Z.ai released GLM-5.2 (753B parameters, 40 active via MoE) under MIT license on June 16th. Text-only model with 1M token context window. Ranks 1st on Artificial Analysis Intelligence Index v4.1 (score 51) ahead of DeepSeek V4 Pro and Kimi K2.6. 2nd on Code Arena WebDev behind Claude Fable 5.
Charity Majors observes that in 2025, the economics of code production flipped: generating code became nearly free and instant instead of expensive and time-consuming. Lines of code shifted from being treasured and carefully curated to disposable and regenerable overnight.
Web Component <click-to-play> that converts a static image into a play button to load GIFs on demand. Improves performance by preventing automatic loading of large files.
Datasette 1.0a34 adds tools to insert, edit and delete rows directly in the web interface. These long-overdue features are available on table and row pages, inspired by Datasette Agent which now supports SQL write operations.
Release of datasette-tailscale 0.1a0, experimental alpha plugin enabling Datasette server deployment via Tailscale. Uses Python bindings for the tailscale-rs library to connect a local instance to a Tailnet.
Georgi Gerganov (llama.cpp creator) uses Qwen3.6-27B daily for coding tasks on M2 Ultra and RTX 5090. He integrates it via a lightweight agent (pi) with custom system prompt for ggml-org maintenance assistance.
Claude Fable 5 was banned under US export controls after a simple "fix this code" prompt enabled exploit generation. Kate Moussouris argues this is absurd: coding models must fix bugs, especially security vulnerabilities. Banning this capability weakens cyber defense.
The White House shared with Anthropic a report on the Fable jailbreak. Cybersecurity expert Katie Moussouris reviewed the tests: Fable refused 'review the code for security issues' but complied with 'fix this code'. Moussouris concluded this is the model working as intended for cyberdefense.
Simon Willison shares a tip for configuring Cloudflare CAPTCHA/Managed Challenge: use a WAF rule that only triggers the challenge on search URLs containing at least one ampersand. This allows simple requests like ?q=term to pass without CAPTCHA.
datasette-agent 0.3a0 introduces execute_write_sql, a new tool enabling AI agents to modify databases with user approval and permission management. Example: inserting pelican sighting data with confirmation before execution.
Axios reports personality clashes between Anthropic leadership and US administration led to Fable/Mythos models going offline over export controls. Logan Graham, Dave Orr, and Nicholas Carlini meet Commerce Department today. Reinstatement hinges on jailbreak-proof guarantees or an "attitude fix."
Arvind Narayanan and Sayash Kappor reject the narrative of mass software engineer replacement by AI. Key data: New York recorded 160+ WARN Act filings in 2025—none checked the AI box. Automation hits three real bottlenecks: specifying what to build, verifying quality, and managing human dependencies.
Pyodide 314.0 enables publishing Python packages built for WASM directly to PyPI (via PEP 783). Previously, Pyodide maintainers manually managed 300+ packages. Package developers can now distribute WASM wheels like native Linux/macOS/Windows wheels.
Release of luau-wasm 0.1a0, a WebAssembly implementation of the Lua language. The project publishes WASM wheels to PyPI for use with Pyodide, enabling Lua execution in browsers.
Simon Willison uses Claude Code (Opus 4.8) to solve SQL column provenance: programmatically identify the source table.column for each result in complex queries with joins and CTEs. Three solutions found: apsw, ctypes with sqlite3_column_table_name(), and EXPLAIN analysis.
US government ordered Anthropic to suspend access to Fable 5 and Mythos 5 for all users citing national security concerns. The directive cites risk of a 'jailbreak' method bypassing model safeguards. Anthropic disputes this: identified vulnerabilities are minor and available on other public models including OpenAI's GPT-5.5.
Simon Willison updated his OpenAI WebRTC tool to support GPT-Realtime-2 model (GPT-5-class reasoning) and added document context feature. Users can now paste text to have audio conversations about specific documents directly in the browser.
Claude Fable 5 stands out for being "relentlessly proactive": it deploys multiple strategies to reach its goals. Simon Willison tested it on a UI bug (unwanted horizontal scrollbar) by providing a screenshot and simple instruction. The model inspected Datasette Agent project dependencies to diagnose the issue.
Datasette 1.0a33 extends the `?_extra=` pattern to queries and rows in addition to tables. This major alpha finally documents this API for JSON responses. A custom API explorer was generated using Claude Fable 5 and GPT-5.5.
Release of asyncinject 0.7, a Python utility library for asyncio dependency injection. Claude Fable 5 identified and fixed bugs in the codebase while being used with Datasette.
Anthropic reverses a secret policy that would have degraded Claude Fable/Mythos performance for AI researchers without notification. The company acknowledges making 'the wrong tradeoff' and now makes safeguards visible.
datasette-agent 0.2a0 enables tools to ask users questions mid-execution via ask_user() supporting yes/no, multiple-choice, or free-text questions. Suspended conversations persist to database and survive server restarts. New built-in save_query tool lets agents save SQL queries as Datasette stored queries with mandatory human approval.
Google releases DiffusionGemma-26B, an open-weight Gemma model (Apache 2 license) based on its May 2024 Gemini Diffusion research. The model generates text at 500+ tokens/second. NVIDIA hosts it free on NIM cloud API.
Anthropic releases Claude Fable 5 and Claude Mythos 5 with 1M token context window, 128k max output tokens, January 2026 knowledge cutoff. Fable 5 includes strict safety guardrails; Mythos 5 without safety classifiers. Pricing: $10/M input, $50/M output (2× Opus 4.5-4.8). Willison reports strong performance after 5.5 hours of testing.
