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

nprofile1qqsyx9dpslszfqvyjtnpjwqf8wng8khxvcjdyqkdgdecmedchgvccrcwpc7rn

Author Public Key

npub1gv26rplqyjqcfyhxryuqjwaxs0dwve3y6gpv6smn3hjm3wse3s8squtlwl

Show more details

Last Notes

2024-08-31 14:40:15

- reply

I think I know enough about programming to assess the tests: I use tricks like changing the implementation, confirming the test breaks, then fixing the implementation and confirming the test passes

2024-08-31 14:07:47

- reply

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

2024-08-24 23:07:27

My notes on Google Research's new paper describing "pipe syntax", their alternative syntax for SQL queries which they've been rolling out internally since February https://simonwillison.net/2024/Aug/24/pipe-syntax-in-sql/
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/113/019/508/753/588/477/original/449a4d33d3bef877.png

2024-07-09 18:01:55

- reply

yeah, I imagine the fingerprinting risk is why they don't expose this functionality to everyone else

2024-07-09 17:54:05

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/

2024-07-08 15:24:18

In case anyone catches wind of the story going around that Reform fielded at least one fake candidate in the UK election, Private Eye tracked down that candidate and confirmed that, while their photo was AI-generated (supposedly due to a wrong coloured tie) they are an actual person that exists: https://twitter.com/PrivateEyeNews/status/1810327043827249452
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/112/751/559/024/686/035/original/5ba2b31f837d09fb.png

2024-07-02 22:17:45

What do you think of Continuous Deployment, where you push code to main and, if the tests pass, that code is automatically deployed to production?

2024-07-01 17:59:46

The new Ladybird initiative is exciting: a new US non-profit initially backed by $1m to support development of Ladybird, an open source web browser that isn't based on a fork of something older and that doesn't depend on Google for funding! https://simonwillison.net/2024/Jul/1/the-ladybird-browser-initiative/

2024-06-25 22:31:30

Worth grepping your source code for "polyfill.io" and taking urgent measures to remove that code if you're linking it into your site - the domain name apparently now intermittently serves malicious JavaScript
My notes here: https://simonwillison.net/2024/Jun/25/polyfill-supply-chain-attack/ - or read this article https://sansec.io/research/polyfill-supply-chain-attack

2024-06-24 04:45:41

It took me an embarrassingly long time to realize that type hints (in Python and in TypeScript) aren't so much about avoiding accidental type errors at runtime... they're much more about enabling really good editor autocomplete

2024-06-23 16:22:32

Here's a brilliant hack... llama.ttf https://fuglede.github.io/llama.ttf/ is a font with a LLM embedded in it (using WebAssembly, which apparently can be used in fonts now)
You type "Once upon a time!!!!!!!!!!" and those exclamation marks are rendered to show the LLM generated text, using a tiny 30MB model
I wrote more notes on how it works (including a link to another font demo that can be used to play Tetris) here: https://simonwillison.net/2024/Jun/23/llama-ttf/

2024-06-22 02:33:54

- reply

(I'm increasingly suspicious that hardly anyone ever clicks on inline links in text like this, but that's not going to stop me from including them!)

2024-06-22 02:32:32

Anyone else spend way too much time thinking about the grammar of linking to things?
In this fragment I'm not at all confident about the "backs up that data" link, maybe I should have included "as JSON" in that link...
Anyone seen a comprehensive style guide that addresses this?
Extract from https://simonwillison.net/2024/Jun/21/search-based-rag/#ingredients-for-rag
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/112/657/927/145/698/410/original/266a67593b28a274.jpeg

2024-06-20 04:11:40

Realised I hadn't checked in on what @npub1k3l…3e8u and Soundslice had been up to recently... and they've only gone and solved OCR scanning of sheet music to import it into their web app player/editor interface!
It's a devastatingly beautiful implementation, packed with Soundslice's signature attention to detail https://www.soundslice.com/sheet-music-scanner/
It's worth reading their release notes to see how the feature has evolved since they launched it https://www.soundslice.com/blog/music-scanning/

