In the new issue of #InfocusAI digest, we’ll discuss the major releases of Grok 3 by Elon Musk and Claude 3.7 Sonnet by Anthropic, as well as learn about full-sized androids from Clone Robotics and an LLM for robot control from Figure.
AI-focused digest – News from the AI world
Issue No. 60, February 14 – 27, 2025
Elon Musk’s xAI Launches Grok 3
The biggest news of recent weeks is the release of Grok 3, which Elon Musk himself called “the smartest AI on Earth.” The model was originally slated for release in 2024, but its launch was delayed by several months. The new LLM is 10 times more powerful than its predecessor, Grok 2, and, according to the company, it has already surpassed GPT-4o (OpenAI), Google Gemini, and DeepSeek V3 in response quality. Grok 3 comes in several versions: for example, Grok 3 Mini responds faster to user questions at the cost of some accuracy, while Grok 3 Reasoning analyzes complex tasks and “reasons” through solutions, explaining its logic. The model also features a special Big Brain mode for solving complex computational and logical tasks, useful for programming or scientific research. Some features of Grok 3 are still in beta testing.
Equally noteworthy is the model’s communication style. For instance, the chatbot sometimes uses profanity, provokes conflicts, flirts with users, and even ends conversations on its own. Initially, the model made negative remarks about Elon Musk but later stopped answering questions about its “creator.”
Anthropic Releases Claude 3.7 Sonnet
The American company is holding its ground in the AI race. It has opened access to its first reasoning LLM, Claude 3.7 Sonnet, which solves complex tasks in mathematics, coding, finance, and law, working faster and more accurately than its predecessor. The model is available via API, Amazon Bedrock, and Google Cloud. Users can control the speed and depth of reasoning. The model excels at “agentic coding” and demonstrates improved performance in tests, including video games. Claude 3.7 Sonnet can be tested for free, but the reasoning mode is only available via subscription.
Clone Robotics Unveils Prototype of Full-Sized Android
Clone Robotics has introduced a prototype of an anatomically accurate humanoid robot, Protoclone V1, which replicates human movements using synthetic muscles. The company aims to create an android that mimics the main biological systems of the human body. Protoclone is equipped with a polymer skeleton that replicates 206 human bones and 1,000 artificial muscles. It also features 500 sensors to track pressure and joint position, as well as a 500W electric pump to circulate hydraulic fluid through its “body.” The company envisions the robot becoming a household assistant—for example, cooking meals or doing laundry. In a video released by the company, the android is shown suspended in the air but already moving its arms and legs like a real human.
Figure Introduces LLM for Robot Control
The startup Figure has developed Helix—the first universal “vision-language-action” (VLA) model for humanoid robots. With Helix, robots can perceive the world around them, analyze and understand human speech, learn, and execute user commands. Thanks to this development, they can synchronize with each other to complete tasks and work together. For example, in a video released by the company, two robots are seen placing groceries in a refrigerator.
The Helix model consists of two components. One is based on an open-source model with 7 billion parameters and is responsible for action planning and processing information from the external world. The other converts these decisions into specific robot movements. The systems run on separate GPUs to avoid interference. Training the model required only 500 hours. This project reportedly inspired Figure so much that the company ended its collaboration with OpenAI in robotic AI shortly before the announcement.
ITMO Develops Digital Assistant for Scientists
Researchers at ITMO University have created a digital assistant, ChemCoScientist, which automates research in chemistry and helps solve routine scientific tasks that require significant time. The neural network generates new compounds, refines existing ones, predicts chemical properties, and develops synthesis algorithms. The assistant can be used via a chatbot, requiring no programming skills. In practice, it will simplify tasks for chemists, such as developing new drugs and materials for pharmaceutical companies and industrial production.
ChemCoScientist consists of three main components: an orchestrator agent for system management, an agent for managing automated machine learning systems, and specialized agents for specific applied tasks in chemistry. The use of a multimodal approach ensures the digital assistant’s high accuracy.
