Simon Willison

https://fedi.simonwillison.net/@simon

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.

Blog: https://simonwillison.net
GitHub: https://simonw.github.io
Twitter: https://twitter.com/simonw
TIL: https://til.simonwillison.net

Public Key

npub1gv26rplqyjqcfyhxryuqjwaxs0dwve3y6gpv6smn3hjm3wse3s8squtlwl

NIP-05 Address

[email protected]

Profile Code

nprofile1qqsyx9dpslszfqvyjtnpjwqf8wng8khxvcjdyqkdgdecmedchgvccrcpz3mhxue69uhhyetvv9ujuerpd46hxtnfduqs6amnwvaz7tmwdaejumr0ds0kg6v4

Publishing to

relay.damus.io

nos.lol

Author Public Key

npub1gv26rplqyjqcfyhxryuqjwaxs0dwve3y6gpv6smn3hjm3wse3s8squtlwl

Show more details

Last Notes

2026-04-03 14:55:37 GMT

Warning to open source maintainers: the Axios supply chain attack started with some
very sophisticated social engineering targeted at one of their developers https://simonwillison.net/2026/Apr/3/supply-chain-social-engineering/

2026-02-27 18:02:38 GMT

Built a fun prototype this morning of binary search using HTTP range requests, in this case to lookup characters in ~77MB of unicode data https://simonwillison.net/2026/Feb/27/unicode-explorer/
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/116/143/905/497/245/336/original/8dd8d900e43105ff.mp4

2025-12-18 14:59:35 GMT

- reply

someone else benchmarked and found it to be quite a lot faster https://github.com/kylehowells/swift-justhtml/blob/master/Benchmarks/BENCHMARK_RESULTS.md#summary

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/

2025-12-17 23:37:49 GMT

- reply

I'm very cautious about submitting AI-generated PRs to any other projects though - I need to be confident that it's not against their policies and also that I'm not introducing 100 lines of code that someone else then has to actively work to understand
I'm more comfortable submitting a dozen or so lines that I've reviewed extensively and which I could have equally written without any AI help at all

2025-12-17 23:36:37 GMT

- reply

I'm weird though because I've reached the "I don't type any code myself" phase - everything I do is "AI-assisted" in that an LLM made the edits to the files, but often those edits are just a few dozen characters and I've designed and dictated exactly what they should look like
Like using the LLM as a partner who types everything for me based on my direct pair-programming guidance

2025-12-17 23:35:19 GMT

- reply

there are definitely many others who have thought about this a lot more than me - my current opinion is that there should be no difference between an AI and non-AI-assisted contribution in that the contributor needs to be accountable for every line of it, and needs to be able to explain and justify all of it
The thing to avoid is anyonee submitting code that they haven't tested or that they don't 100% understand - that's where it's an unfair burden on the maintainers in my opinion

2025-12-17 23:25:38 GMT

- reply

I guess we're calling this "vibespiling" now! https://simonwillison.net/2025/Dec/17/vibespiling/

2025-12-17 22:45:59 GMT

Here's my article on Gemini 3 Flash, with four pelicans (one for each of its thinking effort levels) plus an image gallery Web Component that I had Gemini 3 Flash build specially for this post https://simonwillison.net/2025/Dec/17/gemini-3-flash/

2025-12-17 00:07:55 GMT

Blogged some notes on the new ChatGPT Images model that launched today https://simonwillison.net/2025/Dec/16/new-chatgpt-images/

2025-12-16 03:09:55 GMT

- reply

in this case the project is almost the perfect case for coding agent assistance - it starts with a robust and complete test suite, which is incredibly rare for a project
Emil talks about that in some depth in his post here https://friendlybit.com/python/writing-justhtml-with-coding-agents/#why-html5

2025-12-16 03:07:10 GMT

- reply

that's more a quirk of the parsing library itself - it's just doing the tokenizing and tree-building part, it doesn't get as far as treating CSS or JavaScript differently from other forms of text

2025-12-16 03:06:08 GMT

- reply

one of my great frustrations with this stuff is how hard it is to come up with examples that clearly show one model beating another!
For Opus it's more about how often I find myself nodding in agreement with the result of a requested change - I find it has taste in Python and JavaScript that matches my own really closely
But maybe GPT-5.2 has caught up on that regard? It takes weeks of active use to develop those intuitions to any level of confidence

