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

simon_at_fedi.simonwillison.net@momostr.pink

Profile Code

nprofile1qqsyx9dpslszfqvyjtnpjwqf8wng8khxvcjdyqkdgdecmedchgvccrcpz3mhxue69uhhyetvv9ujuerpd46hxtnfduqs6amnwvaz7tmwdaejumr0ds0kg6v4

Publishing to

atlas.nostr.land

relay.damus.io

nos.lol

Author Public Key

npub1gv26rplqyjqcfyhxryuqjwaxs0dwve3y6gpv6smn3hjm3wse3s8squtlwl

Show more details

Last Notes

2025-10-20 17:28:06 UTC

DeepSeek released a new OCR model - I got it working on an NVIDIA Spark (CUDA + ARM64) by starting a Docker container, running Claude Code as root with "claude --dangerously-skip-permissions" and telling it to figure it out
Took 4 prompts and 40 minutes https://simonwillison.net/2025/Oct/20/deepseek-ocr-claude-code/

2025-10-07 15:05:23 UTC

Vibe coding is irresponsibly building software through dice rolls, not caring what code is produced
What about when engineers at the top of their game use AI tools responsibly to accelerate their work?
I propose "vibe engineering"! (with my tongue only partially in my cheek)
https://simonwillison.net/2025/Oct/7/vibe-engineering/

2025-09-21 15:58:38 UTC

New on Niche Musems: my write-up
of a visit to the Musical Museum in Brentford, London... player pianos, self-playing violins, and orchestrions! https://www.niche-museums.com/115
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/243/103/513/134/111/original/42a3d55d5eafe1d3.jpeg
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/243/104/882/765/342/original/3bdd29da45df90e1.jpeg
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/243/105/787/140/366/original/de031fa5082ccaf0.jpeg
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/243/110/448/055/551/original/2d751f5170ab2845.jpeg

2025-09-21 00:25:31 UTC

New release of my llm-openrouter plugin adds support for reasoning options and tool usage, which means you can now execute tools against 179 models made available via OpenRouter
https://simonwillison.net/2025/Sep/21/llm-openrouter/

2025-09-20 15:57:46 UTC

Leaked memo from Amazon that warns existing H1B holders to avoid travel back into the USA after the September 21st deadline - their lawyers evaluated the new executive order as not just affecting new applications https://www.businessinsider.com/read-memos-sent-big-tech-trump-h-1b-changes-2025-9
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/237/444/190/470/099/original/2ffb3169640e7dec.jpeg
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/237/444/723/759/665/original/47fd3d89485d8cba.png

2025-09-12 07:40:50 UTC

Posted some of my own notes on Shlok Khemani's (excellent and comprehensive) notes on how Claude and ChatGPT's memory implementations differ from each other https://simonwillison.net/2025/Sep/12/claude-memory/

2025-09-12 04:14:37 UTC

Made two pelicans with the new Qwen3-Next-80B-A3B-Instruct and Qwen3-Next-80B-A3B-Thinking models
🐧🦩 Who needs legs?!
https://simonwillison.net/2025/Sep/12/qwen3-next/

2025-09-10 17:28:35 UTC

Some notes on the new Claude API web fetch tool, which I think can be used safely despite the risk of prompt injection exfiltration attacks if you're really carful with the allowed_domains parameter https://simonwillison.net/2025/Sep/10/claude-web-fetch-tool/

2025-09-09 18:41:12 UTC

Claude just released their new version of ChatGPT Code Interpreter! It can run Python and Node.js code in a sandboxed Linux environment, and unlike OpenAI's version the Claude one can install additional packages
Here's my detailed review: https://simonwillison.net/2025/Sep/9/claude-code-interpreter/

2025-09-08 21:00:53 UTC

I got Codex CLI and GPT-5 to help me modify the Transformers.js Llama 3.2 chat demo to enable loading that 1.2GB model from a local folder instead of fetching it from a URL
Full details including the Codex transcript and prompts I used here: https://simonwillison.net/2025/Sep/8/webgpu-local-folder/
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/170/689/051/311/771/original/589a1699cd1461d4.mp4

2025-09-06 22:48:31 UTC

- reply

o3 and GPT-5 Thinking (not necessarily regular non-thinking GPT-5) really do represent a step change in how effective models are at evaluating search results, so I'd like to see clarity from them on exactly which models they evaluated and when

