Microsoft has announced the release of two new language models: MAI-Thinking-1, featuring 1 trillion parameters, and MAI-Code-1-Flash, with 137 billion parameters, both optimized for specific applications like GitHub Copilot (). In other developments, OpenAI has rolled out Lockdown Mode, enhancing security against data exfiltration risks in LLMs, while Uber has capped employee usage of AI coding tools to manage costs effectively (, ). Additionally, the release of datasette-agent-micropython 0.1a0 introduces secure code execution capabilities, enhancing safety in web applications ().
Microsoft has announced two new language models: MAI-Thinking-1, which has 1 trillion parameters with 35 billion active parameters, and MAI-Code-1-Flash, featuring 137 billion parameters with 5 billion active parameters. Both models are designed for specific applications, with MAI-Code-1-Flash optimized for GitHub Copilot and Visual Studio Code. The training for both models utilized clean, commercially licensed data, although it was later clarified that the training data includes a proprietary web crawl, raising questions about licensing practices in the development of large language models.
Simon Willison — 39 d ago · found 30 d agoModels
2.
datasette-agent-micropython 0.1a0
The release of datasette-agent-micropython 0.1a0 introduces a new capability for the Datasette Agent to generate and execute Python code in a secure environment. This alpha version demonstrates promising sandboxing effectiveness, with GPT-5.5 successfully contained without breaking out of the sandbox. This development is significant for practitioners as it enhances the safety of executing dynamic code within web applications, particularly in environments where security is paramount.
Simon Willison — 39 d ago · found 30 d agoOpen Source
3.
Pasted File Editor
The Pasted File Editor, a prototype developed using Codex, allows users to paste large volumes of text into applications like Claude AI, automatically converting them into file attachments. This tool also supports direct file opening, including images displayed as thumbnails, and drag-and-drop functionality for files. Its implementation enhances user experience in AI-assisted programming environments by streamlining text input and file management.
Simon Willison — 40 d ago · found 30 d agoCoding
the full briefing
Models & Releases
Microsoft has launched two new language models: MAI-Thinking-1 with 1 trillion parameters and MAI-Code-1-Flash featuring 137 billion parameters, tailored for applications like GitHub Copilot (). This release highlights the trend towards specialized models in the AI landscape. In a significant move towards enhancing security, OpenAI has introduced Lockdown Mode, which limits data exfiltration risks associated with prompt injection attacks, making it a crucial feature for users with high-risk profiles (). Meanwhile, Uber has implemented a cap on AI tool usage to manage costs effectively, reflecting the financial implications of AI tool integration in large enterprises ().
Tools & Techniques
The release of datasette-agent-micropython 0.1a0 introduces a secure environment for executing Python code, which is essential for enhancing safety in web applications. This alpha version demonstrates promising sandboxing effectiveness, allowing developers to execute dynamic code securely (). Additionally, the Pasted File Editor tool, developed using Codex, streamlines text input and file management in AI-assisted programming environments ().
Research & Evaluation
Recent research highlights the need for robust evaluation frameworks in AI applications. The RECAP benchmark addresses the limitations of existing benchmarks in prompt optimization, emphasizing the necessity for continual adaptation methods in dynamic deployment scenarios (RECAP: Regression Evaluation for Continual Adaptation of Prompts). Another study introduces a novel approach to evaluate the effectiveness of large language models in generating scientific hypotheses, showcasing the importance of human involvement in scientific AI applications (Towards Diverse Scientific Hypothesis Search with Large Language Models).