2025-12-16 00:40:37 GMT

- reply

I still like Claude Opus 4.5 the best, but GPT-5.2 is clearly just as capable and I've not spent enough time with Gemini Pro 3 to decide one way or the other

2025-12-16 00:39:50 GMT

- reply

I'm sharing this because it's quite a stark demonstration of what's possible using the latest frontier LLMs at the tail end of 2025
Is it a GOOD idea to churn out a 9,000 line direct port while watching Daniel Craig chew on the scenery? That's a whole other set of questions!

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 17:51:34 GMT

- reply

if the model was just repeating what it saw in its training data it wouldn't have taken two months and 1,193 commits to build the new library
What's clever here is running a coding agent in a loop against 9,200 existing tests to brute force the creation of a new library that passes all of them

2025-12-14 17:16:50 GMT

- reply

Here's Emil's 17 step guide to how he used VS Code agent mode plus Claude 3.7 Sonnet, Gemini Pro 3 and Claude Opus to build the new library - it's a fantastic case study in using LLMs for serious, prediction quality code (vibe engineering, not vibe coding) https://friendlybit.com/python/writing-justhtml-with-coding-agents/

2025-12-14 17:13:59 GMT

JustHTML by @npub1ukc…r8n2 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 23:32:18 GMT

OpenAI aren't talking about it yet, but it turns out they've adopted Anthropic's brilliant "skills" mechanism in a big way
Skills are now live in both ChatGPT and their Codex CLI tool, I wrote up some detailed notes on how they work so far here: https://simonwillison.net/2025/Dec/12/openai-skills/

2025-12-12 20:10:56 GMT

- reply

I've been upgrading all of my projects to use pyproject.toml and dependency groups for cleaner compatibility with uv, which means you can check them out and run the test suite like this:
cd /tmp
git clone http''s://github.com/simonw/llm
cd llm
uv run pytest
https://til.simonwillison.net/uv/dependency-groups

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

2025-12-12 03:25:00 GMT

- reply

apparently free users get 10 GPT 5.2 messages every five hours and then it goes back to "mini" (presumably GPT-5.1 mini since they didn't announce a 5.2 mini model) https://help.openai.com/en/articles/11909943-gpt-52-in-chatgpt
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/704/456/354/712/362/original/36174141c29a1bd5.jpeg

2025-12-12 00:06:55 GMT

Initial notes on GPT-5.2 - I'm excited to see the knowledge cut-off bump up to August 2025, interesting that it's priced at 1.4x over GPT-5.1 https://simonwillison.net/2025/Dec/11/gpt-52/

2025-12-10 22:17:15 GMT

- reply

I made that because now if a Python package doesn't have good release notes I can click "copy diff" and paste into Claude/ChatGPT/Gemini and get genuinely useful release notes a few seconds later
No need to even add a prompt, the models guess that's what you want from the diff
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/697/583/486/757/895/original/d228d4091b37923e.jpeg
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/697/584/001/153/309/original/41dc50612c8c5143.jpeg

2025-12-10 21:59:45 GMT

- reply

My most recent HTML tool is this one for loading code from the Python packaging index and displaying a diff between two versions - all running entirely in the browser, created using this prompt (which remixed a previous project) https://tools.simonwillison.net/pypi-changelog?package=llm&compare=0.27...0.27.1
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/697/514/716/099/226/original/01612be22d47c83b.jpeg
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/697/515/206/052/454/original/1dc3e5253ca1e02d.jpeg

2025-12-10 21:11:58 GMT

- reply

Here's the table of contents, - by patterns I'm talking about things like hitting CORS-enabled APIs, using localStorage and URLs to store state, loading dependencies from CDNs and taking extensive advantage of rich copy and paste for both input and output to the tools you build
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/697/327/527/463/881/original/72168b1bc69fe621.png

2025-12-10 21:08:25 GMT

I put together a detailed collection of useful patterns I've collected after vibe-coding 150 different single-file HTML tools over the past couple of years https://simonwillison.net/2025/Dec/10/html-tools/