2025-09-06 19:45:05 UTC

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 17:03:59 UTC

My notes on Kimi-K2-Instruct-0905, aka Kimi K-2.1 - an incremental improvement on Moonshot's previous trillion parameter open weights model, now with twice the context length (256k up from 128k) https://simonwillison.net/2025/Sep/6/kimi-k2-instruct-0905/

2025-08-25 09:47:07 UTC

- reply

There is some good news here: the Hacker News thread about this topic is refreshingly packed with people who clearly understand the nature and severity of the problem and get why none of the proposed mitigations are likely to work out https://news.ycombinator.com/item?id=45004846

2025-08-22 22:42:53 UTC

Notes and a pelican for DeepSeek 3.1 https://simonwillison.net/2025/Aug/22/deepseek-31/

2025-08-19 23:47:36 UTC

Notes on running the new Qwen-Image-Edit model on my Mac, using the prompt "Give the pelicans rainbow colored plumage" against a photo I took of some pelicans
It took almost ALL of my 64GB of RAM and needed 54GB of disk space https://simonwillison.net/2025/Aug/19/qwen-image-edit/

2025-08-14 17:27:21 UTC

The new Gemma 3 270M open weights model from Google is really fun - it's absolutely tiny, just a 241MB download
I asked it for an SVG of a pelican riding a bicycle and it wrote me a delightful little poem instead
The real value should become apparent with fine-tuning - models this small will never be useful for general purpose tasks but could absolutely fly when fine-tuned for something specific
https://simonwillison.net/2025/Aug/14/gemma-3-270m/
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/115/028/291/578/611/632/original/6818f2fd80b0765e.png

2025-08-13 05:42:03 UTC

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 00:16:26 UTC

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

2025-08-08 17:53:26 UTC

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/

2025-05-06 18:27:30 UTC

New Gemini 2.5 Pro preview model today, released in advance of Google I/O. Google claim it is state of the art in both frontend code generation and video understanding. https://simonwillison.net/2025/May/6/gemini-25-pro-preview/

2025-05-05 02:35:43 UTC

- reply

if you're interested in AI ethics here are the notes I've collected so far https://simonwillison.net/tags/ai-ethics/

2025-05-05 02:26:43 UTC

- reply

no need to be rude

2025-05-04 23:23:25 UTC

- reply

that website is EXACTLY why this issue matters to me - they start the page with the claims that 10kg of coal is needed to produce 200 ChatGPT emails
Is that true? It had better be true for their campaign to make sense
I don't think it is
(For one thing, how is 10,000 ChatGPT "searches" using the same energy as composing 200 emails? What's a "search" in this case anyway?)
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/114/452/135/867/205/711/original/3ea50a1ab07b202f.jpeg

2025-05-03 01:01:17 UTC

- reply

mainly still just experiments - here to justify using it frequently when the API Models are still cheap and a whole lot better

2025-05-02 23:45:12 UTC

My favorite local model right now is a bit of surprise to me: I'm really enjoying the relatively tiny Qwen3-8B, running the 4bit quantized version on my Mac using MLX
It's surprisingly capable given it's a 4.3GB download and uses just 4-5GB of RAM while it's running
https://simonwillison.net/2025/May/2/qwen3-8b/

2025-04-30 00:17:27 UTC

- reply

I base my estimates of their efficiency on their API prices, which have almost all continued to drop - with the exception of o1 Pro and GPT-4.5 which are both diabolically expensive https://tools.simonwillison.net/llm-prices
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/114/424/036/888/924/564/original/3db4a1e8dc0fa292.png

2025-04-30 00:12:38 UTC

- reply

where did you get the thing about OpenAI losing money on every user?
I know Sam made a throwaway remark about that for the $200/month OpenAI Pro thing a few months ago but I pretty much wrote that off as hyperbole and a way of bragging about their "most powerful" (aka least efficient) models - which they've already replaced with the more efficient o3/o4-mini combo

2025-04-30 00:00:30 UTC

- reply

I don't have any certainty at all over usage numbers, but I'm confident ChatGPT has more consumer users because anecdotally
most people haven't even heard of the competition (see also Google Trends)
I use DeepSeek as an example to illustrate that spending $100m on a training run may not actually be necessary after all, and optimistically they might be a lesson the other big AI labs learn

2025-04-29 23:54:31 UTC

- reply

I'm genuinely not comfortable giving people advice on which LLM to use if they are environmentally conscientious because none of the companies release enough energy usage data for me to make a credible suggestion

