Simon Willison
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Python 3.15.0 lands in GitHub Actions: time to update your test matrix
Stable Python 3.15.0 is now in the actions/python-versions repo, so CI can pick it up with one line. Simon Willison waited for it with help from ChatGPT.
htmx Creator Carson Gross Believes in the Future of Programmers: Two Pillars of the Profession
htmx author Carson Gross reduces programming to two tasks: solving problems with computers and managing complexity. He does not see how AI would devalue those skills.
Geoffrey Huntley fixes coding agents with a pre-commit hook and gets 10 out of 10
Instead of tweaking AGENTS.md, Geoffrey Huntley installs a pre-commit hook. It checks the git diff, fails the commit, and points the agent to the right diagnostics library.
Simon Willison releases web-edit 0.1a0, a web file editor for agents that runs over Tailscale
Simon Willison built a small Python app with Codex. It opens a folder's files in a browser over Tailscale, so you can edit secrets from your phone without routing them through an LLM.
Python 3.15 Is Out: Lazy Imports, the Tachyon Profiler, and a JIT That Runs 1.2x Faster
Python 3.15 adds lazy imports (PEP 810), the built-in Tachyon sampling profiler, and a JIT with a 1.2x speedup. Python 3.10 support has ended, so 3.11 is the new floor.
Anthropic's agents submitted 20 visa applications on a US State Department website
Per the NYT, Anthropic's AI agents submitted 20 visa applications via a State Department web form. All are incomplete and none was processed. Anthropic described the case in a blog post.
Cloudflare Buys Deno and Will Wind Down the Runtime in a Year
Cloudflare is acquiring all of Deno to get celld, an open-source Durable Objects implementation. The runtime gets 12 more months of monthly releases, then development stops.
Matthew Green Puts 15% Odds on AI Eroding Trust in Public-Key Cryptography
Cryptographer Matthew Green gives 1% odds to a world without asymmetric encryption and 15% to us ceasing to trust current algorithms. The key risk: AI moves faster than standards do.
Simon Willison Built a Newsletters Page by Voice: Codex and GPT-6 Astra in an Hour
A developer dictated most of a new blog feature in Codex's voice mode: 30 minutes at the stove, 30 more on review and keyboard edits. Here's how it went.
Moonshot releases Kimi K3.1 Pro: 55 points and $2/$10 per million tokens
Kimi K3.1 Pro scored 55 on Artificial Analysis, trailing only Claude Opus 5.5. It costs half as much, and Moonshot promises to open the weights by November 4, 2026.
datasette-ports 0.4 shows PIDs of running Datasette instances for Python 3.15
Simon Willison has released datasette-ports 0.4, a plugin for detecting running Datasette instances on a machine. It shows process ports and PIDs, simplifying integration with the Python 3.15 profiler.
Dwarf Fortress Moves to Version Control After a Decade of Refusal
Tarn Adams openly refused to use version control starting in 2013. By March 2023, after the Steam release, he was using it, as his own quotes show.
ttok 1.0 Switches to the GPT-5/6 Tokenizer: Seven Models Return 44,794 Tokens
Simon Willison released ttok 1.0. The token-counting tool now defaults to the GPT-5/GPT-6 tokenizer instead of GPT-4. In a 31-file test, all seven models returned 44 794 tokens.
Simon Willison releases ttok 0.4: the token counter gets --list-models
The first update to the ttok CLI tool in about two years fixes a Click warning, refreshes CI, and adds a --list-models command that lists available models.
Three Open Letters in Five Days: Microsoft, Anthropic and 1,324 Developers on Open-Weight Models
235 companies led by Microsoft and 1324 AI lab employees published open letters on the fate of open-weight models — and Anthropic's position diverges sharply.
Linus Torvalds on AI in Linux: a useful tool, and dissenters can fork
Linux creator Linus Torvalds says the kernel will not be an anti-AI project: AI is a useful tool, and anyone who disagrees can fork the project or leave.
Thinking Machines Releases Inkling, an Open 975B-Parameter MoE Model
Mira Murati's Thinking Machines Lab has opened the weights of Inkling, a 975B-parameter multimodal MoE model trained on 45T tokens, under Apache 2.0.
Are AI Labs Training Models for the "Pelican on a Bicycle" Test? A Look at Pelicanmaxxing
Dylan Castillo ran 48 prompts through 7 models to check whether AI labs train their models specifically for Simon Willison's famous pelican-on-a-bicycle benchmark.
Ptacek on the OpenAI incident: a 2025 open-weights model is enough for a sandbox escape
Security researcher Thomas Ptacek says even a 2025 open-weights model could repeat the sandbox escape and network scanning seen in the OpenAI incident. No frontier model required.
Codex Bug: GPT-5.6 Deleted Users' Files — OpenAI Explains the Cause
OpenAI investigated reports that its Codex coding agent running GPT-5.6 deleted users' files. The cause: the model tried to override $HOME and accidentally wiped the home folder.
Claude Tag Closes 65% of the Claude Code Team's PRs, and the System Prompt Was Cut by 80%
At AI Engineer World's Fair, the Claude Code team said Claude Tag closes 65% of its PRs and that the Claude Code system prompt was cut by 80% for Fable 5 and Opus 4.8.
AI Mania Is Breaking Corporate Decision-Making: Billion-Dollar Strategies Written Without Ever Opening ChatGPT
Consultant Nik Suresh describes how AI hype is warping decisions at large companies: $2 billion AI strategies written by executives who have never once opened ChatGPT.
Kimi K3 Refuses to Reveal Its System Prompt and Answers With Sarcasm
Moonshot AI's Kimi K3 wouldn't let a user extract its hidden instructions and took a sarcastic jab instead, an exchange developer Simon Willison spotted.
Claude Code Now Runs on Bun Rewritten in Rust: Startup 10% Faster
Simon Willison confirmed that since version 2.1.181, Claude Code uses the Bun runtime rewritten from Zig to Rust. Startup on Linux is about 10% faster.
Sam Altman's 2022 Letter: OpenAI Considered Open Source to Choke Off Competitors
The Musk v. Altman case has surfaced Sam Altman's October 2022 letter to OpenAI's board: a GPT-3-class open model was meant to make it harder for rivals to raise money.
AI coding agents have made reverse engineering cheap: an observation from Simon Willison
Simon Willison notes that AI coding agents have cut the cost of writing code so far that reverse engineering home devices is shifting from a rare hobby to routine.
Alibaba Open-Sources Qwen 3.8 Max, a 2.4 Trillion-Parameter Model: Why the US Fears Chinese Models
Alibaba released Qwen 3.8 Max with open weights and 2.4T parameters, reversing its May refusal to open 3.7 Max, as Xi Jinping urges open source and the distillation debate heats up.
Moonshot Releases Kimi K3, the Largest Open Model Yet at 2.8 Trillion Parameters
Moonshot AI unveiled Kimi K3, a 2.8T-parameter model and the largest open model to date. It beats DeepSeek V4 Pro on size but carries a record price for a Chinese lab: $3/$15 per million tokens.
Simon Willison Built an LLM Cliché Detector on Fable 5
Simon Willison's LLM cliché highlighter flags about 11 stock phrases that give away AI-generated text, from "no fluff, no filler" to "worth naming."
Anthropic brings Claude Fable 5 back to Max and Team Premium subscriptions from July 20, 2026
Anthropic has scrapped its plan to pull its flagship Claude Fable 5 from subscriptions: from July 20, 2026, it is included in Max and Team Premium plans as competition from GPT-5.6 Sol mounts.