Welcome to the first spring issue of of #InfocusAI digest. In this edition, we’ll discuss the remote competition among AI agent developers, the benefits of causal machine learning, a new method for analyzing medical images from the USA, and a neural network for psychological testing from Russia.
AI-focused digest – News from the AI world
Issue No. 61, February 28 – March 13, 2025
Chinese Startup Releases AI Agent for Solving Complex Tasks
According to its developers, Manus can independently solve complex tasks—from travel planning to financial analysis. The name of the agent carries deep meaning: it derives from the Latin words Mens et Manus (“mind and hand”). This highlights its enhanced planning capabilities compared to Operator by OpenAI and Claude Computer Use by Anthropic. The system works with text prompts—users describe what they want to do, and Manus converts natural language instructions into actionable steps. The developers claim that Manus outperforms DeepResearch by OpenAI on the GAIA benchmark, which evaluates AI agents on practical task execution. However, gaining access to the agent is not easy, and technical details remain undisclosed. Those who have tried Manus report that it sometimes misses obvious information and struggles with tasks like booking tickets and ordering food.
OpenAI Introduces New Toolkit for Developing AI Agents
The American company is also keeping pace in the tech race. OpenAI has released a comprehensive toolkit for developing AI agents—Responses API, which will replace the Assistants API. The new platform enables the creation of assistants that can independently search the internet, extract files, and analyze corporate documents through a search tool, as well as automate data entry. At the core of the solution is the Computer-Using Agent (CUA), the same model used to create the earlier Operator agent. The developers acknowledge its imperfections and plan further refinements. OpenAI also released the Agents SDK, an open-source toolkit that allows integration of models with internal systems, implementation of protective mechanisms, and monitoring of agent activity.
European Researchers Advocate for Wider Use of Causal ML
MIT Sloan Management has published an article on causal machine learning (Causal ML). The authors—researchers from the University of Lausanne, Ludwig Maximilian University of Munich (LMU), and ETH Zurich—highlight the potential of this method. Unlike traditional ML, which makes predictions based on correlations, Causal ML answers questions like “What will happen if we do X?” —similar to how marketers conduct A/B tests to determine which ad drives more sales. According to the authors, Causal ML has significant potential for businesses, as it can improve decision-making processes and provide answers to practical questions such as “What if we increase advertising?” or “What effect will a new employee training program have?” More details on Causal ML can be found not only in the MIT Sloan Management article but also in the developers’ original paper.
US Researchers Develop a Faster, More Accurate Method for Analyzing Medical Images
A team of researchers from leading US universities has introduced a new machine learning method for analyzing temporal changes in medical images—Learning-based Inference of Longitudinal Image Changes (LILAC). This method automatically analyzes changes in examination results obtained at different times. Developed by researchers from Cornell University and Stanford University, LILAC uses a convolutional architecture to compare pairs of images, identifying significant changes while ignoring irrelevant variations like noise or differences in the field of view. The method is trained on two main tasks: predicting which image was taken earlier and forecasting specific changes in medical indicators. According to the scientists, LILAC stands out for its versatility—it requires no complex preprocessing and can be applied to various data types, from microscopic images to MRIs. The researchers hope LILAC will find widespread use in clinical practice and improve understanding of complex processes such as embryo development, wound healing, and brain aging.
Russian Researchers Create Neural Network to Detect Anxiety in Children
Researchers from the Moscow State University of Psychology and Education (MSUPE) have developed an ML model to help psychologists identify personality traits in children based on their drawings. The neural network is trained to detect signs of increased anxiety or aggression. To use it, a specialist asks the child to draw a person with a pencil on a sheet of paper, photographs the result, and uploads it to the application. The system analyzes the drawing and provides a personality assessment of the subject. This solution is expected to assist novice educators and psychologists in obtaining a second opinion in ambiguous cases and avoiding diagnostic errors.
