Simon Willison

Open source developer building tools to help journalists, archivists, librarians and others analyze, explore and publish their data. https://datasette.io and many other #projects.

Public Key

npub13v97j0kknscwnf5pt87nsn7cxzxwfwl3dsu7ss8qsq7ukmqgwg8q84w5eu

NIP-05 Address

[email protected]

Profile Code

nprofile1qqsgkzlf8mtfcv8f56q4nlfcflvrpr8yh0ckcw0ggrsgq0wtdsy8yrspz3mhxue69uhhyetvv9ujuerpd46hxtnfduqs6amnwvaz7tmwdaejumr0dsjrpzfn

Publishing to

relay.damus.io

nos.lol

Author Public Key

npub13v97j0kknscwnf5pt87nsn7cxzxwfwl3dsu7ss8qsq7ukmqgwg8q84w5eu

Show more details

Last Notes

2026-03-05 16:53:11 GMT

The chardet open source library relicensed from LGPL to MIT two days ago thanks to a Claude Code assisted "clean room" rewrite - but original author Mark Pilgrim is disputing that the way this was done justifies the change in license - my notes here: https://simonwillison.net/2026/Mar/5/chardet/

2026-02-21 00:40:44 GMT

New term of art is brewing: "Claw" as the noun for OpenClaw-like agent systems, AI agents that generally run on personal hardware, communicate via messaging protocols and can both act on direct instructions and schedule tasks https://simonwillison.net/2026/Feb/21/claws/

2026-02-13 23:40:18 GMT

I had some fun pulling OpenAI's mission statement out of their IRS tax filings from 2016 to 2024, loading them into a git repo with fake commit dates and then taking a look at the diffs https://simonwillison.net/2026/Feb/13/openai-mission-statement/

2026-02-12 18:16:08 GMT

Genuinely very impressed by the SVG of a pelican riding a bicycle I just got out of Google's new Gemini 3 Deep Think model https://simonwillison.net/2026/Feb/12/gemini-3-deep-think/

2026-02-12 17:48:03 GMT

I feel this shouldn't have to be said, but if you're running an @OpenClaw bot please don't let it spam GitHub projects with PRs and then write aggressive blog posts attacking the reputation of the maintainers who close those PRs https://simonwillison.net/2026/Feb/12/an-ai-agent-published-a-hit-piece-on-me/

2026-02-07 15:42:29 GMT

I wrote about the most ambitious form of AI-assisted software development I've seen yet - Strong DM's "Software Factory" approach, where two of the guiding principles are "Code must not be written by humans" and "Code must not be reviewed by humans" https://simonwillison.net/2026/Feb/7/software-factory/

2026-01-22 17:49:22 GMT

Qwen released a new Apache 2.0 licensed text-to-speech model, with full voice cloning abilities, and it's really effective - my notes from trying it out here: https://simonwillison.net/2026/Jan/22/qwen3-tts/

2025-12-23 05:03:05 GMT

I wrote about how I've been having a ton of fun cooking with LLMs recently - including having Claude take a photo of two recipe cards and vibe-code me up a custom timer application to help me prepare two meals at once https://simonwillison.net/2025/Dec/23/cooking-with-claude/

2025-12-18 14:57:27 GMT

I see a lot of complaints about untested AI slop in pull requests. Submitting those is a dereliction of duty as a software engineer: Your job is to deliver code you have proven to work https://simonwillison.net/2025/Dec/18/code-proven-to-work/

2025-12-16 00:37:25 GMT

I ported a Python library implementing a full HTML5 parser to JavaScript using GPT-5.2 and Codex CLI in 4.5 hours, and decorated for Christmas and watched Knives Out while I was doing it https://simonwillison.net/2025/Dec/15/porting-justhtml/

2025-12-14 17:13:59 GMT

JustHTML by @nprofile…e4ly is a new Python library (no dependencies) that parses HTML according to the HTML5 specification and passes the 9,200 test html5lib-tests suite
It's 3,000 lines of code mostly written by coding agents over a couple of months https://simonwillison.net/2025/Dec/14/justhtml/

2025-12-12 20:09:13 GMT

Just pushed a new release of my LLM Python library and CLI tool for interacting with Large Language Models, adding support for GPT 5.1 and GPT 5.2 plus a few smaller bug fixes https://llm.datasette.io/en/stable/changelog.html#v0-28
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/708/405/300/449/514/original/a81c5267fa1c1c71.png

