AI Developments and Innovations
Google DeepMind’s FACTS Grounding Test
Google DeepMind has introduced a new benchmark called FACTS Grounding to evaluate AI models’ ability to provide fact-based and detailed answers based on given texts. The test includes 1,719 carefully selected examples from various fields such as finance, technology, retail, medicine, and law. Three leading AI models, Gemini 1.5 Pro, GPT-4o, and Claude 3.5 Sonnet, act as judges, assessing the responses based on whether the query is sufficiently answered and the factual correctness of the answer. To prevent manipulation, the benchmark is divided into two halves, with 860 examples available publicly and 859 withheld. The final evaluation is based on the average of both sets.
Nvidia’s Jetson Orin Nano Super
Nvidia has unveiled the Jetson Orin Nano Super, a compact and affordable “supercomputer” for generative AI. Priced at approximately $250, the developer kit offers up to 70% increased performance and 50% higher memory bandwidth compared to its predecessor. It is suitable for commercial AI developers, hobbyists, and students to develop skills in generative AI, robotics, or computer vision. The software updates for performance enhancement are also available for owners of the old Jetson Orin Nano Developer Kits.
AI Filters in Nvidia App
Nvidia’s GeForce graphics cards reportedly experience performance loss in games when the new Nvidia app is installed. Users have reported up to a 15% reduction in performance, even when the app’s features are not actively used. Nvidia recommends disabling certain AI functions in the app until an update is available. The app, which replaced the previous GeForce Experience, serves as a central hub for hardware control, driver downloads, and game optimizations.
Germany’s Energy-Efficient AI Ambitions
The Information Technology Society (ITG) highlights the importance of addressing the challenges of large language models to maximize the benefits of AI sustainably. They emphasize making data processing more efficient and reducing energy consumption to use AI sustainably. New approaches to overcome traditional chip design limitations are being explored, and Germany aims to maintain a leading position in technological innovation.
YouTube’s Control Over AI Training
YouTube introduces new settings that allow content creators to decide whether AI companies like OpenAI, Meta, or Google can use their videos to train models. The option is available under “Third-party training” in YouTube Studio settings and requires the consent of all rights holders. It applies only to publicly accessible videos that comply with YouTube’s guidelines.
Salesforce’s Agentforce 2.0
Salesforce has announced Agentforce 2.0, expanding its AI solution with pre-built workflows for sales and customer service. The new version includes a skill library for sales and marketing, which can be integrated into the Agent Builder. The skills also support integration with Tableau for data visualization and Slack for automated channel summaries. Agentforce 2.0 is set to launch in February 2025, with Slack integration planned for January.
OpenAI’s API Update
OpenAI has released a new version of the o1 model for its API and ChatGPT, which requires 60% fewer tokens for reasoning tasks compared to its predecessor. It also excels in math and programming tasks and function calling. OpenAI has reduced prices for the Realtime API for voice assistants by 60% and introduced “Preference Fine-Tuning” to adapt AI models to user preferences.
SwagBot – The Autonomous Herding Robot
An autonomous herding robot named SwagBot is being developed by Agerris, a startup from the University of Sydney, to assist in livestock and pasture management. The robot is equipped with sensors and cameras to monitor cattle and pasture conditions. It can recognize the health of cows and assess whether a pasture has been grazed. SwagBot aims to help farmers by collecting data and overseeing cattle and pasture areas.