2025-12-10 16:16:27 GMT

I've never been particularly invested in light mode v.s. dark mode but I got enough complaints from people that my blog was "blinding" that I had Claude Code knock out a dark mode theme for it, and it seems to work pretty well https://simonwillison.net/2025/Dec/10/dark-mode/

2025-12-09 20:22:05 GMT

Put together some notes on Mistral Vibe, Mistral's new CLI coding agent (think Claude Code / OpenAI Codex) which is Apache 2 licensed, written in Python+Pydantic+Textual and has a neat set of system and tool definition prompts that are fun to read https://simonwillison.net/2025/Dec/9/mistral-vibe/

2025-12-08 05:07:48 GMT

- reply

OK here's a real feature I just built, reviewed and deployed on my phone while riding on a cab
Now live on https://www.niche-museums.com/116 - code is here: https://github.com/simonw/museums/pull/48
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/682/211/316/693/595/original/d3f7f239d3fa8c9a.png

2025-12-08 04:09:27 GMT

- reply

all three of these are research projects - they are not code I intend to ship, but they give me a very solid starting point for when I get back to a laptop

2025-12-08 04:08:10 GMT

- reply

two more coding research projects I ran from my phone over in-flight WiFi just now: https://github.com/simonw/research/pull/38 and https://github.com/simonw/research/pull/37

2025-12-08 03:52:02 GMT

- reply

turns out that one's on my list of museums I want to visit already!
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/681/913/451/806/173/original/890e1e7f944598f2.jpeg

2025-12-08 03:32:36 GMT

- reply

the prompt I used for that is the quoted bit in the PR https://github.com/simonw/research/pull/36

2025-12-08 03:31:47 GMT

- reply

I tried it out but haven't written up my experiment yet, it didn't work as well as I hoped

2025-12-08 03:31:10 GMT

- reply

I used my phone to write and test a C extension for SQLite this morning, just out of idle curiosity https://github.com/simonw/research/tree/main/sqlite-ripgrep-function

2025-12-08 03:29:34 GMT

- reply

I really liked his take on why AI art feels so eerie and incomplete

2025-12-08 03:27:54 GMT

- reply

I think the bestiary, but I also loved the oil paintings of Soviet space dogs and the general feeling of overwhelming confusion, also the trailer park dioramas

2025-12-08 03:26:58 GMT

- reply

I couldn't see any evidence of the fire, I only found out about that after my visit when I read the Wikipedia page

2025-12-08 02:51:44 GMT

I finally made it to The Museum of Jurassic Technology today! It's been top of my list of want-to-go museums for years https://www.niche-museums.com/116

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/

2025-12-02 00:22:13 GMT

Notes and pelicans for today's pair of DeepSeek v3.2 model releases https://simonwillison.net/2025/Dec/1/deepseek-v32/

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

2025-11-30 22:21:52 GMT

It's ChatGPT's third birthday today, so I pulled other some quotes from the early
months of ChatGPT on my blog
OpenAI really did not expect it to be as popular as it was!
https://simonwillison.net/2025/Nov/30/chatgpt-third-birthday/

2025-11-30 16:11:17 GMT

I think one of the underrated things about LLMs for coding is that they make writing code on a mobile phone massively more feasible
I can build small but useful pieces of software entirely on my phone now - without even needing to pay full attention, so it's compatible with being out in the world
I knock out small tools on my phone several times a week now, but my most ambitious phone coding project is still this one https://simonwillison.net/2025/Jul/17/vibe-scraping/

2025-11-29 02:27:35 GMT

- reply

The Netflix number is five years old and the ChatGPT number pre-dates GPT-5 by a couple of months so there are plenty of caveats here, but I expect the order of magnitude remains about correct

2025-11-29 02:16:50 GMT

Out of curiosity I decided to try and run the numbers on how much Netflix you can watch for the energy cost of a ChatGPT prompt
As far as I can tell it's between 5.1 and 10.2 seconds, depending on which end of the 2019 IEA Netflix energy usage estimate you use
https://simonwillison.net/2025/Nov/29/chatgpt-netflix/
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/630/578/559/731/460/original/db9e6db69f43a853.png