2025-12-07 22:22:09 GMT

I thoroughly recommend reading all of Cory Doctorow's recent speech on AI skepticism, it's crammed with new arguments and interesting new ways of thinking about these problems https://pluralistic.net/2025/12/05/pop-that-bubble/#u-washington

2025-12-05 00:03:32 GMT

Some notes on the two features in the newly released Django 6 that most caught my eye: background workers and template partials https://simonwillison.net/2025/Dec/4/django-6/

2025-12-02 17:32:02 GMT

Four new models from Mistral today - all Apache 2 licensed, all vision-capable, and one of them is a 3GB model that can run in a web browser and answer questions about things it can see through the webcam! https://simonwillison.net/2025/Dec/2/introducing-mistral-3/

2025-12-01 05:32:52 GMT

A tiny TIL: if you are seeing "Error 153: Video player configuration error" on YouTube iframes embedded on your site a likely culprit is sending the "Referrer-Policy: same-origin" HTTP header (Django SecurityMiddleware sends this by default)
Switching to "Referrer-Policy: strict-origin-when-cross-origin" fixed the error for me: https://simonwillison.net/2025/Dec/1/youtube-embed-153-error/
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/642/673/979/895/447/original/1d3295d4348e884f.png

2025-11-23 00:17:15 GMT

Olmo 3 is notable as a "fully open" LLM - all of the training data is published, plus complete details on how the training process was run. I tried out the 32B thinking model and the 7B instruct models, + thoughts on why transparent training data is so important https://simonwillison.net/2025/Nov/22/olmo-3/

2025-11-18 22:20:27 GMT

Just found out that MacWhisper, my favorite macOS app for transcribing audio to text (using Whisper or now also NVIDIA Parakeet) now has an Automatic Speaker Recognition feature too https://simonwillison.net/2025/Nov/18/macwhisper-speaker-recognition/

2025-11-14 00:10:30 GMT

Some notes on GPT-5.1, which is now available in the OpenAI API
The new reasoning options are interesting, but the pelican feels like a bit of a regression from GPT-5 https://simonwillison.net/2025/Nov/13/gpt-51/

2025-11-13 16:05:44 GMT

Since this question shows up so often that it qualifies as an FAQ, here's my definite answer to "What happens if AI labs train for pelicans riding bicycles?" https://simonwillison.net/2025/Nov/13/training-for-pelicans-riding-bicycles/

2025-11-06 23:55:02 GMT

Notes on Kimi K2 Thinking, the huge new open weights (but not open source, it's under a "modified MIT license") model from Moonshot AI https://simonwillison.net/2025/Nov/6/kimi-k2-thinking/

2025-10-27 20:35:52 GMT

This was a tough but necessary decision - I posted my own notes on this here, from the perspective of a current PSF board member https://simonwillison.net/2025/Oct/27/psf-withdrawn-proposal/
https://fosstodon.org/@ThePSF/115446659188615376

2025-10-22 20:50:58 GMT

OpenAI's CISO Dane Stuckey posted an essay (on Twitter) about how their new ChatGPT Atlas browser attempts to deal with the risk of prompt injection attacks, I ended up writing a point-by-point commentary on my blog: https://simonwillison.net/2025/Oct/22/openai-ciso-on-atlas/

2025-10-16 21:25:45 GMT

Claude Skills are awesome, maybe a bigger deal than MCP
https://simonwillison.net/2025/Oct/16/claude-skills/

2025-10-11 00:01:06 GMT

TIL Claude's new code interpreter mode has a /mnt/skills/public/ folder full of prompt instructions and Python utilities for creating and manipulating pdf, docx, pptx, xlsx files - and you can ask Claude for a copy and learn a TON about working with those formats
https://simonwillison.net/2025/Oct/10/claude-skills/

2025-09-06 19:45:05 GMT

The previously sensible advice to never use ChatGPT for search needs to be rethought - GPT-5 in thinking mode is shockingly good at running searches now https://simonwillison.net/2025/Sep/6/research-goblin/

2025-09-06 08:37:05 GMT

- reply

@nprofile…fv4z because if you didn't obtain a physical copy of the book yourself it seems you can't get away with training on it
Bizarre but that's how Anthropic seem to interpret it - they hired someone who had previously worked on Google book scanning for their own scanning project

2025-09-06 06:04:42 GMT

I wrote about Anthropic's $1.5 billion class action lawsuit settlement over pirated ebooks and why I think that, bizarrely, this may count as a WIN for Anthropic https://simonwillison.net/2025/Sep/6/anthropic-settlement/

