Simon Willison — Week of 2026-05-16 to 2026-05-22#

Highlight of the Week#

The most impactful milestone this week is the official announcement of Datasette Agent, merging Simon’s three years of work on his LLM library directly into Datasette. This conversational AI interface allows users to naturally interrogate their databases, boasting an extensible plugin architecture for charts, image generation, and secure code execution.

Key Posts#

[The last six months in LLMs in five minutes] · Source Simon shared annotated slides from his PyCon US 2026 lightning talk capturing a major inflection point in AI developer tooling. He highlights how coding agents crossed the threshold to become reliable daily drivers, and points to the astonishing capabilities of massive local models running on consumer hardware like Mac Minis.

[Google I/O, Gemini Spark, Antigravity] · Source Digging into Google’s Gemini Spark workspace integration, Simon questions the security of their opaque “Antigravity” architecture. He highlights the growing tension between building powerful agents with deep access to user data and the persistent, fundamental risks of prompt injection vulnerabilities.

[GDS weighs in on the NHS’s decision to retreat from Open Source] · Source Simon covers a rare public dispute within the UK civil service regarding AI security. After the NHS closed its open-source repositories over AI-discovered vulnerabilities, the Government Digital Service pushed back, arguing that “open by default” policies are essential for cost reduction, code reuse, and security scrutiny.

[The memory shortage is causing a repricing of consumer electronics] · Source Highlighting the massive physical footprint of AI, Simon links to analysis showing how data center demands for High-Bandwidth Memory (HBM) are eating up fixed silicon wafer capacity. This shift is severely constraining consumer RAM supplies and threatens to significantly drive up the cost of affordable electronics like sub-$100 smartphones globally.

Project Updates#

This was a massive week for releases, dominated by the launch of the datasette-agent ecosystem, which included plugins like datasette-agent-charts for data visualization and datasette-agent-sprites for secure sandbox execution via Fly Sprites. Simon focused heavily on transparency, shipping UX updates to expose the underlying SQL queries generated by the agent and ensuring robust permission checks. Additionally, he released datasette-llm-limits 0.1a0 for strict API cost management, and aggressively updated his llm-gemini CLI plugin to support Google’s new Gemini 3.5 Flash model and streaming reasoning tokens on the exact day of the model’s announcement.

Themes#

A recurring theme this week is the hidden economic and infrastructural costs of the AI boom. Simon kept returning to the realities of AI pricing at every scale: from his own need to build datasette-llm-limits to control local API spend, to analyzing Google’s 3x price hikes for Gemini 3.5 Flash, noting SpaceX paying Anthropic $1.25 billion monthly for compute, and tracking how HBM data center demands are pricing out cheap consumer electronics.