2024-06-20 03:59:08

- reply

thanks, hadn't tried either of those - added them to my TIL as alternatives

2024-06-20 02:35:35

New TIL: Running Prettier against Django or Jinja templates https://til.simonwillison.net/npm/prettier-django

2024-06-16 00:57:15

The 6th example I've seen of the same prompt injection attack against LLM chatbots: https://embracethered.com/blog/posts/2024/github-copilot-chat-prompt-injection-data-exfiltration/
The attack involves tricking an LLM chatbot with access to both private and untrusted data to embed a Markdown image with a URL to an attacker's server where that URL leaks private data extracted from the session.
We've now seen this same attack in ChatGPT itself, Google Bard, Writer.com, Amazon Q and Google NotebookLM (all now fixed, thankfully).
My collection: https://simonwillison.net/tags/markdownexfiltration/

2024-06-13 03:50:09

The https://phanpy.social/ desktop timeline browsing experience is genuinely the first time I've actively enjoyed scrolling through my Mastodon timeline, it feels really fresh and the way it groups boosts together is a huge improvement on the default chronological lists from other tools

2024-06-10 22:56:59

- reply

Here are some of the technical details of the Private Cloud Compute platform and WOW that is some cutting edge computer science https://security.apple.com/blog/private-cloud-compute/

2024-06-10 20:23:49

Wrote up some initial thoughts on the Apple Intelligence WWDC announcements today:
- On-device models that can outsource to Apple’s servers with cryptographic guarantees of privacy are really interesting
- Restricting image generation to three approved styles (sketch, illustration and animation) avoids a whole host of complicated ethical concerns
- App Intents look like a minefield for prompt injection
https://simonwillison.net/2024/Jun/10/apple-intelligence/

2024-06-09 00:06:06

Woohoo! @npub176w…vm5y has a new Link Blog https://laughingmeme.org/links/
Way more people should have one of these, they're great
> Like many people I’ve been dealing with the collapses of the various systems I relied on for information over the previous decades. [...] And particular with the collapse of the social spaces many of us grew up with, I feel called back to earlier forms of the Internet, like blogs, and in particular, starting a link blog.
https://fiasco.social/@kellan/112583726435885054

2024-06-06 14:11:03

I wrote about a delightful example of accidental prompt injection against RAG applications, where a RAG application against my LLM project's documentation suddenly decided that it was a gerbil
https://simonwillison.net/2024/Jun/6/accidental-prompt-injection/

2024-06-04 20:13:30

I got a good quote in this story about that dumb Zoom Idea to have "digital twins" attend meetings in your place https://arstechnica.com/information-technology/2024/06/zoom-ceo-envisions-ai-deepfakes-attending-meetings-in-your-place/
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/112/560/177/388/682/676/original/8ca1212303e69487.jpeg

2024-05-29 11:03:57

I wrote about a common misconception I see people have about LLM tools like ChatGPT
Training is not the same as chatting: ChatGPT and other LLMs don’t remember everything you say
https://simonwillison.net/2024/May/29/training-not-chatting/

2024-05-21 17:01:53

- reply

The value of implementing "Sign in with Facebook" dropped like a stone a few years ago when they made social graph features much less available in response to the whole Cambridge Analytica thing

2024-05-21 16:59:40

It feels like "Sign in with Facebook" is a whole lot less common than it used to be, not sure I remember the last time I saw a new service that had that, whereas "Sign in with Google" and "Sign in with Apple" still show up a whole lot

2024-05-18 15:46:05

Here's the code for the AI counter app I used in my #PyConUS keynote today - it uses vosk and pyaudio to listen through the microphone and increment a visible Tkinter counter any time it detects "AI" or "Artificial Intelligence" https://github.com/simonw/count-ai
Prompts I used to write it are here: https://chatgpt.com/share/58f2352d-1f17-495b-94f1-4eb44cd574b9
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/112/462/865/912/319/052/original/b9b59a13b78f7253.png

2024-05-17 13:43:47

The #pyconus tag on here is super active right now, great to have a functioning social media back channel again!