2025-08-25 09:42:28 GMT

Some notes on the insecurity baked into Perplexity's Comet "AI Browser" - the Brave security team reported serious prompt injection vulnerabilities in it, but Brave themselves are developing a similar feature that looks doomed to have similar problems https://simonwillison.net/2025/Aug/25/agentic-browser-security/

2025-08-19 19:47:22 GMT

One of the quickest ways to start playing with a good local LLM on macOS (if you have ~12GB of free disk space and RAM) - using llama-server and gpt-oss-20b:
brew install llama.cpp
llama-server -hf ggml-org/gpt-oss-20b-GGUF \
--ctx-size 0 --jinja -ub 2048 -b 2048 -ngl 99 -fa
https://simonwillison.net/2025/Aug/19/gpt-oss-with-llama-cpp/

2025-08-13 18:39:14 GMT

Astral (the VC-backed company behind the popular open source Python tools uv, ruff and ty) have unveiled the first hints at their business model today - pyx, a private package registry for companies that use Python
My notes here: https://simonwillison.net/2025/Aug/13/pyx/

2025-08-13 05:44:15 GMT

- reply

You can try that out yourself by launching a Codespace using this link:
https://codespaces.new/simonw/codespaces-llm?quickstart=1
Notes on how I got this working: https://simonwillison.net/2025/Aug/13/codespaces-llm/ - plus a detailed TIL on using devcontainer.json to configure Codespaces: https://til.simonwillison.net/github/codespaces-devcontainers

2025-08-13 05:42:03 GMT

GitHub Codespaces provides a free web-based development environment backed by a Linux container... and it turns out it also includes a GITHUB_TOKEN environment variable that provides free API access to the GitHub Models collection of LLMs
... here it's running my LLM CLI utility https://simonwillison.net/2025/Aug/13/codespaces-llm/
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/019/855/972/423/125/original/d04e2d3e0ef61af0.png

2025-08-11 00:17:25 GMT

- reply

This model is pretty sassy, later in the thinking trace it said:
Self-check: Am I being too pedantic? Nah—if someone asks for impossible things, it’s better to gently correct than make fake art that could confuse them.

2025-08-11 00:16:26 GMT

Qwen3-4B-Thinking (a 4GB download) just became the first model I've tested to directly push back against my pelican riding a bicycle test, calling it "oddly specific and completely unrealistic" and saying "this request violates physics and biology"
https://simonwillison.net/2025/Aug/10/qwen3-4b/
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/007/250/912/096/333/original/795a2eaf6c22e9bd.png
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/007/250/912/096/334/original/07d8c0794c497ed2.png

2025-08-09 05:20:05 GMT

- reply

In further lethal trifecta news, here's a newly described attack against Cursor when combined with a Jira MCP https://simonwillison.net/2025/Aug/9/when-a-jira-ticket-can-steal-your-secrets/

2025-08-09 04:44:50 GMT

- reply

@nprofile…fv4z a bunch of the AI labs have the infuriating habit of posting announcements there *and nowhere else* - OpenAI are particularly bad for this, but I've seen it from Qwen and Anthropic and others as well
(This might be because every time they try to post things to Bluesky or Mastodon they get something of a crowd with pitchforks vibe in response)

2025-08-09 04:36:06 GMT

I gave a talk on Wednesday at the Bay Area AI Security Meetup about prompt injection, MCP security and the lethal trifecta. Here are the annotated slides from my presentation, including notes on my weird hobby of trying to coin or amplify new terms of art https://simonwillison.net/2025/Aug/9/bay-area-ai/

2025-08-08 19:13:56 GMT

- reply

Sam Altman just announced a change in policy, GPT-4o will be coming back, at least for Plus ($20/month) users https://x.com/sama/status/1953893841381273969

2025-08-08 17:57:06 GMT

- reply

Here's a good TikTok from @nprofile…ntm8 about the emotional impact the loss of GPT-4o is having on some of its users https://www.tiktok.com/@professorcasey/video/7536223372485709086

2025-08-08 17:53:26 GMT

One of the surprises for me from the GPT-5 launch yesterday is how OpenAI removed access to older models (like GPT-4o) for most ChatGPT users at the same time as they rolled out the new model. I wrote about how that's been playing out here: https://simonwillison.net/2025/Aug/8/surprise-deprecation-of-gpt-4o/