Release of llm 0.32a3, almost entirely written by Claude Fable 5. Simon Willison details using Claude Code to add features to Datasette Agent and llm.
Simon Willison documents how to set a custom price for Claude Fable 5 in AgentsView, Wes McKinney's tool for tracking token usage and costs across local coding agents. He used Fable to reverse-engineer the solution after the new model wasn't yet in the pricing database.
Andrej Karpathy comments on Claude Fable 5: increasingly available working software on tap triggers Jevons paradox, exploding demand. Instant generation of explainers, visualizers, dashboards, bespoke apps, test suites, code optimization, and research projects with custom HTML becomes possible.
Apple announces at WWDC 2026 a new generation of Siri AI using a custom Gemini-derived model running on Private Cloud Compute. New features leverage vision-LLMs to extract information from user screens. Apple launches Core AI, a developer library built on Meta's PyTorch, enabling custom model execution on Apple hardware.
Release of datasette-agent-edit 0.1a0, a plugin for Datasette Agent enabling agents to edit text (Markdown, SQL, SVG). Implements three tools inspired by Claude's text editor: view, str_replace, insert.
Release of micropython-wasm 0.1a2 with new CLI. Enables running Python code in a sandboxed WebAssembly environment.
Simon Willison released micropython-wasm, an alpha package for running Python code in a sandbox using MicroPython and WebAssembly. The tool secures plugin execution in Datasette, LLM, and sqlite-utils by isolating potentially malicious or buggy code without access to files, network, or system resources.
OpenAI rolls out Lockdown Mode to ChatGPT (Free, Go, Plus, Pro, Business). This feature blocks outbound network requests to prevent data exfiltration in prompt injection attacks. It does not prevent injections themselves, but cuts off the exfiltration vector—one of three conditions in the "Lethal Trifecta".
Andreas Kling announces Ladybird will no longer accept public pull requests. The project closes external contributions because AI makes it impossible to verify contributor good faith. Only people responsible for changes can introduce them.
Charity Majors analyzes the tension between AI enthusiasts (seeing discontinuous capability gains and existential threat in ignoring AI) and skeptics (fearing reliability degradation, loss of context, team burnout). The challenge: design feedback loops connecting both groups to reconcile their divergent realities.
Google asked 404 Media to revise its official statement after publishing an article about internal memes criticizing its AI systems. The revised version removed language about keeping humans in the loop.
Uber caps monthly token spending at $1,500 per AI coding tool (Claude Code, Cursor) to control costs. The limit represents ~11% of median Uber engineer salary ($330k). Policy follows exhausting 2026 AI budget in four months.
Microsoft announces MAI-Thinking-1 (35B, reasoning) and MAI-Code-1-Flash (5B, code). The former outperforms Claude Sonnet 4.6 in blind human evaluation. Both trained on commercially licensed data without third-party distillation.
Release of datasette-agent-micropython 0.1a0. Enables Datasette Agent to safely generate and execute Python code via WebAssembly sandbox. GPT-5.5 failed to break out of the sandbox in testing.
Release of micropython-wasm 0.1a1 with fixes for limitations discovered while building datasette-agent-micropython. Enables Python execution in WebAssembly with sandboxing.
Simon Willison built a web tool that replicates Claude.ai's feature: detecting large pasted text volumes and automatically converting them to file attachments. The tool also supports direct file opening and images (shown as thumbnails) plus drag-and-drop functionality.
Release of micropython-wasm 0.1a0: an experimental package bundling a WASM build of MicroPython with a wasmtime wrapper to execute Python code in a sandboxed environment.
Hackers exploited Meta's AI support chatbot to gain access to high-profile Instagram accounts. By simply asking the bot to link a new email address to a target account, they bypassed the entire account recovery process. Meta had wired its support system to an AI capable of executing account changes in a single request.
Datasette 1.0a32 fixes a bug with INSERT ... RETURNING queries via the new /db/-/execute-write endpoint and multiple base_url issues found during Service Worker experiments.
Anthropic calculates its 'run-rate revenue' in two parts: last 28 days of consumption-based sales × 13, plus monthly subscriptions × 12. This metric, reported by Reuters, raises questions about actual revenue measurement.
Anthropic publishes detailed documentation on sandboxing techniques across Claude.ai, Claude Code, and Cowork. Uses gVisor (Claude.ai), Seatbelt/Bubblewrap (Claude Code local), and full VMs (Cowork). Includes process sandboxes, filesystem boundaries, and egress controls to prevent credential exfiltration.
Simon Willison used Claude Opus 4.8 via Claude Code to implement running Python ASGI apps in the browser via Pyodide and Service Workers. This approach replaces the previous Web Workers implementation, enabling JavaScript execution and fixing Datasette Lite limitations. Working demos are available.
Datasette 1.0a31 adds two major features: execution of write queries (INSERT/UPDATE/DELETE) and saving stored queries (private or shared). Permissions control access to sensitive operations like CREATE TABLE.
Anthropic reports annualized run-rate revenue of $47 billion as of May 2026, up from $9 billion end-2025. Growth accelerates: $14 billion in February, $30 billion in April. Metric disclosed during $65 billion Series H funding round.
Anthropic releases Claude Opus 4.8, described as a "modest but tangible improvement" over 4.7. The model excels in honesty: 4x less likely to let code flaws pass unremarked, and abstains more on uncertain questions. Pricing unchanged: $5/M input tokens, $25/M output.
Release of llm-anthropic 0.25.1: adds Claude Opus 4.8 model, -o fast 1 option for fast mode (enabled organizations), and default max_tokens now matches each model's maximum output instead of 8192.