https://simonwillison.net/ · Editorial weight 9/10
Matt Webb argues headless APIs will proliferate as personal AI agents become the preferred interface to services, bypassing traditional GUIs. Salesforce's new Headless 360 product signals enterprise adoption of this architectural shift.
Anthropic released Claude Opus 4.7 on April 16, 2026, with updated system prompts compared to the February 4.6 version. The company continues its practice of publicly archiving system prompt changes, enabling transparency into how model instructions evolve across releases.
Simon Willison converted Anthropic's published Claude system prompts into a git repository with timestamped commits, enabling visual tracking of how the prompts have evolved across model versions and families.
Simon Willison demonstrates a practical agentic engineering pattern using a short LLM prompt to automate adding new content types to his blog-to-newsletter tool, showcasing how minimal prompting can accomplish substantial work in a single API call.
Simon Willison released llm-anthropic 0.25, adding support for Claude Opus 4.7 with extended thinking capabilities and new display options for reasoning output. The plugin now defaults to maximum token limits per model and introduces thinking_display and thinking_adaptive toggles.
Simon Willison compared Qwen3.6-35B-A3B and Claude Opus 4.7 using his informal "pelican riding a bicycle" benchmark, finding Alibaba's model produced superior image generation on a MacBook Pro M5 despite being smaller and quantized.
Google launched Gemini 3.1 Flash TTS, a new text-to-speech model accessible via the Gemini API that accepts natural language prompts to control voice characteristics and output style, expanding multimodal capabilities beyond text generation.
Kyle Kingsbury argues that companies will increasingly employ people as accountability holders for AI system failures—whether as internal reviewers, external legal representatives, or convenient scapegoats—shifting responsibility rather than ensuring genuine safety.
OpenAI launched GPT-5.4-Cyber, a fine-tuned variant designed for defensive cybersecurity applications, as part of a broader program to enable trusted access for high-capability models in sensitive domains.
The UK AI Safety Institute independently evaluated Claude Mythos's cybersecurity capabilities, confirming Anthropic's claims about its vulnerability-detection prowess. Analysis suggests performance scales with computational investment, framing security testing as a resource-intensive verification process.
Steve Yegge reports that Google's internal AI adoption mirrors the broader industry pattern: 20% power users leveraging agentic tools, 20% non-adopters, and 60% using chat interfaces like Cursor. An 18-month hiring freeze has stalled knowledge transfer across the sector.
Bryan Cantrill argues that LLMs lack the economic incentive to optimize systems efficiently, unlike humans whose finite time forces elegant design choices; this structural difference means unchecked LLM-assisted development risks bloat over improvement.
Simon Willison shares a practical recipe for running Gemma 4 E2B (10.28 GB) on macOS to transcribe audio files using MLX and mlx-vlm, with a ready-to-use uv command demonstrating local inference.