Last Week in AI: Liquid AI, Hyperscience, Orsini, Cloudflare, Mistral AI, Crowbotics, More -- Pure AI

Last Week in AI: Liquid AI, Hyperscience, Orsini, Cloudflare, Mistral AI, Crowbotics, More

By Pure AI Editors
10/07/2024

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 Liquid AI's first GenAI models, new capabilities in Crowdbotics' AI-powered app dev platform, Hyperautomation's Hypercell platform update, and more.

Orsini, a leader in rare disease pharmacy solutions, unveiled ORBIT (Orsini Rare Business Intelligence Technology), an advanced AI-powered reporting and analytics platform aimed at improving care for patients with rare diseases. The new platform leverages artificial intelligence to provide pharmaceutical manufacturers with deeper insights, personalized support, and data-driven decision-making. ORBIT's capabilities include real-time data exchange and robust analytics to enhance patient engagement and adherence, according to Ashok Singh, Orsini's Chief Information Officer. ORBIT's launch marks the first of several initiatives Orsini has planned to drive innovation in rare disease care.

Liquid AI, a startup founded by researchers from the Massachusetts Institute of Technology, announced the launch of its first set of generative AI models built on a new architecture. The Liquid Foundation Models (LFMs) are designed to rival top-tier large language models in both performance and efficiency. The startup, led by pioneers in the concept of "liquid neural networks," said its models are capable of handling a variety of data types with minimal memory usage. These systems, rooted in dynamical systems and numerical linear algebra, aim to cater to industries ranging from mobile applications to cloud servers. The company has already raised $37.6 million in seed funding and said it will provide early access to its models via several platforms, as it continues optimizing for hardware from major tech companies like Nvidia and Apple.

Hyperautomation company Hyperscience announced updates to its AI-powered Hypercell platform. The updates focus on accelerating automation for back-office processes and improving digital transformation efforts. The platform now features enhanced deep learning models, workflow orchestration, and infrastructure upgrades to help organizations embed AI and machine learning more deeply into their operations, the company said. The updates include a new deep learning model that automates long-form document extraction for complex materials, such as contracts, boosting efficiency, and decision-making. Additionally, the platform improves AI model lifecycle management and governance, allowing for easier adaptation to new document types while ensuring compliance.

The Allen Institute for AI (Ai2) unveiled Molmo, a family of open-source language models capable of processing both text and images, today. Molmo includes four neural networks, ranging from 1 billion to 72 billion parameters. The models can identify objects in images, count them, and explain visualized data. In internal evaluations, the 72-billion-parameter model outperformed OpenAI’s GPT-4o in object recognition tasks, the company said. Ai2’s models are trained on high-quality image-text pairs, enabling them to deliver efficient and accurate results. The nonprofit said the smallest model, with 1 billion parameters, is compact enough to run on mobile devices while outperforming some larger models.

IBM Research and NASA introduced Prithvi WxC, a 2.3 billion-parameter foundation model designed to improve weather and climate forecasting. Built on a transformer-based architecture, Prithvi WxC incorporates 160 variables from NASA’s MERRA-2 dataset, covering global atmospheric conditions. The model captures local and global dependencies, handling long-range atmospheric interactions with high precision. Prithvi WxC demonstrated exceptional accuracy in several key benchmarks, the company said, including predicting Hurricane Ida’s track with a mean error of just 63.9 km. The model also excelled in downscaling tasks, significantly outperforming traditional methods in temperature prediction, and proved effective in parameterizing gravity wave fluxes in the upper troposphere, the company said.

Cloudflare launched new features for its AI Audit tool that allows website owners to control how their content is used by artificial intelligence models and set prices for access. The tool helps users understand how AI models are scraping their content and offers options to block or monetize this usage. The move follows growing controversy over AI developers scraping content from websites without permission. Cloudflare’s tool aims to give creators more control and transparency, providing a one-click option to block AI scrapers and helping smaller websites set fair prices for content access.

Crowdbotics introduced new capabilities for its AI-powered application development platform aimed at solving common challenges in app creation, including poor requirements and inconsistent architecture. The enhancements target issues responsible for 70% of project failures and provide tools to predict cloud costs more accurately. Key updates include Enterprise Context, which enforces corporate standards during project planning, and Technical Recommendations, offering AI-driven architectural suggestions. Crowdbotics also announced Application Code Generation, automating up to 50% of code development, and cloud cost estimation features to inform early design decisions. Additionally, new Jira integration and Azure deployment tools streamline DevOps workflows and simplify app deployment.

Data Dynamics unveiled Zubin, its new AI-powered self-service data management software designed to help organizations optimize risk management, data privacy, sovereignty, and sustainability. Zubin was designed to promote digital trust and data democracy while addressing challenges posed by the growing influence of AI on unstructured data, which accounts for 80% of enterprise data. The software offers centralized governance with decentralized control, enabling data and application owners to manage and audit data through a low-code, self-service interface. Key features include risk management, data privacy, sovereignty, and role-based access control, which foster transparency and accountability across enterprises.

Mistral AI unveiled Pixtral 12B, its latest model capable of processing both images and text. The model uses 12 billion parameters and includes a 400-million-parameter vision adapter, enabling it to handle image inputs alongside text. Based on Mistral’s previous Nemo 12B model, Pixtral 12B offers multimodal capabilities similar to those of AI models from OpenAI and Google. Mistral released Pixtral 12B’s code and parameters via GitHub and Hugging Face, encouraging developers to download and fine-tune the model for various applications. The licensing details for Pixtral 12B remain unconfirmed.

Digital transformation company UST launched its new UST Retail GenAI Platform at its London Innovation Lab. The platform was designed to transform retail operations with generative AI, the company said. The platform integrates proven business models with key generative AI features, the company said, such as search, summarization, automation, and content creation to enhance retail operations. The platform allows retailers to safely pilot AI-driven solutions, accelerating innovation and improving operational efficiency.