The Week in AI: New Cohere LLM, Google GenAIOps, and Adobe AI, Averlon Out of Stealth, More

This edition of our weekly roundup of AI products and services includes a new version of Cohere's Aya family of multilingual LLMs, the new GenAI Ops service from Google Cloud, Adobe's new Generative Remove tool, Averlon's emergence from stealth, and more!

TrueFoundry introduced a new open-source Retriever-Augmented Generation (RAG) framework called Cognita, which was designed to provide robust, scalable solutions for deploying AI applications. Cognita was developed on top of Langchain and LlamaIndex to provide a structured and modular approach to AI application development. Each component of the RAG, from data handling to model deployment, is designed to be modular, API-driven, and extendable, the company says. The framework's user interface was designed to allow both technical and non-technical users to experiment with different RAG configurations in real-time. The UI supports document uploads and interactive question-answering, making it a valuable tool for developers and end-users alike.

The non-profit research lab, Cohere For AI, announce Aya 23, a new family of multilingual, generative large language research models (LLM) covering 23 different languages. The company said it is releasing both the 8-billion and 35-billion parameter Aya 23 models as open weights as part of its "continued commitment to multilingual research." The Aya 23 model family builds on Aya, an "open science" movement that brought together 3,000 collaborators from around the world to build the world's largest multilingual instruction fine-tuning dataset and state-of-the-art massively multilingual model. Aya 101 covered 101 languages and is focused on breadth; Aya 23 focuses on depth by pairing a highly performant pre-trained models with the recently released Aya dataset collection. The result is a powerful multilingual large language research model serving 23 languages, the said, which expands state-of-the-art language modeling capabilities to nearly half of the world's population.

Google Cloud announced a new service called Generative AI Ops. Delivered via the company's Google Cloud Consulting arm, as well as its partner ecosystem, the new service was developer to "help organizations mature their gen AI prototypes into production-grade solutions," and to provide support in areas like security, model tuning and feedback, and optimization. The new Generative AI Ops services "move customers through the steps required to make AI applications production-ready." The steps include Prompt engineering, design, and optimization; Performance and system evaluation; Model optimization and continuous tuning; Monitoring and observability; and Business integration and testing.

Averlon, a cloud security startup, came out of stealth mode and announced its official launch. The company makes an AI-driven platform designed to enhance threat prediction and neutralization, and its founders plan to use $8 million in new funding to expand its capabilities. Averlon specializes in a cybersecurity platform tailored to prevent malicious attacks. Averlon's technology utilizes the same AI that powers cyberattacks to counteract them. The platform offers "panoptic visibility" into all cloud assets, continuously analyzing vulnerabilities to predict and neutralize attacks. This proactive strategy enables teams to grasp the broader threat landscape and address issues with precision, the company says.

Adobe unveiled Generative Remove in Adobe Lightroom, bringing the capabilities of Adobe Firefly directly into everyday photo editing workflows across Lightroom mobile, web, and desktop surfaces. Generative Remove, Lightroom’s remove tool gives users the ability to remove unwanted objects from any photo non-destructively in a single click; it intelligently matches the removed area with "pixel perfect" generations. "From removing distractions in family photos, to empowering professionals with speedier retouching workflows and more fine-grain control, Generative Remove empowers exciting capabilities for all photographers." Generative Remove is available today as an early access feature across the Lightroom ecosystem for millions of users. 

Nightfall AI, a cloud data protection startup, announced the release of Firewall for AI, a new detection engine designed to prevent sensitive data exposure in AI apps and data pipelines. Firewall for AI was built to address critical issues of sensitive data exposure and prompt injection, which pose significant risks to companies utilizing public large language models(LLMs) or generative AI services from providers such as OpenAI and Google LLC. Nightfall's new service aims to mitigate these concerns with robust security, operational, and content guardrails for AI models and applications. The Firewall for AI functions as a client wrapper, safeguarding company and customer interactions with generative AI-based applications and data pipelines. It prevents sensitive data from being exposed to LLMs by scanning automation workflows and data pipelines to remove personally identifiable information, payment card information, personal health information, and secrets. It also protects against prompt injection attacks by detecting malicious content and ensuring proper language use, code use, response relevancy, and sentiment analysis. And it maintains data quality by filtering out proprietary, malicious, toxic, and irrelevant content from datasets.

Pinecone, provider of a fully managed vector database designed to make it easy to add vector search to production applications, announced the general availability of Pinecone serverless on AWS, and it announced the public preview of Private Endpoints for AWS PrivateLink to users on the Enterprise plan for advanced security. Pinecone serverless is a "completely reinvented" vector database designed to let users easily build fast and accurate GenAI applications at up to 50x lower cost, the company says.