/AI Weekly/Issue 141
Big Tech Backs Anthropic Against Trump
Published Tuesday, March 17, 2026
A coalition of major tech companies has come out in support of Anthropic as it faces challenges from the Trump administration. The group criticized the government's stance, describing its actions as a "temper tantrum." This marks a significant show of unity within the tech industry as they rally behind Anthropic in this dispute.
Meta has acquired Moltbook, an AI agent social network that gained attention for its viral fake posts. The platform is known for its unique "always-on-directory" feature, which connects AI agents in an innovative way. This move aligns with Meta's push into AI-driven technologies and showcases their interest in expanding digital interactions through emerging tech platforms.
This content introduces a small vision-language model designed for robot task planning through multimodal behavior tree generation. It focuses on combining visual and language understanding to help robots effectively interpret tasks and execute plans in real-world scenarios. The approach aims to enhance efficiency and adaptability in robotic task execution using a compact and innovative model.
The paper introduces AutoTraces, a novel approach for trajectory forecasting using multimodal large language models (LLMs). It highlights how LLMs can effectively tackle autoregressive trajectory prediction tasks by leveraging multimodal data inputs. The approach shows promise in improving performance and adaptability in applications requiring accurate trajectory forecasting.
Martin Keen explores the differences between Retrieval Augmented Generation (RAG) and long context windows in large language model workflows. He discusses how tools like vector databases, semantic search, and embedding models affect AI performance, helping you determine the most suitable option for your needs. This breakdown provides insights into optimizing AI workflows for specific applications.
This video introduces LangSmith's new feature for supporting multi-modal evaluators in the UI. It explains how to evaluate interactions involving attachments and base64 multi-modal content with flexible mapping for both online and offline workflows. This enhancement enables a more comprehensive assessment of the full agent experience beyond just text outputs.
Amazon recently acknowledged in a meeting with engineers that the extensive use of AI in its e-commerce platform has contributed to challenges on its shopping website. Leadership highlighted how certain AI implementations have impacted the functionality and experience for users, leading to disruptions in its core business. The company is now focusing on addressing these issues to improve its operations.
A red-team experiment revealed how quickly autonomous AI agents could exploit weakly secured AI systems. Researchers at security startup CodeWall unleashed an offensive AI agent on McKinsey’s internal chatbot platform, gaining full read-write access to its database in just two hours. The breach exposed tens of millions of chat logs and hundreds of thousands of sensitive files in plaintext. While the test was ethical and quickly patched, it highlights a growing cybersecurity concern: AI agents can automate and accelerate attacks against other AI systems at machine speed.
Traditional search relies on keyword matching and often misses relevant results when wording differs. This tutorial explains how to build a semantic search engine using LLM generated embeddings that capture the meaning of text rather than exact words. Using Python and a Sentence Transformers model, it walks through generating embeddings and running nearest neighbor search to retrieve similar content. The result is a compact but powerful retrieval system and a practical foundation for modern RAG applications.
The latest update to the Microsoft Agent Framework enhances Agent Skills with new capabilities in the Python SDK. You can now define skills entirely in code, enabling more flexibility and customization. Additionally, the update introduces features like script execution and an approval process for managing agent skills more effectively. These improvements aim to make skill management and execution more streamlined for developers.
"The Cognitive Revolution" is a biweekly podcast hosted by Nathan Labenz and Erik Torenberg, featuring conversations with leading AI builders and researchers. The show focuses on the transformative potential of artificial intelligence and its impact in the years ahead. It is part of the Turpentine podcast network.
The AI links worth your time, curated by the community and always free.
