News

Last Week in AI: Akamai, Familia.AI, Vertiv, H2O.ai, Preamble and More

This edition of our roundup of AI products and services announced last week that didn't make headlines, but should be on your radar includes Akamai's Behavioral DDoS Engine, Familia.AI's AI-powered family app, DeepSeek-AI's Janus, Preamble's system for mitigating prompt injection threats, and more.

Vertiv, a global provider of critical digital infrastructure, released a complete 7MW reference architecture for NVIDIA's GB200 NVL72 platform. The collaboration aims to transform traditional data centers into AI factories capable of powering enterprise-wide AI applications. The reference architecture, co-developed with NVIDIA, supports up to 132kW per rack and is designed to accelerate deployment of NVIDIA's liquid-cooled rack-scale platform. It offers an end-to-end approach to infrastructure design, optimizing deployment speed, performance, resiliency, cost, energy efficiency, and scalability. The collaboration is part of Vertiv's efforts to reshape critical power and cooling infrastructure to support accelerated computing and AI applications, the company said.

Researchers from DeepSeek-AI, the University of Hong Kong, and Peking University introduced Janus, an AI framework designed to enhance multimodal understanding and visual content generation. The model uses two distinct visual encoding pathways for understanding and generation, addressing conflicts found in prior approaches that relied on a single encoder.
Janus outperforms earlier models across various benchmarks, offering superior results in both tasks. The architecture uses SigLIP for semantic feature extraction and a VQ tokenizer for visual generation, processed through a unified transformer. The results show improved performance over models like LLaVA-v1.5 and DALL-E 2. The new model's ability to decouple understanding and generation tasks makes it a promising candidate for future AI developments across multiple modalities, researchers say.

H2O.ai Inc. unveiled two new small language models, Mississippi 2B and Mississippi 0.8B, designed for multimodal tasks like text extraction from scanned documents. The models, now available on Hugging Face under an open-source license, were developed to streamline AI-powered optical character recognition (OCR) and image analysis. Mississippi 2B, with 2.1 billion parameters, can process images based on user instructions, generate high-level descriptions, and extract details from visual data. Mississippi 0.8B, a scaled-down version with 800,000 parameters, is optimized for OCR tasks and outperforms similarly sized models in benchmark tests. H2O.ai, based in Mountain View, Calif., says the models provide high-performance AI solutions for businesses with limited processing power or latency-sensitive needs.

The Software Engineering Institute (SEI) at Carnegie Mellon University (CMU) announced the release of a new tool, Machine Learning Test and Evaluation (MLTE), to help teams developing ML-enabled software systems mitigate the common problem of ML models failing in production. ML model developers often work in isolation, lacking knowledge of the overarching system or its operational environment, SEI explained in a statement. Without this context, they can only evaluate a model on its accuracy or predictability, leaving software engineers and quality assurance teams with no specifications or knowledge to guide its testing. This disconnect can lead to models failing operational tests once deployed. To address this challenge, the SEI collaborated with the U.S. Army Artificial Intelligence Integration Center (AI2C) and CMU's Christian Kästner to create MLTE. The tool applies best practices from traditional software development to ML model test and evaluation, bringing together all stakeholders to negotiate the model's quality attribute requirements based on system needs. MLTE's semi-automated process and infrastructure enable the specification and testing of ML model and system qualities, incorporating an earlier SEI tool, TEC, that detects mismatched expectations among the teams building an ML component. Both TEC and MLTE are part of the SEI's effort to establish integrated test and evaluation of ML capabilities throughout the Department of Defense.

Familia.AI announced the launch of its AI-powered family app designed to create virtual family members, preserve family legacies, and provide emotional support. The app utilizes advanced AI to develop virtual family members capable of interacting with users, offering companionship and family support. This serves as a lifeline for individuals who have experienced family-related challenges, potentially helping to heal emotional wounds and family trauma. Beyond family support, Familia.AI acts as a digital family heirloom, capturing the essence, wisdom, and memories of real family members through old photos, video footage, and family stories. The launch of Familia.AI comes as studies show an increasing need for family support and connection, with family trauma linked to higher rates of loneliness, depression, and anxiety, the company says.

Akamai Technologies announced two major updates to its security portfolio that enhance protection against sophisticated threats and streamline security operations. The first update is the Behavioral DDoS Engine, which uses machine learning to provide automatic, proactive protection against application-layer distributed denial-of-service attacks. The service analyzes traffic patterns to detect and mitigate these attacks in real-time, reducing the need for manual intervention. The second update is a new AI Assistant, a chat-based tool integrated into Akamai's web security analytics platform. The AI Assistant allows users to quickly filter through security data, identify key information, and access relevant insights to speed up threat investigations. The Behavioral DDoS Engine is now available in limited availability, with plans to expand access platform-wide by 2025. The AI Assistant is immediately available as part of Akamai's web security analytics platform.

Preamble, a Pittsburgh-based AI security startup, has been granted a comprehensive patent for its system for mitigating prompt injection threats in artificial intelligence (AI) models. The patent was awarded by the US Patent and Trademark Office. Prompt injections, identified as a top risk by the Open Worldwide Application Security Project (OWASP), present significant vulnerabilities in AI. Preamble's technology provides a robust mitigation strategy, offering enterprises and users granular control over AI interactions. CEO Jeremy McHugh described Preamble’s platform as a "true end-to-end solution" for managing AI inputs and safeguarding sensitive data.

Featured

Upcoming Training Events

0 AM
Live! 360 Orlando
November 17-22, 2024
TechMentor @ Microsoft HQ
August 11-15, 2025