News

The Week in AI: Nvidia's Lightweight AI Model, Wiliot's GenAI Chatbot, Salesforce's AI Agents, More

• By Pure AI Editors
• 08/26/2024

This edition of our weekly roundup of AI products and services includes Nvidia's new Mistral-NeMo-Minitron 8B lightweight language model, Wiliot's new "WiliBot" GenAI chatbot that facilitates natural-language interactions with IoT-connected devices, Salesforce's new Einstein Sales Development Rep  and Einstein Sales Coach Agent AI-powered agents, and more.

Nvidia launched Mistral-NeMo-Minitron 8B, a lightweight language model designed to outperform similar-sized neural networks across a range of tasks. Mistral-NeMo-Minitron 8B is a scaled-down version of the Mistral NeMo 12B model that Nvidia debuted last month. The new model was created using machine learning techniques known as pruning and distillation, which reduce hardware requirements while maintaining output quality. Pruning involves removing less active components from a neural network’s code base, optimizing the model’s efficiency. Distillation transfers the knowledge from a larger AI model to a smaller, more efficient one, in this case reducing the model's parameters from 12 billion to 8 billion. Nvidia’s approach allows Mistral-NeMo-Minitron 8B to run on an Nvidia RTX-powered workstation while excelling in benchmarks for chatbots, virtual assistants, content generators, and educational tools. The release came a day after Microsoft open-sourced its own set of language models, similarly focused on hardware efficiency. Nvidia’s new model aims to offer developers a powerful yet compact tool for AI applications, the company said.

Internet of Things (IoT) platform provider Wiliot unveiled WiliBot, a GenAI chatbot designed to facilitate natural-language interactions with IoT-connected products. The company is billing the chatbot as a significant step in merging generative AI with real-time ambient data from the physical world. The goal is to make it possible for companies—and eventually consumers—to engage in meaningful conversations with the products they manufacture, source, distribute, and purchase, the company said. WiliBot was designed to leverage the vast amount of data generated by Wiliot's Ambient Data Platform, which utilizes stamp-sized, self-powered IoT Pixels affixed to various products, packaging, and containers. These IoT Pixels transmit critical information, such as location, temperature, humidity, and carbon footprint, to the Wiliot cloud, where businesses can analyze the data to optimize operations, the company said.

Salesforce introduced two new AI-powered agents for its customer relationship management platform. The new tools, Einstein Sales Development Rep (SDR) and Einstein Sales Coach Agent, were developed to streamline the sales process by automating tasks and providing personalized training for sales representatives.
The Einstein SDR Agent uses advanced "agentic AI" to autonomously engage with potential sales prospects, sorting out the most promising leads and taking such actions answering questions, handling objections, and scheduling meetings. This tool minimizes the need for human intervention in the early stages of the sales process, the company said, allowing sales teams to focus on closing deals. The Einstein Sales Coach Agent provides tailored training for sales reps through role-playing scenarios. It simulates real-life sales interactions, offering feedback and tips to help reps refine their pitches and negotiation strategies. Managers can connect these coaching sessions to actual sales outcomes to assess the effectiveness of the training. Both AI agents are built on the Einstein 1 Agentforce platform and will be available to Salesforce CRM customers in October, the company said. Users can customize workflows using prebuilt templates and integrate external data to enhance the agents' effectiveness.

LinkedIn unveiled the Liger (LinkedIn GPU Efficient Runtime) Kernel, a collection of highly efficient Triton kernels designed specifically for large language model (LLM) training. This new technology represents an advancement in machine learning, particularly in training large-scale models that require substantial computational resources. The Liger Kernel is poised to become a pivotal tool for researchers, machine learning practitioners, and those eager to optimize their GPU training efficiency. The Liger Kernel was crafted to address the growing demands of LLM training by enhancing both speed and memory efficiency, the company said. The Liger Kernel comes with a number of advanced features, including Hugging Face-compatible RMSNorm, RoPE, SwiGLU, CrossEntropy, FusedLinearCrossEntropy, and more. These kernels are efficient and compatible with widely used tools like Flash Attention, PyTorch FSDP, and Microsoft DeepSpeed, making them highly versatile for various applications, the company said.

Recogni, a leader in GenAI inference, introduced Pareto, a logarithmic number system designed to optimize AI chip performance. Pareto promises to revolutionize AI computing, the company said, by significantly reducing power consumption, chip size, and latency while maintaining high accuracy. Pareto simplifies AI computations by converting multiplications into additions, a breakthrough that allows Recogni's chips to be smaller, faster, and more energy-efficient. This innovation addresses the increasing demands of modern GenAI models, the company said, which require immense computational power. By reducing the need for power-intensive operations, Pareto outperforms traditional number systems like FP8 and FP16, delivering high accuracy with minimal energy usage. Recogni's extensive testing on models such as Llama3-70B and Stable Diffusion XL demonstrated that Pareto achieves over 99.9% accuracy compared to high-precision baselines, with significantly lower power consumption. This efficiency enables developers to deploy AI models quickly without the need for time-consuming retraining.

D-ID, a leader in digital human technology, has announced the general availability of its latest tool, D-ID Video Translate. Designed to help businesses and content creators engage multilingual audiences more effectively, the tool leverages cutting-edge GenAI to automatically translate videos into multiple languages, cloning the speaker's voice and adapting their lip movements from a single upload. With video content playing a central role in modern communication, translation often becomes a costly and time-consuming barrier to reaching broader audiences, the company said. D-ID Video Translate addresses this challenge by offering a simple, cost-effective solution that allows creators to transform their videos into multilingual experiences instantly. The tool supports a wide range of languages, including Arabic, Mandarin, Japanese, Hindi, Spanish, and French, and enables bulk translation for simultaneous multilingual output, the company said.