In short
Today: access risks in Microsoft Copilot, a new memory architecture for AI accelerators, and tools for agentic development. Plus, the latest research and community discoveries.
Today: risks of access in Microsoft Copilot, a new memory architecture for AI accelerators, and tools for agentic development. Plus, the latest research and community discoveries.
🔥 Hot:
🔹 Microsoft Copilot may show users excess data from SharePoint — The assistant respects access permissions, but overly broad sharing settings can legally allow it to find and summarize more than expected. 🔹 Samsung unveiled a zHBM prototype with memory placed directly on the AI accelerator — The new approach aims to integrate memory and the compute chip more closely. 🔹 Stanford released a new AI Index Report — The report compiles up-to-date data on the development and adoption of artificial intelligence.
➡️ News:
🔹 Gotcha 1.0 combines errors, traces, metrics, logs, and profiling — The self-hosted Go platform supports the Sentry SDK and OTLP, stores data in PostgreSQL and ClickHouse, and is distributed under Apache-2.0. 🔹 Nitter and XCancel resumed operation after consulting with lawyers — The services once again let users view posts from X without registration; the Nitter repository was restored from archive mode.
➡️ Useful materials:
🔹 An author compared LLMs for estimating calories from food photos — The benchmark used Nutrition5k photos and USDA and MEXT databases. 🔹 A study proposes two ways to explore the psychology of AI and humans through persona vectors — The project focuses on analyzing personas and their influence on interactions with models. 🔹 NodaLogic described an agentic pipeline for creating business systems — The approach automates the process from requirements formalization to the development and testing of web and mobile clients.
➡️ Discussions and case studies:
🔹 A developer turned old SOAP APIs into tools for AI agents — The project demonstrates a practical way to connect legacy enterprise interfaces to agents. 🔹 GuidedReview helps review AI-generated code before publication — The tool is designed to review a model’s output before a developer signs off on it. 🔹 Engrim adds local SQLite memory to AI command-line interfaces — It is a universal memory engine focused on local operation.
📝 If you would like to add other news and materials to the list, leave a comment.