2025-11-27 16:03:19 GMT

DeepSeek-Math-V2 means we now have an open weights (Apache 2) model that can achieve gold medal performance on this year's International Mathematical Olympiad - previously proprietary models from OpenAI and Google DeepMind had achieved that
689GB from Hugging Face! https://simonwillison.net/2025/Nov/27/deepseek-math-v2/

2025-11-26 00:33:38 GMT

I'm in an episode of the new Data Renegades podcast with CL Kao and Dori Wilson talking about data journalism and Datasette and related topics - it's a fun conversation, here are some extracted highlights https://simonwillison.net/2025/Nov/26/data-renegades-podcast/

2025-11-25 21:00:05 GMT

- reply

The best approach I know of for reducing the risk here is to make sure that any credentials that are visible to coding agents - like AWS keys - are tied to non-production accounts with strict spending limits
That way if the credentials are stolen the blast radius is limited

2025-11-25 20:59:42 GMT

PromptArmor have a nasty description of a prompt injection attack against Google's new Antigravity AI IDE which can result in credentials (like AWS keys) being stolen by an attacker: https://www.promptarmor.com/resources/google-antigravity-exfiltrates-data
My notes here: https://simonwillison.net/2025/Nov/25/google-antigravity-exfiltrates-data/

2025-11-24 20:59:45 GMT

- reply

Updated my post with this section about their improved protection against prompt injection attacks - definitely better, but the problem is that if an attacker gets 10 tries they'll still succeed 1/3rd of the time!
https://simonwillison.net/2025/Nov/24/claude-opus/#still-susceptible-to-prompt-injection
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/606/682/338/707/989/original/41adf58409836c2f.png

2025-11-24 19:48:03 GMT

- reply

I had access over the weekend but with hindsight didn't do a great job of evaluation because I used it for real work: Most of the sqlite-utils 4 alpha release was written with help from Opus
I learned that production coding like that doesn't help identify new model strengths!

2025-11-24 19:40:31 GMT

- reply

I'm glad to see it's 1/3rd of the cost of the previous Claude Opus - still more expensive than GPT-5.1-Codex-Max and Gemini 3 Pro but much more competitive with them than if they had stuck to their previous pricing!

2025-11-24 19:38:37 GMT

Initial impressions (and pelicans) of Claude Opus 4.5, Anthropic's new "best model in the world for coding" released this morning. https://simonwillison.net/2025/Nov/24/claude-opus/

2025-11-24 15:04:38 GMT

- reply

I prototyped one of the new features using a Claude code for web running on my phone while out walking the dog! Here's the prompt I used for that https://github.com/simonw/research/pull/31

2025-11-24 14:54:29 GMT

Annotated release notes for the new sqlite-utils alpha, which includes several minor backwards incompatible changes that will be released in the stable 4.0 version once the new alpha has baked for a while https://simonwillison.net/2025/Nov/24/sqlite-utils-40a1/

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/

2025-11-20 19:07:34 GMT

- reply

no, and it looks like the watermark spans the whole image if this Reddit post is anything to go by https://www.reddit.com/r/nanobanana/comments/1o1tvbm/nano_bananas_watermark/

2025-11-20 17:11:44 GMT

- reply

I'm a freeloader - Ghost would cost me $200/month already

2025-11-20 17:10:29 GMT

- reply

the goal with the second prompt is to see if they can follow instructions like that - I'll keep running the old prompt too, which is more vague and hence gives them a bit more creative freedom

2025-11-20 17:09:46 GMT

- reply

that seems to me more like a feature of the generation harness as opposed to the model itself - Nano Banana Pro in AI Studio lets you pick an aspect ratio (1:1, 9:16, 16:9, 3:4, 4:3, 3:2, 2:3, 5:4, 4:5, 21:9) and 1K, 2K or 4K for the output size

2025-11-20 16:37:34 GMT

- reply

Also notable is the new SynthID feature of the Gemini app - you can upload a photo to it and ask if it was generated by AI and it will detect the invisible SynthID watermarks added by Nano Banana Pro
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/583/002/223/469/275/original/a71e1e1a5f72767f.png

2025-11-20 16:34:28 GMT

Nano Banana Pro, released this morning, is clearly the best image generation model. Superb instruction following, plus it can generate full infographics (with correct spelling and properly rendered text!) from a short prompt based on running extra searches https://simonwillison.net/2025/Nov/20/nano-banana-pro/

