September 25, 2025

AI

MIT’s MultiverSeg lets users outline organs in medical images with scribble clicks, learning from prior annotations to predict masks without further input. It surpasses existing methods, hitting about 90 % accuracy after one or two examples, supporting contexts and 3‑D applications.

Meta FAIR unveiled the 32‑billion‑parameter Code World Model, an open‑weight transformer combining code generation with real‑world execution. Trained through pre‑training, interpreter‑trace injection, instruction fine‑tuning, and reinforcement learning, it achieves impressive benchmarks and supports reproducible research.

Using a diabetes dataset, the tutorial scales data, trains a gradient‑boosting model, and evaluates it with residual plots and RMSE. It calculates permutation importance, plots partial dependence, then leverages Gemini AI to query, report JSON, critique risks, and suggest experiments.

Text‑RAG parses PDFs into OCR text, losing table layout and figures; Vision‑RAG retains whole‑page images, embedding them with models like ColPali, yielding 25‑39% performance boosts, and is preferred for graphic‑rich documents, whereas Text‑RAG suits low‑cost, text‑only cases.

Tech giants are rolling out new products and services, including Meta's AI-powered dating assistant Amanda Silberling and Google's plan to slash subscriptions. Other companies like Amazon and Facebook are updating their streaming services, while OpenAI publishes research on deceptive AI models.

Mindstate Design Labs has created MSD-001, a psychedelic-like drug developed with AI technology, which induces specific mental states without hallucinations, and showed promise in Phase I trial results, potentially treating mood disorders, compulsive disorders, and phobias.

ENEOS Materials implemented ChatGPT Enterprise, speeding research, improving plant safety, and streamlining HR workflows. Over 80% of employees report better daily tasks, strengthening the firm’s competitive edge in the manufacturing sector.

Alibaba's Qwen3-Max is a large-scale language model with over 1 trillion parameters, leveraging expert design and sparse activation per token, achieving strong performance in coding, agentic control, and math benchmarks, marking Alibaba's largest and most capable model to date.

Google's Search Live feature is now available to everyone in the US, allowing users to search through real-time conversations with an AI assistant via the Google app or Lens app, responding to queries and sharing relevant web links.

Microsoft is integrating Anthropic's Claude models into its Microsoft 365 Copilot platform, expanding model options and giving customers greater flexibility for Researcher and Studio, while following recent announcements favoring Anthropic over OpenAI for Visual Studio Code and other tools.

SAP and OpenAI have launched "OpenAI for Germany", a collaborative initiative aiming to introduce secure AI solutions to Germany's public sector by 2026, promoting a trusted and transparent AI ecosystem through safe and efficient public services enabled by AI technology.

Generative AI enables machines to create original content such as images, music, and text based on learned patterns, using complex algorithms to produce unique outputs indistinguishable from human creations, with applications in art, design, marketing, and media production.

Sam Altman's decision to privately pitch a $7 trillion AI initiative and publicly announce it on Twitter may have been driven by his psychological dependence on OpenAI, raising concerns about "ChatGPT Psychosis" and systemic risk when leaders become trapped in their delusions.

The FreeGF AI Chat user guide provides step-by-step instructions on using the chatbot's features, including creating characters and "uncensored" conversations, while emphasizing responsible use and caution with terms of service and privacy policy.