2025-04-29 23:40:24 UTC

- reply

I don't think I ever said I was pro-gold rush
I have complicated opinions about all of this! Hence my contradicting titles in attached screenshots from https://simonwillison.net/2024/Dec/31/llms-in-2024/#the-environmental-impact-got-better
The one opinion I'll stick by is that telling someone "every time you run a ChatGPT prompt you're using engi in mentally irresponsible amounts of energy" is misleading
(People really do say that)
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/114/423/890/708/015/359/original/d52abe40198b5152.jpeg
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/114/423/891/121/906/324/original/08d442de28381f3d.jpeg

2025-04-29 23:28:56 UTC

- reply

given everything we've learned about the transformer architecture in the last five years my strong intuition is that if OpenAI hadn't figure it out someone else would have

2025-04-29 22:46:51 UTC

- reply

it's using loads of energy and relieves. But it's also serving (latest claim) 500 million weekly active users

2025-04-29 22:42:09 UTC

- reply

they're losing money because that's how this stage in a gold rush works - everyone in the hot space raises a ton of money and burns through it trying to grow bigger than their competitors so that when they figure out the business model they can do it better than anyone else
Silicon Valley explained it pretty well https://www.youtube.com/watch?v=BzAdXyPYKQo&pp=0gcJCdgAo7VqN5tD

2025-04-29 19:01:33 UTC

- reply

I feel like in that case it's pretty easy to argue that if just a few hundred people take home a pretty rock it will reduce the supply of pretty rocks by an amount that meaningfully impacts that particular environment
I don't see a few thousand or even a few million people using ChatGPT as comparable to that, assuming they also spend time watching Netflix / YouTube / scrolling social media media etc

2025-04-29 16:31:21 UTC

- reply

the "but LLMs aren't actually useful" argument is covered by that piece too (it's very well constructed) https://andymasley.substack.com/p/a-cheat-sheet-for-conversations-about?#%C2%A7this-is-all-a-gimmick-anyway-why-not-just-use-google-chatgpt-doesnt-give-better-information

2025-04-29 15:35:04 UTC

I'm glad somebody out there is brave enough to push back against the "personal ChatGPT usage is terrible for the environment" message https://andymasley.substack.com/p/a-cheat-sheet-for-conversations-about?utm_campaign=posts-open-in-app&triedRedirect=true
"If you want to prompt ChatGPT 40 times, you can just stop your shower 1 second early."
"If I choose not to take a flight to Europe, I save 3,500,000 ChatGPT searches. this is like stopping more than 7 people from searching ChatGPT for their entire lives."

2024-08-31 14:07:47 UTC

- reply

I certainly won’t deny that there are an incredible new array of footguns now available to anyone who wants them

2024-08-31 13:44:14 UTC

- reply

I thought that too, but the more work I get done with LLMs myself the less worried I am about that
I have a Go project I wrote from scratch in production now, despite not being remotely fluent in Go. It has comprehensive test coverage and even implements continuous integration and continuous deployment, which is why I’m confident it’s not a spectacularly bad idea
Would other people YOLO something like that to production without tests? Maybe, and that would definitely be a bad idea!

2024-08-31 13:03:27 UTC

- reply

This meshes with my more optimistic take on AI-assisted programming from last year: AI-enhanced development makes me more ambitious with my projects https://simonwillison.net/2023/Mar/27/ai-enhanced-development/

2024-08-31 13:01:41 UTC

Forrest Brazeal:
“I think that AI has killed, or is about to kill, pretty much every single modifier we want to put in front of the word “developer.”
“.NET developer”? Meaningless. Copilot, Cursor, etc can get anyone conversant enough with .NET to be productive in an afternoon … as long as you’ve done enough other programming that you know what to prompt.”
From https://newsletter.goodtechthings.com/p/the-death-of-the-modified-developer
https://indieweb.social/@fatrat/113056013841232498

2024-08-10 17:47:06 UTC

- reply

any chance it could handle tweet threads in the future?
I have some /epic/ threads I want to archive, 100+ tweets chained together, most of which are themselves quote-tweets
This one for example (only properly visible if you are signed in to Twitter) https://twitter.com/simonw/status/1077737871602110466

2024-07-09 17:54:05 UTC

It turns out Google Chrome ships a default, hidden extension that allows code on `*.google.com` access to private APIs, including your current CPU usage
You can test it out by pasting the following into your Chrome DevTools console on any Google page:
chrome.runtime.sendMessage(
"nkeimhogjdpnpccoofpliimaahmaaome",
{ method: "cpu.getInfo" },
(response) => {
console.log(JSON.stringify(response, null, 2));
},
);
More notes here: https://simonwillison.net/2024/Jul/9/hangout_servicesthunkjs/