2025-11-19 23:16:12 GMT

Notes on GPT-5.1-Codex-Max - remember when GPT-5 was meant to signal and end to confusing model names? https://simonwillison.net/2025/Nov/19/gpt-51-codex-max/

2025-11-19 22:05:28 GMT

- reply

Here's a post on my blog with additional notes and links to the source code of the various components that make it all work https://simonwillison.net/2025/Nov/19/how-i-automate-my-substack-newsletter/

2025-11-19 22:04:30 GMT

Made a video demonstrating the slightly absurd collection of digital duct tape I use to automate the sending of my Substack newsletter, despite Substack not offering an API https://www.youtube.com/watch?v=BoPZltKDM-s

2025-11-18 23:01:55 GMT

New llm-gemini plugin release for my LLM CLI tool, in addition to Gemini 3 support it also adds the ability to pass in a YouTube URL as an attachment in order to analyze that video using Gemini
https://simonwillison.net/2025/Nov/18/llm-gemini/

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/

2025-11-18 20:53:44 GMT

My notes on Google's new Antigravity IDE - at first glance it looks like yet another VS Code fork but there are some interesting new ideas in there https://simonwillison.net/2025/Nov/18/google-antigravity/

2025-11-18 19:12:59 GMT

- reply

For reference, here's a photo I took myself of a California brown pelican in breeding plumage a couple of weeks ago
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/572/288/783/052/659/original/2ea8283294599095.png

2025-11-18 19:07:37 GMT

- reply

Here's the upgraded SVG pelican riding a bicycle prompt - I've added stricter requirements around the bicycle and specified that the pelican should be a California brown pelican in full breeding plumage
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/572/267/607/338/797/original/ff4864aa4fc97775.png
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/572/267/618/135/098/original/e56a97864a27802d.png
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/572/267/593/363/931/original/937d4941962d2293.png
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/572/267/592/852/323/original/f9b762ac5d70cf55.png

2025-11-18 19:02:13 GMT

My notes on Gemini 3, including analyzing a 3.5 hour council meeting audio recording and performance on a new, improved version of my pelican on a bicycle benchmark https://simonwillison.net/2025/Nov/18/gemini-3/

2025-11-15 20:52:34 GMT

New release of my llm-anthropic plugin adding support for structured outputs (via the new official API - previously I faked it with tool calls) and Anthropic's web search feature https://github.com/simonw/llm-anthropic/releases/tag/0.22
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/555/693/351/053/224/original/7b5908a461794d60.png

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/

2025-11-11 23:01:59 GMT

I've now reached the "six coding agents in six terminal windows at once" phase of parallel agent delirium
https://simonwillison.net/2025/Nov/11/six-coding-agents-at-once/
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/533/552/659/003/724/original/36eabcdf2a41f3bc.mp4

2025-11-11 19:05:00 GMT

- reply

personally I count them as "open source" if they use an OSI-approved license for the model weights themselves
There is a credible argument that they're not open source if the training data isn't available - but since the training data for almost every model is full of copyright stuff it's not actually possible for them to publish that data in most cases
I wrote a bit about that here: https://simonwillison.net/2024/Aug/27/open-source-ai/

2025-11-11 16:39:33 GMT

- reply

If you've been following my blog you'll have seen all of this already - I send out the newsletter by copy-pasting in my blog content since the last newsletter using a custom tool I built in an Observable notebook https://simonwillison.net/2023/Apr/4/substack-observable/

2025-11-11 16:38:02 GMT

Sent out my weekly-ish newsletter - I left it a bit too long this time and it ended up with three full articles, two YouTube videos, 5 SVGs of pelicans riding bicycles and 3 POV-Ray renders of pelicans riding bicycles https://simonw.substack.com/p/code-research-projects-with-async

2025-11-09 17:16:09 GMT

- reply

I ran their "Write a POV-Ray file that shows a pelican riding on a bicycle" prompt against the new gpt-5-codex-mini and got the worst result yet! https://simonwillison.net/2025/Nov/9/pelican-on-a-bike-raytracer-edition/

