# Aug 10, 2025 - **Simon Willison's talk** at the Bay Area AI Security Meetup focused on **prompt injection** and the **lethal trifecta**, highlighting the urgent need for robust security in AI systems that utilize Model Context Protocol (MCP). - MCP overlooks hard-won lessons from distributed systems - **MCP’s design prioritizes simplicity over robustness**, neglecting essential lessons from 40 years of RPC best practices, which could lead to severe operational failures in enterprise environments. - GPT-OSS vs. Qwen3 and a detailed look how things evolved since GPT-2 - **OpenAI's release of gpt-oss-120b and gpt-oss-20b marks the first open-weight models since GPT-2, featuring optimizations like MXFP4 for local execution on consumer GPUs.** This advancement reflects a significant evolution in model architecture, emphasizing efficiency and accessibility for developers and researchers alike. - Diffusion Language Models Are Super Data Learners - **Diffusion Language Models** demonstrate exceptional capabilities as **data learners**, effectively leveraging vast datasets to enhance their performance in various tasks, as detailed in the article [here](https://jinjieni.notion.site/Diffusion-Language-Models-are-Super-Data-Learners-239d8f03a866800ab196e49928c019ac). - An AI-first program synthesis framework built around a new programming language - **Universalis** is a novel programming language designed for **knowledge workers**, enabling them to interact with AI through natural language, thus democratizing programming for non-developers. - Microsoft POML – Prompt Orchestration Markup Language - **POML (Prompt Orchestration Markup Language)** enhances prompt engineering for Large Language Models (LLMs) by providing a structured, modular approach that improves **readability**, **reusability**, and **maintainability** of prompts. - [P] I used YOLOv12 and Gemini to extract and tag over 100,000 scientific plots. - The author developed **Plottie.art**, a searchable library of **over 100,000 scientific plots**, utilizing a **custom YOLOv12 model** for subplot segmentation and the **Google Gemini API** for classification and keyword extraction, streamlining the visualization design process in research. - [P] From GPT-2 to gpt-oss: Analyzing the Architectural Advances And How They Stack Up Against Qwen3 - OpenAI's **gpt-oss-120b** and **gpt-oss-20b** are their first open-weight models since GPT-2, featuring optimizations like **MXFP4** for local execution on consumer GPUs, enhancing accessibility for developers. - [R] Adaptive Classifiers: Few-Shot Learning with Continuous Adaptation and Dynamic Class Addition - **Adaptive Classifiers** enable **few-shot learning** with as few as **5-10 examples per class**, achieving **90-100% accuracy** on enterprise tasks while continuously adapting to new data without catastrophic forgetting. - The Magic of Herding - **Dogs utilize indecisiveness in sheep to effectively guide them**, revealing insights into controlling chaotic collectives, which has implications for robotics and swarm behavior. - Conversations remotely detected from cell phone vibrations, researchers report - Researchers at Penn State have demonstrated that **conversations can be remotely detected** from cell phone vibrations using a millimeter-wave radar sensor, achieving up to **60% accuracy** in transcribing speech from a distance of **three meters**. - [R] Associative memory inspires improvements for in-context learning using a novel attention residual stream architecture - The **AMICL** algorithm enhances in-context learning by identifying incomplete patterns, searching for similar complete patterns, and completing them, achieving **near-perfect performance** on classification tasks.