In this issue of #InfocusAI, we discuss how minor text quirks can affect LLMs’ medical recommendations, introduce a new genome research tool by Google and present the next VLA model for robots. There is also news about a new Claude feature that enables creating and publishing AI apps and recent research by the Max Planck Institute for Human Cognitive and Brain Sciences on differences between human and AI perception of the world.
AI-focused digest – News from the AI world
Issue 67, June 11-26, 2025
Extra spaces and typos in user queries can affect the accuracy LLMs’ medical recommendations
MIT researchers discovered that sometimes, large language models can issue different treatment recommendations not because of varying clinical data in user queries, but due to typos, extra spaces or choice of phrasing. In other words, simply changing the style or grammar in your message to the model can result in it recommending that you simply monitor your health, while previously it would send you to a doctor with the exact same symptoms. Curiously, these text variations more often cause changes in the model’s recommendations for female patients… This finding underscores the need to thoroughly audit LLMs before using them in healthcare. You can read the full scientific report here.
Google released new AI model for studying the human genome
Google presented AlphaGenome, a new AI tool that uses data from long DNA chain analysis to predict how their variants or mutations can influence biological processes. The training data was taken from large public bases, including ENCODE, GTEx, 4D Nucleome and FANTOM5. The developers say the model outperforms competitors in a number of benchmarks. Its draft version is now available for non-profit research via API AlphaGenome.
Anthropic expanded Claude with AI app creation and publication functionality
Claude now allows users to create AI-based apps and share them publicly without leaving the chatbot. The new feature builds on the Anthropic Artifacts function presented by Anthropic back in 2024. So far, it has been functioning in the beta version. However, the company’s message says its early users have already created a video game, data analysis apps and text-handling AI assistants. See details here.
Google released VLA-model that functions locally on robotic devices
Another piece of news from Google: the company presented Gemini Robotics On-Device, the most powerful VLA (Vision, Language, Action) model in its lineup that can run locally on robotic devices. The model operates independently of data transfer networks that makes it suitable for lag-sensitive applications and environments with unstable connectivity. Access to Gemini Robotics SDK requires registration. See details at the developer’s website.
Max Planck Institute for Human Cognitive and Brain Sciences compared human and AI perception of the world
New research from Max Planck Institute revealed qualitatively different approaches to world perception between humans and AI. In short, humans focus on semantic (conceptual) features of the surrounding objects while AI tends to prioritize their visual characteristics. This shift can be significant and define the difference between human and model reasoning. Considering the growing role of AI in critical sectors like healthcare and industry, this aspect should be taken into account to improve model reliability. Scientists studying differences between humans and AI may find insightful the approach the researchers followed there. You can read about it in the Nature Machine Intelligence article.
