I couldn't resist recreating this using the new GPT-5 Codex Mini, which produced the worst ray-traced pelican on a bicycle yet! simonwillison.net/2025/Nov/9/p...

Doing this easily was very much not the point!

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

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 www.youtube.com/watch?v=9o1_...

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 simonwillison.net/2025/Nov/9/g...

Provided the publishing industry can resist the temptation to bankrupt the Internet Archive we might be OK

I love this project infinitemac.org

I think old language execution is more or less a solved problem now thanks to emulators - we can run code from 50 years ago on an emulator, or even an emulator inside an emulator

A lot of those emulators run in the browser now thanks to WebAssembly!

Good error messages are so important these days! The past couple of Python versions included some big improvements to their error messages for which I'm very grateful

Yes I'd love to se an example of this too, challenge is finding an interesting novel language that definitely hasn't made it into any LLM training data yet!

Yes I'm here for good I think

I have some undisclosed ones that I use - I'm hoping some day I'll catch a model drawing an amazing pelican on a bicycle and failing horribly at some other creature on some other mode of transport!

Which means it's not OSI compliant, which is my bar for calling something "open source" over just "open weights"

Then I wanted to make it easier to play with, so another hour with Claude Code and I had a plugin for @simonwillison.net's llm: github.com/btucker/llm-...

What's cool this is you don't have to install anything other than some python packages & you have full access to a reasonably capable LLM.

Ooh I haven't tried that yet, thanks for the tip!

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 til.simonwillison.net/python/uv-te...

No, because it's not creating the data - it's writing code that then produces that data.

It might make mistakes in the code it writes but those are a lot easier to spot.

I doubt it, personally. I spend a lot of time experimenting with local LLMs and I've not yet experienced any that I'd trust with a task remotely as complex as "search for PyPI markdown packages, download them and design and implement a comparative benchmark that tries them out"

I find Claude Sonnet 4.5 fits my own personal coding style slightly better, but that might just be that I have more practice in prompting it - they're very close to each other!

I like Claude Code for web more than Codex Cloud because you can send Claude new instructions while it's still running