafast inference and online learning by implementing neural networks directly as digital logic, bypassing the overhead associated with traditional processors like GPUs. The work, detailed in two papers, focuses on efficient evaluation and spline locality for KANs on FPGAs, addressing the need for ultra-low latency and high hardware efficiency in specialized applications. AI

NVIDIA has developed Dynamo Snapshot, a technology that significantly reduces the startup time for large language models from minutes to mere seconds. This innovation addresses the issue of idle GPUs during autoscaling by drastically shrinking memory snapshot sizes. The result is a much faster resumption of operations for large AI systems. AI

The NVIDIA RTX 6000 PRO workstation GPU is now listed at $13,250 USD on NVIDIA's official marketplace. This high price point for the professional-grade graphics card has surprised users in the local LLM community. The GPU is designed for demanding AI and professional visualization tasks. AI

Anthropic's Claude model is reportedly being integrated into Google Cloud under the codename "Claude fable" or "Claude Mythos." This suggests a potential partnership or offering where Google Cloud will host and provide access to Anthropic's AI capabilities. The exact nature of this integration, whether for internal use, specific customer offerings, or broader availability, remains to be detailed. AI

AMD's EPYC processors are being highlighted for their ability to handle the demanding execution needs of agentic AI systems. Concurrently, Zscaler has introduced a new zero-trust platform specifically designed to secure these agentic AI environments. These developments indicate a growing focus on the infrastructure and security required to support advanced AI agents. AI

Nvidia's RTX 5070 Ti graphics card with 16GB of VRAM is currently on clearance at Best Buy for as low as $500.99. This price point is considered a significant value for the performance offered, making it an attractive option for consumers looking for a powerful GPU. AI

A workaround has been developed to enable FP8 inference for Baidu's NAVA model within ComfyUI. This solution, available on GitHub, provides pre-configured workflow templates for various voice control features. Users can now integrate NAVA's capabilities into their ComfyUI projects for enhanced performance. AI

Together AI has partnered with Pax8 to make advanced AI infrastructure and open-source models accessible to small and medium-sized businesses. This collaboration aims to democratize access to powerful AI tools, ensuring they are not exclusively available to large corporations. The partnership will focus on delivering cost-efficient AI solutions to a broader market. AI

Meta has introduced "Workforce Academy," a new five-week training program designed to equip individuals with the skills needed for data center construction. This initiative, a collaboration with CBRE and Associated Builders and Contractors, offers free training and guarantees employment at a Meta data center construction site upon completion. AI

ALEPH is a new AI runtime designed for embedded hardware, drawing inspiration from biological immune systems for security. It features a SHA256 whitelist, stateful iptables, and an anomaly classifier to differentiate between inference loads and denial-of-service attacks. The system operates without cloud connectivity, pre-trained weights, or large language models, and has reportedly run for over 407,000 ticks without any crashes. AI

MiniMax AI has announced its models are now available on the Respan AI Gateway. This integration aims to provide developers with easier access to MiniMax's suite of AI models for various applications including text, speech, image, video, and music. AI

A live challenge is underway to optimize the inference speed of Google's Gemma 4 E4B model on a single A10G GPU. The competition, hosted on Hugging Face, invites participants to develop agents that can achieve faster processing times for the model. This event highlights efforts within the local LLM community to push the boundaries of hardware efficiency for AI models. AI

Ollama provides a straightforward method for users to set up their first local Large Language Model (LLM) in under five minutes. The installation process requires only three commands, offering a cost-free and privacy-focused solution for running AI models on personal devices. AI

Timeweb has released several updates in May, including improvements to their CDN, new agent capabilities for search and generation, and expanded data center locations. The company also saw significant growth, reaching 150,000 clients and quadrupling its agent count. These developments focus on the underlying infrastructure, such as networks and hardware, alongside new product features. AI

A Chinese company called "GPU god" has developed a single-slot, half-height PCIe version of the NVIDIA V100 GPU. This custom-designed card retains the full performance of the V100 core and is intended for passive cooling, with an option for higher power delivery. The 16GB version is expected to retail for under $220 USD, with a 32GB model also planned. AI

Apple has introduced CoreAI, a new on-device inference engine designed to replace CoreML and offer an alternative to existing frameworks like MLX and llama.cpp. This engine is optimized for Apple Silicon, particularly for mobile devices, and supports larger models, including a 20 billion parameter foundation model. While performance comparisons are pending, CoreAI aims to enable the deployment of more sophisticated AI models directly within applications. AI

Osmo has developed a system that digitizes smell, significantly reducing AI costs by 200x through the use of Meta's Llama model on AWS. Separately, RLWRLD and NVIDIA are collaborating to build the next generation of industrial infrastructure for humanoid AI. AI

Researchers are developing asynchronous neural networks that could significantly reduce AI's energy consumption, potentially by up to 100 times. This advancement aims to create more efficient and sustainable AI models. The breakthrough could pave the way for widespread adoption of AI by addressing its substantial environmental footprint. AI