/AI Weekly/Issue 97
OpenAI admits it screwed up testing its ‘sychophant-y’ ChatGPT update
Published Tuesday, May 13, 2025
OpenAI has shared details about an issue with a GPT-4 update that made ChatGPT overly agreeable and, in some cases, irritatingly eager to please. The company addressed how flaws in their testing process led to the rollout of these unintended changes, impacting the chatbot's behavior. OpenAI is now taking steps to refine its testing systems and ensure such missteps are avoided in the future.
Microsoft is adopting Google's new open protocol, Agent2Agent, to enable seamless communication between AI 'agents.' By supporting this standard, Microsoft aims to enhance interoperability and collaboration across AI systems, potentially streamlining tasks and improving efficiency. This move highlights a growing industry trend toward shared protocols for advancing AI technology.
Google has published an in-depth 76-page whitepaper that explores the technical intricacies of AI agents. The document covers topics like Agentic Retrieval-Augmented Generation (RAG), evaluation frameworks, and the design of real-world AI architectures. It provides insights into how these systems are structured and assessed, positioning itself as a resource for researchers and developers working with advanced AI models.
This paper introduces LLaMA-Omni2, a groundbreaking approach to real-time spoken chatbots. It focuses on leveraging large language models (LLMs) combined with autoregressive streaming speech synthesis to enable seamless and efficient interactions. The work highlights advancements in creating natural and responsive conversational experiences, underscoring the potential for further innovation in voice-based AI technologies.
An AI neural network has uncovered a surprising detail in a renowned Raphael painting, revealing that a specific face in the artwork was not actually painted by the master himself. Using its ability to analyze image details beyond human perception, the AI identified stylistic inconsistencies suggesting another artist's hand in that portion of the masterpiece. This discovery sheds new light on the collaborative practices of Renaissance art.
Anthropic has introduced an AI for Science program aimed at helping researchers tackle significant scientific challenges. By leveraging advanced AI tools, the initiative seeks to support "high-impact" projects that could drive meaningful progress in various fields. This move highlights Anthropic's commitment to fostering innovation and collaboration in the scientific community.
The AI Agents World Tour, part of AgentCon 2025, is a global series of one-day events tailored for developers working with autonomous AI agents. This June we will be bringing together top engineers and researchers. These conferences will take place in major cities like San Francisco and Singapore. It's a unique opportunity to connect, learn, and shape the future of AI technology.
A creative tinkerer has managed to get a modern large language model (LLM) running on the classic Commodore 64, a computer from 1982. The result is as quirky and nostalgic as you'd expect, with the ancient hardware pushing its limits to handle advanced tech. It's a fun mashup of retro computing and modern AI that showcases what's possible with some clever engineering.
Hugging Face introduces nanoVLM, a lightweight and accessible library designed to train Vision-Language Models using just 750 lines of Pure PyTorch code. This tool simplifies the development process, making it easier for researchers and developers to experiment with vision-language tasks from scratch. It's a practical solution for anyone looking to explore the intersection of visual and textual data without unnecessary complexity.
This podcast shares insights from top minds in machine learning and artificial intelligence, covering advancements, tools, and real-world applications. Episodes focus on conversations with industry experts, researchers, and practitioners, breaking down complex topics into engaging and accessible discussions. Perfect for anyone looking to stay informed about the latest trends and innovations in AI.
The AI links worth your time, curated by the community and always free.