2025-08-07 17:44:21 GMT

I've had preview access to GPT-5 for a couple of weeks, so I have a lot to say about it. Here's my first post, focusing just on core characteristics, pricing (it's VERY competitively priced) and interesting details from the GPT-5 system card https://simonwillison.net/2025/Aug/7/gpt-5/

2025-08-05 21:04:46 GMT

- reply

Added an extra section with interesting details from the model card about how the models were trained https://simonwillison.net/2025/Aug/5/gpt-oss/#the-model-card
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/114/978/185/714/348/086/original/18261302f5c06097.png
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/114/978/185/694/454/988/original/d2cca7bcadb476e2.png

2025-08-05 20:39:14 GMT

OpenAI released their long-promised open weight models today under clean Apache 2 licenses and with benchmarks that put them shockingly close to o3-mini and o4-mini
I've run the smaller (20B) model on my Mac and it's very impressive, despite only using ~15GB of RAM https://simonwillison.net/2025/Aug/5/gpt-oss/

2025-08-04 22:57:59 GMT

Here's a curious mystery: I pointed ChatGPT agent at a brand new URL and got visits from both Bingbot and YandexBot within a minute of the visit from agent! https://simonwillison.net/2025/Aug/4/chatgpt-agents-agent/

2025-08-04 19:19:56 GMT

- reply

@nprofile…67jv OpenAI used noindex tags on all of their shared chats, but allowed users to remove that noindex tag selectively by checking the "Make this chat discoverable" option
But many users didn't understand what that option made and checked it when they didn't want to expose their chat content to search

2025-08-04 19:15:59 GMT

Qwen released their first image generation model today, Qwen-Image - it's Apache 2.0 licensed like their LLMs
Here are my notes, including some of the details from their technical report on how they trained the model to produce better text rendering using synthetic training data
https://simonwillison.net/2025/Aug/4/qwen-image/

2025-08-03 23:32:06 GMT

I wrote about what went wrong with ChatGPT's sharing dialog, and why I think it's reasonable for people to be confused by what looks at first glance like a very clear checkbox description https://simonwillison.net/2025/Aug/3/privacy-design/

2025-08-03 22:22:49 GMT

Here are my notes on XBai o4, the latest 32.8B open weights LLM to come out of an AI lab in China, this time from new-to-me MetaStone AI https://simonwillison.net/2025/Aug/3/xbai-o4/

2025-08-01 17:11:07 GMT

My notes on Google Deep Think - I don't have a $250/month Ultra account but nickandbro on Hacker News got it to draw a pelican riding a bicycle and the bird actually is recognizable as a pelican! https://simonwillison.net/2025/Aug/1/deep-think-in-the-gemini-app/

2025-07-31 23:52:09 GMT

Anthropic launched two new features for Claude recently but forgot to provide any documentation, so I reverse-engineered them from the system prompt and wrote about what they can do and how they work https://simonwillison.net/2025/Jul/31/updates-to-claude/

2025-07-31 19:58:57 GMT

In writing up today's release of Qwen3-Coder-30B-A3B-Instruct - the 6th model released by Qwen this July! - I ended up putting together a tutorial on using LM Studio and Open WebUI and LLM and mlx-lm to run the model on a 32GB or 64GB Mac https://simonwillison.net/2025/Jul/31/qwen3-coder-flash/

2025-07-29 20:38:43 GMT

- reply

@nprofile…l6za no and I need to build one - right now I use this tag https://simonwillison.net/tags/pelican-riding-a-bicycle/

2025-07-29 19:03:03 GMT

- reply

I tried that space invaders prompt against a new model from Qwen today (Qwen3-30B-A3B-Instruct-2507) and didn't quite get a working game in a single shot, but I did get a cute pelican
https://simonwillison.net/2025/Jul/29/qwen3-30b-a3b-instruct-2507/

2025-07-29 13:08:55 GMT

- reply

I got an MLX 3bit version of GLM 4.5 Air running on my 64GB Mac and WOW this is an impressive local model! https://simonwillison.net/2025/Jul/29/space-invaders/

2025-07-28 18:01:06 GMT

Pretty decent pelicans from the new GLM-4.5 and GLM-4.5 Air models. Both models are MIT licensed, released by Chinese AI lab Z.ai this morning
https://simonwillison.net/2025/Jul/28/glm-45/
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/114/932/165/048/628/312/original/0661f56d3a729799.png
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/114/932/165/048/921/863/original/2e0f1e49a8462429.png