2025-11-09 17:15:28 GMT

This is really fun - [@beetle_b](https://mastodon.xyz/@beetle_b ) ran a variant of my pelican riding a bicycle SVG test using POV-Ray instead and got terrible ray-traced pelicans out of a bunch of different models: https://blog.nawaz.org/posts/2025/Oct/pelican-on-a-bike-raytracer-edition/ and https://mastodon.xyz/@beetle_b/115436481779600909

2025-11-09 03:48:00 GMT

- reply

For comparison, here are the pelicans riding bicycles drawn by GPT-5-Codex-Mini (the new model), GPT-5-Codex and full GPT-5 - all produced via the same hacked version of the Codex CLI tool
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/517/690/756/285/663/original/b8ee40cbaf135193.png
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/517/690/756/285/664/original/c3baa68a9862bab3.png
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/517/690/789/659/966/original/d106183a8cc713e2.png

2025-11-09 03:39:35 GMT

- reply

I also recorded a 7 minute YouTube video showing how I got Codex to reverse-engineer and then extend itself in order to draw me that pelican https://www.youtube.com/watch?v=9o1_DL9uNlM

2025-11-09 03:38:33 GMT

- reply

Since there's no API access yet I got OpenAI's Codex coding agent to rewrite itself (in Rust) to add a new "codex prompt ..." command which I could use to run prompts against the private models that are only available within that tool - full details here https://simonwillison.net/2025/Nov/9/gpt-5-codex-mini/

2025-11-09 03:37:16 GMT

OpenAI partially released a new model yesterday called GPT-5-Codex-Mini
No API access yet, but I did some truly horrible things to their Codex CLI app to get it to spit out this SVG of a pelican riding a bicycle
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/517/648/555/892/162/original/96746136afbe1b53.png

2025-11-08 01:54:16 GMT

- reply

... this got fixed in Mastodon 4.5! https://blog.joinmastodon.org/2025/11/mastodon-4.5/
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/511/581/314/047/678/original/cb95e707d8bb811a.png

2025-11-07 16:13:54 GMT

I have a hunch that current LLMs might make it easier to launch a brand new programming language, provided you can describe it in a few thousand tokens and ship it with a compiler and linter that coding agents can use https://simonwillison.net/2025/Nov/7/llms-for-new-programming-languages/#atom-everything

2025-11-07 07:23:54 GMT

New TIL: Using Codex CLI with gpt-oss:120b on an NVIDIA DGX Spark via Tailscale https://til.simonwillison.net/llms/codex-spark-gpt-oss

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/

2025-11-06 18:44:02 GMT

- reply

uv makes testing different projects against upgraded dependencies so much easier - no need to think about virtual environments, uv handles them almost invisibly
I wrote more about my uv testing tricks in this TIL https://til.simonwillison.net/python/uv-tests

2025-11-06 18:32:08 GMT

RE: https://mastodon.social/@Sarahp/115504173496002298
Glad to see this become a universal feature as opposed to one that was limited to specific clients
#note1zaa…em8k

2025-11-06 18:31:05 GMT

Made a new video demonstrating my process for upgrading a Datasette plugin using uv and an OpenAI Codex bash one-liner https://www.youtube.com/watch?v=qy4ci7AoF9Y
Here are detailed notes to accompany the video on my blog: https://simonwillison.net/2025/Nov/6/upgrading-datasette-plugins/

2025-11-06 16:02:53 GMT

- reply

And here's an example of one of my code research prompts
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/503/593/602/501/577/original/8788f703f5cc2a39.png

2025-11-06 15:59:17 GMT

- reply

Here's my research repo - each of the 13 folders is a different research project, and the README is automatically updated by an LLM to include summaries describing each one https://github.com/simonw/research?tab=readme-ov-file#research-projects-carried-out-by-ai-tools
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/503/579/386/023/767/original/3919cfe1f161ec5f.png

2025-11-06 15:56:18 GMT

I've been getting a lot of value using coding agents for code research tasks recently - I have a dedicated simonw/research GitHub repo and I frequently have them run detailed experiments and write up the results. Here's how I'm doing that + some examples:
https://simonwillison.net/2025/Nov/6/async-code-research/

2025-11-05 15:23:07 GMT

- reply

It's a great TikTok! https://www.tiktok.com/@belligerentbarbies/video/7568380008633257271