News

Last Week in AI: PyTorch, Crowdbotics, ManageEngine, Harness, Clupoints, 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 PyTorch's torchao native library, Crowdbotics new AI features, ManageEngine's new rec engine, Harness' new AI-powered agents, AMD's new language model, and more.

PyTorch introduced torchao, a new native library designed to improve the speed and efficiency of AI models by leveraging techniques like low-bit data types, quantization, and sparsity. The torchao toolkit, primarily written in PyTorch code, supports both inference and training, offering various methods for optimizing performance on models like LLaMA 3 and diffusion models. Benchmark tests show up to 97% speedup for LLaMA 3 8B inference and a 50% speedup for LLaMA 3 70B pretraining using float8 training on H100 GPUs, with minimal drops in accuracy. The library includes Quantization Aware Training (QAT) and low-bit optimizers to maximize accuracy and efficiency, alongside integrations with popular open-source projects like Hugging Face.

Crowdbotics introduced new AI-driven features to its application development platform aimed at resolving common challenges such as poor requirements and inconsistent architectural decisions, which lead to 70% of project failures. The new tools include Enterprise Context for enforcing development standards, Technical Recommendations for architecture guidance, and Application Code Generation, which automates up to 50% of code. The platform also provides a Cloud Consumption Cost Estimation tool and integrates with DevOps tools like Jira, streamlining deployment for Azure.

ManageEngine, a division of Zoho Corporation, launched Spotlight, an AI-driven recommendations engine, as part of the upgrade to its Analytics Plus solution, version 6.0. Spotlight is designed to identify inefficiencies in IT operations and suggest remediation strategies. According to ManageEngine, the tool helps IT managers analyze key metrics and make quicker decisions. For example, it can evaluate the age and failure rates of IT assets to recommend optimal replacement times. The AI also detects inefficiencies like bottlenecks in processes and suggests corrective actions.

Harness announced the release of three new AI-powered agents designed to enhance software delivery: AI QA Assistant, AI Code Assistant, and AI DevOps Assistant. The AI QA Assistant automates testing, promising 10x faster test creation and reducing maintenance needs by 70%. The AI Code Assistant accelerates developer productivity by offering intelligent code suggestions and natural language function generation. The AI DevOps Assistant simplifies pipeline creation and management with natural language commands and proactive optimization suggestions. These tools aim to streamline software development, testing, and delivery, reinforcing Harness’s commitment to AI-driven innovation.

Risk-based quality management (RBQM) and data quality oversight software provider CluePoints introduced a new application, Medical & Safety Review (MSR), designed to improve the efficiency of clinical trial reviews. The MSR tool enhances medical oversight by simplifying data analysis, query management, and ensuring transparency in study data. Key features include automated checks to reduce human error, a standard visualization library, and integrated review workflows for improved collaboration and faster decision-making. The tool is part of CluePoints’ ongoing effort to optimize clinical trials and improve patient safety.

AMD introduced a new AI language model, AMD-135M, built on the LLaMA2 architecture and optimized for use on its latest MI250 GPUs. The model, featuring 135 million parameters, is designed to enhance text generation and language comprehension tasks, providing developers and researchers with high efficiency and performance. Pretrained on datasets like SlimPajama and Project Gutenberg, AMD-135M is accessible through the Hugging Face Transformers library, making it easy to integrate into various applications. Early benchmarks show the model performs competitively in natural language processing tasks.

Airtable, creator of a no-code platform for building applications and workflows, introduced two new AI tools aimed at the enterprise, including App Library and HyperDB. App Library enables organizations to create standardized AI-driven applications that can be customized by different departments. HyperDB allows integration of datasets with over 100 million records from systems like Snowflake and Salesforce, offering improved accessibility while maintaining governance. Airtable’s cloud-based “Digital Operations Platform” allows teams to store, organize, and collaborate on structured data in real time, often described as "spreadsheets on steroids."

Redbird unveiled a new AI-driven platform designed to transform enterprise data analytics. The platform utilizes conversational AI agents to deliver advanced business intelligence, enabling users to perform complex data analytics through simple chat interactions, similar to a Google search. The platform addresses challenges companies face in implementing secure and accurate chat-based business intelligence, offering seamless integration with organizational data ecosystems. It also provides on-premises deployment options to ensure data privacy. Redbird’s AI agents handle tasks like data collection, SQL analysis, and reporting, empowering users with self-serve analytics beyond traditional tools like Tableau and PowerBI.

Google rolled out new features for its AI-powered note-taking assistant, NotebookLM, allowing users to upload videos from YouTube and audio files directly, alongside text, PDFs, Google Docs, and web pages. The updates further enhance the multimodal capabilities of the experimental tool, which leverages Google's Gemini 1.5 large language model. NotebookLM, part of Google Labs, now supports up to 50 sources and 1,000 notes per notebook, enabling users to analyze video essays, lectures, and audio recordings from meetings or study sessions. The tool generates summaries and source guides with key topics and suggested questions to facilitate deeper engagement with the materials.

HaystackID announced the integration of Relativity's aiR for Review into its Core Intelligence AI platform, further enhancing its generative AI eDiscovery solution. The collaboration aims to provide legal professionals a faster and more accurate approach to managing complex document reviews, addressing growing data volumes and strict legal requirements. HaystackID’s Core Intelligence AI combines advanced AI, data science, and machine learning to provide deep insights and improve document classification, relevance determination, and overall review accuracy.

Salesforce AI Research introduced SFR-Judge, a family of AI-powered judge models designed to evaluate large language models (LLMs). The SFR-Judge models, available in 8B, 12B, and 70B parameter sizes, can perform pairwise comparisons, single ratings, and binary classifications, while also generating explanations for their judgments. In extensive testing across 13 benchmarks, the company says, SFR-Judge outperformed other judge models, including proprietary ones such as GPT-4o, achieving top performance on 10 benchmarks. The models also excel at reward modeling, used to improve LLMs through reinforcement learning from human feedback (RLHF). SFR-Judge models ranked first, second, and fourth on the RewardBench leaderboard, demonstrating high accuracy and reduced evaluation biases.

Featured