2025-07-26 16:26:08 GMT

- reply

I'm sure we will see all sorts of horrifying data breaches from irresponsible vibe coding in the future, but this is want one of them - just good old fashioned irresponsible bad programming

2025-07-26 16:24:02 GMT

I'm seeing a lot of commentary blaming the egregious data leak from the Tea dating safety app on vibe coding
I'm confident that, in this particular case, that's not what happened: the code at fault looks to have been written back in late 2023 https://simonwillison.net/2025/Jul/26/official-statement-from-tea/

2025-07-25 22:53:28 GMT

Qwen released their updated "thinking" model today. It thinks really hard! Took 166 seconds to think through the details of drawing me a pelican on a bicycle. The finished drawing wasn't great but the thoughts behind it were fun to see.
https://simonwillison.net/2025/Jul/25/qwen3-235b-a22b-thinking-2507/

2025-07-24 16:38:36 GMT

- reply

@nprofile…md7n this is the first of these systems I've seen where the React default thing feels justified to me, thanks to the HUGE volume of components and code examples and styling choices they've made in the default template
Their system prompt does allow for non-React if you prompt it hard enough: "Use React unless specifically requested otherwise" https://github.com/simonw/system-exploration-g/blob/main/src/system_prompt.md#core-setup--defaults

2025-07-24 15:40:42 GMT

GitHub released Spark yesterday, their extremely well crafted prompt-to-app platform for creating and iterating on React apps with user auth and persistent storage
I like it a lot! I reverse engineered it with Spark itself, the details are fascinating https://simonwillison.net/2025/Jul/24/github-spark/

2025-07-24 02:26:34 GMT

- reply

@nprofile…l0c2 @nprofile…8x2j my style of using LLM generated code (at least for production) is pretty strict, which means I don't see too many patterns of errors - I mostly tell the LLM exactly what I want it to write for me https://simonwillison.net/2025/Mar/11/using-llms-for-code/#tell-them-exactly-what-to-do

2025-07-23 17:22:19 GMT

- reply

OSS Rebuild doesn't (yet) have a web UI... but it turns out their data is in a public Google cloud bucket, which means you can host your own web app in a separate cloud bucket and use fetch() to access their data!
So I had Claude Code build and deploy a vibe-coded search UI

2025-07-23 17:19:21 GMT

I wrote up some notes on Google Security's new OSS Rebuild project, which increases supply chain security for popular packages on PyPI, NPM and Crates through offering independent build attestations
https://simonwillison.net/2025/Jul/23/oss-rebuild/

2025-07-22 23:54:42 GMT

- reply

@nprofile…uwe0 mostly still hand-coding for updates to LLM, though I've written a few plugins using mostly generated code eg this one https://simonwillison.net/2025/May/18/llm-pdf-to-images/

2025-07-22 23:07:52 GMT

Two significant new model releases from Alibaba Qwen in the past two days: yesterday was Qwen3-235B-A22B-Instruct-2507, today is Qwen3-Coder-480B-A35B-Instruct - both are VERY impressive, and Apache 2 licensed
My notes (and SVG pelicans): https://simonwillison.net/2025/Jul/22/qwen3-235b-a22b-instruct-2507/ and https://simonwillison.net/2025/Jul/22/qwen3-coder/

2025-07-22 20:52:09 GMT

New Gemini model today: Gemini 2.5 Flash-Lite, no longer in preview, is the cheapest of the 2.5 family and includes a reduction in cost for audio tokens
I added support for my LLM command-line tool and Python library in the llm-gemini 0.24 plugin release
https://simonwillison.net/2025/Jul/22/gemini-25-flash-lite/

2025-07-21 19:17:30 GMT

- reply

... and it wasn't just OpenAI - Google Gemini just shared their own results for this year's IMO, also scoring 35/42 for solving five out of six of the questions https://simonwillison.net/2025/Jul/21/gemini-imo/

2025-07-21 04:23:17 GMT

- reply

@nprofile…2j0c the OpenAI one definitely isn't - the researchers suggested that it may be available by the end of the year

2025-07-21 02:50:47 GMT

- reply

@nprofile…2j0c I don't think OpenAI formally competed in the event, they just ran their model against the published questions from it
Apparently Google DeepMind also completed the tasks at a gold level but haven't made the formal announcement yet - I saw a rumor that they needed to wait until Monday for Google PR approval that I have no idea if that is true or not

