Large language models (LLMs) are being used to answer complex queries, leveraging tool-use strategies to augment capabilities and external knowledge. This approach helps answer questions requiring up-to-date knowledge or intricate computations, but LLMs still face challenges in few-shot tool-use applications.

Source: https://blog.research.google/

ChatDirector developed a decision-tree algorithm to enhance video conferencing by adjusting the layout and avatar behaviors based on ongoing conversations, aiming to minimize cognitive load and improve user experience.

Source: https://blog.research.google/

Large language models (LLMs) have revolutionized natural language processing, enabling better understanding of user behavior and improved personalization services. With user consent, LLMs can analyze interactions with digital systems and enhance summarization and question-answering capabilities.

Source: https://blog.research.google/

Large language models have made significant progress in capabilities like multi-hop reasoning, plan generation, and tool usage. However, their reliability in real-world deployment is sometimes compromised by the issue of 'hallucination', where models generate plausible but non-factual information.

Source: https://blog.research.google/

Google has introduced an experimental feature called Ask Photos, which allows users to ask for specific information within photos. The feature uses multiple Gemini models to deliver helpful responses. This feature was previewed at Google I/O 2024 and aims to simplify photo searching by allowing users to ask for specific memories or information.

Source: https://blog.research.google/

LANISTR has made a breakthrough in multimodal learning, focusing on structured data such as tabular or time-series formats, which is prevalent in real-world scenarios. This advancement integrates structured and unstructured data, with applications in healthcare and other fields.

Source: https://blog.research.google/

Researchers have developed methods to generate differentially private synthetic data, providing a strong guarantee that personal data will not be revealed by predictive models or their predictions. This technology is used in machine learning algorithms to protect user information in industries and academia.

Source: https://blog.research.google/

Med-Gemini is advancing medical AI to assist clinicians, researchers, and patients. The AI models require advanced reasoning and specialized medical knowledge to perform diverse tasks, such as generating radiology reports and summarizing health information, and can process complex multimodal data like images and videos.

Source: https://blog.research.google/

Graph visualization plays a crucial role in machine learning development, helping researchers and engineers debug models, identify performance bottlenecks, and optimize model architecture. This process is particularly useful for deploying models to devices with limited resources, such as mobile phones and browsers.

Source: https://blog.research.google/

Google's Connectomics team published maps of the 302 neurons in the Caenorhabditis elegans worm's nervous system, a milestone achieved after 10 years of work. The maps were created using AI and large datasets, building upon the 1986 connectome published for the same species.

Source: https://blog.research.google/