About Simon Willison

LLM Capabilities and Applications

Simon Willison extensively explores LLMs' prowess in code generation, multimodal tasks, and agentic systems. He highlights models like GLM-5.1 for autonomous 8-hour operation and advanced SVG/CSS animation generation with self-debugging [9][7]. Gemma 4's efficient multimodal variants enable local audio processing on Macs [38][33], while Qwen 3.5 supports secure local inference for sensitive journalism [57]. He demonstrates 'vibe coding' for SwiftUI macOS apps without Xcode [52][49] and profiles users from Hacker News comments [60][55]. Tools like llm-mrchatterbox run Victorian-era nano models locally [45][46][48].

AI Security Risks and Safety Measures

Willison warns of LLM vulnerabilities like prompt injection [41], data leaks in cloud agents [58], and Anthropic's system prompt filtering for third-party apps [22][19][23]. He covers supply chain attacks on Axios via social engineering [34][43] and AI-driven zero-day discovery [35]. New tools like scan-for-secrets detect leaked credentials [12][27][28][14], and he tracks AI-powered security research [31]. Anthropic's Project Glasswing restricts powerful models for vulnerability research [10][5].

Agentic Engineering and Development Workflows

A core theme is 'agentic engineering patterns' using LLMs as prototyping accelerators, not architects [29][72]. He adapts Red-Green TDD for agents [47], introduces Showboat for verifiable demos [63][79][74], and Rodney for browser automation [65]. AI shifts engineering to ambition and risk management [40][39][64], with tools like Pydantic's Monty for sandboxing [81]. He critiques cognitive limits in managing multiple agents [32] and coins 'Deep Blue' for AI-induced developer dread [76].

Datasette, Python Tools, and Open-Source Development

Willison maintains Datasette ecosystem tools like datasette-ports for instance management [16][17] and datasette-showboat [74]. Python innovations include uvx workflows [25][45], go-to-wheel for Go binary distribution [78][83], and chartroom for CLI visualization [75]. He discusses Starlette 1.0 [56], relicensing challenges with AI rewrites [68], and README-driven development [27].

AI Ecosystem Trends and Critiques

He analyzes model releases like Gemma 4 [38][8], Qwen departures [69], OpenAI acquisitions [61], and mission shifts [77]. Local inference optimizations like MoE streaming [54][3] and on-device apps [15] are key. Critiques include AI tropes in spam [59] and surveillance risks [55]. Videos cover AI in development [90][40].