Simon Willison is focused on tracking and experimenting with cutting-edge AI models and tools, particularly multimodal systems, memory architectures, and developer utilities for working with documents and code.
Recent Activity
Fine-tune Gemma 4 and 3n with audio, images and text on Apple Silicon, using PyTorch and Metal Performance Shaders.
Highlights: This project enables fine-tuning of Gemma models (versions 4 and 3n) with multimodal data including audio, images, and text, specifically optimized for Apple Silicon hardware using PyTorch and Metal Performance Shaders. It provides a practical solution for running advanced multimodal fine-tuning on consumer Apple hardware.
Worth reading: It demonstrates how to leverage Apple's Metal framework for efficient multimodal training on accessible hardware, making advanced fine-tuning techniques more approachable for developers with Apple Silicon machines.
The highest-scoring AI memory system ever benchmarked. And it's free.
Highlights: MemPalace is a free, high-performance AI memory system that achieved top benchmark scores for memory retention and recall. It integrates with ChromaDB for vector storage and supports MCP (Model Context Protocol) for enhanced LLM interactions.
Worth reading: Its exceptional benchmark performance and open-source availability make it a valuable tool for developers building AI applications requiring reliable memory systems.
A friendly library for working with PDFs
Highlights: Natural-pdf provides a user-friendly Python library for PDF processing with a focus on simplicity and accessibility. It offers intuitive functions for common PDF tasks like text extraction, page manipulation, and basic transformations without requiring deep PDF specification knowledge.
Worth reading: The library's clean API and practical examples make it approachable for developers who need to work with PDFs but want to avoid complex low-level libraries.
@simonwillison.net
@simonwillison.net
@simonwillison.net
@simonwillison.net
@simonwillison.net
Simon Willison·Apr 7, 2026
<p>Anthropic <em>didn't</em> release their latest model, Claude Mythos (<a href="https://www-cdn.anthropic.com/53566bf5440a10affd749724787c8913a2ae0841.pdf">system card PDF</a>), today. They have instead made it available to a very restricted set of preview partners under their newly announced <a...
Highlights: Anthropic is taking a cautious approach with Claude Mythos by restricting access to security researchers through Project Glasswing, rather than releasing it publicly. This reflects growing industry awareness of AI safety risks and the need for controlled testing before broader deployment.
Worth reading: The post offers timely insight into how leading AI companies are balancing innovation with safety, providing context on current industry practices around responsible AI deployment.
Simon Willison·Apr 8, 2026
<p>Meta <a href="https://ai.meta.com/blog/introducing-muse-spark-msl/">announced Muse Spark</a> today, their first model release since Llama 4 <a href="https://simonwillison.net/2025/Apr/5/llama-4-notes/">almost exactly a year ago</a>. It's hosted, not open weights, and the API is currently "a...
Highlights: Meta's Muse Spark represents their first major model release in about a year, following Llama 4. Unlike previous models, Muse Spark is a hosted service rather than open weights, indicating a shift in Meta's AI deployment strategy.
Worth reading: The post provides timely analysis of Meta's strategic pivot in AI model distribution and highlights new tools available through meta.ai chat that developers and researchers should explore.