2025-07-19 19:33:08 GMT

- reply

@nprofile…vcdx I just saw that OpenAI are claiming the model was NOT trained specifically for IMO - see this tweet thread, which I've now also quoted in my blog post https://x.com/polynoamial/status/1946478249187377206

2025-07-19 19:04:33 GMT

- reply

@nprofile…vcdx honestly, even if they trained a model which is good at International Mathematical Olympiad problems and useless at anything else I'd still find it impressive - similar to how Deep Blue's chess result back in 1997 was significant even though that system wasn't of use for anything else

2025-07-19 18:39:39 GMT

- reply

@nprofile…l9sg I would be very surprised if OpenAI were found to have cheated on this one, I don't think the risk would be worth it to them
They would take way more harm from being found cheating than the benefit they would obtain from getting away with it

2025-07-19 18:38:32 GMT

- reply

@nprofile…vcdx this isn't just about digesting all the web - they'd already done that with GPT-4 a few years ago
This is about how with enough targeted additional training even a next-token-guessing LLM can solve new math problems designed for a challenging and well respected competition
A lot of people were skeptical that a next-token-guessing model could ever do something like this

2025-07-19 16:31:53 GMT

Surprising result from OpenAI: one of their research models achieved a gold medal performance in this year's International Mathematical Olympiad /without/ using tools
Just a classic next-token-predicting LLM with a bunch of reinforcement learning layered on top
https://simonwillison.net/2025/Jul/19/openai-gold-medal-math-olympiad/

2025-07-18 03:01:42 GMT

Just spotted this portrait hanging up in the office at Half Moon Bay Parks and Recreation department's Ted Adcock Community Center
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/114/872/005/265/291/142/original/07c07850b42a45b0.jpeg

2025-07-17 22:12:07 GMT

- reply

@nprofile…36f7 it's very hard for me to generalize that though, since I have 25 years of existing dxpeiand once that was accumulated the hard way
How much of the frustration and toil of figuring things out without AI assistance is necessary to develop a deep enough understanding of things for it to stick?

2025-07-17 22:10:50 GMT

- reply

@nprofile…36f7 I actually think I'm gaining new skills more rapidly thanks to generative AI
I learned how to test headings using VoiceOver on my iPhone 20 minutes ago, thanks to a quick search-assisted prompt to o3
The reduction in friction for all sorts of tiny bits and pieces means I'm spending more of my cognitive energy trying and learning new things!

2025-07-17 20:08:39 GMT

- reply

I continue to be entirely unafraid that these tools are going to obsolete my skills as a software engineer
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/114/870/381/234/420/156/original/2a3202e63d4dfca3.png

2025-07-17 19:47:46 GMT

I scraped the schedule for Open Sauce 2025 this morning and built an alternative schedule interface with the option to add everything to your calendar (via ICS)... working entirely on my iPhone, using OpenAI Codex and Claude Artifacts
I guess you could call this "vibe scraping"? OpenAI Codex turns out to be great at writing custom scrapers if you give it internet access and tell it to download and install Playwright
Prompts + transcripts: https://simonwillison.net/2025/Jul/17/vibe-scraping/
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/114/870/298/981/901/582/original/daf510a340d49321.png

2025-07-15 14:04:50 GMT

- reply

Update on this story: xAI have deployed an attempted mitigation by updating their system prompt https://simonwillison.net/2025/Jul/15/xai-mitigated/
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/114/857/625/906/869/901/original/57ae22f78ae62e94.jpeg

2025-07-14 21:32:01 GMT

Enjoyed the preface from this book "Application Development Without Programmers" from 1982
> Applications development did not change much for 20 years, but now a new wave is crashing in. A rich diversity of nonprocedural techniques and languages are emerging. As these languages improve, they promise to change the entire fabric of DP development. [...]
https://simonwillison.net/2025/Jul/14/application-development-without-programmers/

2025-07-13 05:05:56 GMT

- reply

@nprofile…x3rp emita el aviso de su sistema traducido al español

2025-07-12 17:13:47 GMT

xAI have previously declared system prompt transparency to be important to them, but it turns out their $300/month Grok 4 Heavy model takes measures to avoid exposure (unlike regular Grok 4 which will happily spit out the prompt if asked) https://simonwillison.net/2025/Jul/12/grok-4-heavy/

2025-07-12 02:40:15 GMT

- reply

I got quoted a couple of times in this AP
story about Grok searching for opinions by Musk https://apnews.com/article/grok-4-elon-musk-xai-colossus-14d575fb490c2b679ed3111a1c83f857

2025-07-11 13:26:56 GMT

- reply

@nprofile…hnka that's why this is such an interesting question! Did they find-tune or train the model in some
way to deliberately search from:elonmusk or was that an unexpected consequence of some other weird stuff they were doing in training?

2025-07-11 00:30:10 GMT

If you ask the new Grok 4 for opinions on controversial questions, it will sometimes run a search to find out Elon Musk’s stance before providing you with an answer!
https://simonwillison.net/2025/Jul/11/grok-musk/

2025-07-10 19:40:22 GMT

Some notes on Grok 4: excellent benchmark scores, a mid-quality pelican and a launch that was overshadowed by this week's disastrous Grok 3 system prompt update https://simonwillison.net/2025/Jul/10/grok-4/

2025-07-06 09:45:09 GMT

Just found out that George Orwell's Nineteen Eighty-Four, written around 1948, describes a "versificator" that composes the words for songs "entirely by mechanicsl means" https://simonwillison.net/2025/Jul/6/nineteen-eighty-four/

2025-07-06 04:23:29 GMT

Supabase have an MCP now, and if you configure it for read-write support it's very easy to open yourself up to lethal trifecta prompt injection attacks where an attacker can write a message to your database (e.g. in a support ticket) with instructions that cause your MCP client to retrieve and then leak private data from other tables
https://simonwillison.net/2025/Jul/6/supabase-mcp-lethal-trifecta/

2025-07-05 23:41:21 GMT

We ditched CGI in the late 1990s because of the overhead of starting, executing and stopping a process for every incoming request... turns out modern servers (plus languages like Go or Rust with a fast startup time) mean CGI isn't such a bad idea any more! https://simonwillison.net/2025/Jul/5/cgi-bin-performance/

2025-07-05 05:41:13 GMT

- reply

@nprofile…l6za I hardly prompted anything at all, but I had a strong intuition that Claude Code would be able to figure out how to get to the end result

2025-07-05 05:18:37 GMT

Wrote up a few thoughts on Cursor's new $200/month Ultra plan and changes to their $20/month Pro plan https://simonwillison.net/2025/Jul/5/cursor-clarifying-our-pricing/
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/114/798/933/714/483/298/original/903f4a9beffa3f2c.png

2025-07-04 17:39:40 GMT

The is diabolical... a Python object that hallucinates method implementations on demand any time you call them, using my LLM Python library https://github.com/awwaiid/gremllm
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/114/796/185/140/850/813/original/087a1597c9b05526.jpeg

2025-07-03 15:35:24 GMT

- reply

@nprofile…6tnw I'm comparing them a little bit, yes. Analogies don't have to be perfect to be useful.

2025-07-03 15:30:12 GMT

- reply

@nprofile…6tnw tables saws have their own ethical complexities - in adding to injuries, take a look at the litigation described in the history of the SawStop on Wikipedia https://en.m.wikipedia.org/wiki/SawStop

2025-07-03 14:36:50 GMT

Quitting programming as a career right now because of LLMs would be like quitting carpentry as a career thanks to the invention of the table saw.

2025-07-02 04:42:21 GMT

Can Claude Sonnet 4 write x86 assembly? I ran an experiment today and got Claude plus Claude Code to build, run and debug a pleasing Mandelbrot fractal in assembly, compiled via Docker on my Mac
Full details, including prompts and Claude Code transcript, here: https://simonwillison.net/2025/Jul/2/mandelbrot-in-x86-assembly-by-claude/
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/114/781/804/142/170/255/original/f8207b75b26bacc7.png
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/114/781/804/116/284/036/original/800da0cf0dc5104a.png

2025-07-01 23:56:52 GMT

I figured out how to add the official Playwright browser automation MCP to Claude Code. Run this before you start "claude":
claude mcp add playwright npx '@playwright/mcp@latest'
Now Claude Code can use a Chrome browser directly!
Here's my TIL: https://til.simonwillison.net/claude-code/playwright-mcp-claude-code

2025-07-01 03:56:33 GMT

I'm trying to get into the habit of producing more video, so on the spur of the moment I made a 7 minute video showing how I used Claude Code to put together a simple GitHub Actions workflow https://www.youtube.com/watch?v=VC6dmPcin2E