2024-05-13 19:44:46

- reply

Notes here! https://simonwillison.net/2024/May/13/gpt-4o/

2024-05-12 14:07:09

Several of the major social media platforms - Instagram, TikTok, LinkedIn, Twitter - have effectively declared war on linking to things and I absolutely hate it
"Link in my bio" / "Link in thread" / "Link in first comment"... or increasingly no link at all, just an unsourced screenshot of a page

2024-05-08 00:16:17

Here's a brilliant neologism: "slop", for text generated entirely by LLMs and published, unwanted, on the Internet
> Watching in real time as "slop" becomes a term of art. the way that "spam" became the term for unwanted emails, "slop" is going in the dictionary as the term for unwanted AI generated content
Source: https://twitter.com/deepfates/status/1787472784106639418
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/112/402/582/374/153/670/original/db9cf6c465c1574d.png

2024-05-01 02:56:48

Do you understand what is happening in this TikTok? https://www.tiktok.com/@donghuajinlong/video/7363487878598495531

2024-04-20 02:37:04

Oakland food tip: we tried the pecan caramel cheesecake from Gregory's Gourmet Desserts yesterday and it was spectacular, like worth-a-special-trip spectacular https://maps.app.goo.gl/iYNXU36ZybBtUSFu5
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/112/301/214/639/168/048/original/1ac134c143264605.jpeg
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/112/301/216/959/885/926/original/adf626f0c8452ba4.jpeg
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/112/301/219/535/359/738/original/8c94ba07ff573b99.jpeg

2024-04-09 14:28:09

- reply

one of the many reasons I'm so interested in smaller models that can run directly on a laptop or phone - like Mistral 7B - is that the inference energy costs are then reduced to the cost of running one of those devices - and the one-time training cost gets shared between everyone using that model

2024-04-09 14:26:21

- reply

a result of that is that innovations which help train better models with less power usage are celebrated, and can be used to drive down energy costs
In crypto mining any innovation that lets you mine more efficiently just means you can make even more money by burning even more energy with it

2024-04-09 14:24:11

- reply

there is an important difference with how crypto mining works: crypto mining is deliberately designed as a competition
To win the coin mining lottery you need to burn more energy than anyone else does - it's a revolting scheme which directly incentivizes energy wastage
Training models is different in that you still get a model you can use even if someone else burns more energy to train a bigger one

2024-03-30 18:04:07

I built a new tool: https://tools.simonwillison.net/ocr - it runs OCR against images and PDFs entirely in your browser (no file upload needed) using Tesseract.js and PDF.js
I wrote more about the tool and how I built it (with copious amounts of Claude 3 Opus and a little bit of ChatGPT) here: https://simonwillison.net/2024/Mar/30/ocr-pdfs-images/
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/112/185/950/303/038/617/original/496d8a233536b3e4.mp4

2024-03-22 19:52:44

- reply

I wrote this up in part because I'm tired of hearing people complain that LLMs aren't useful. There are many valid criticisms of them as a technology, but "not being useful" should not be one of them https://simonwillison.net/2024/Mar/22/claude-and-chatgpt-case-study/#llms-are-useful
https://cdn.masto.host/fedisimonwillisonnet/media_attachments/files/112/141/083/922/567/463/original/be7c0155293c09e1.png

2024-03-22 19:50:09

I put together some detailed notes showing how I use Claude and ChatGPT as part of my daily workflow - in this case describing how I used them for a 6 minute side quest to create myself a GeoJSON map of the boundary of the Adirondack Park in upstate New York
https://simonwillison.net/2024/Mar/22/claude-and-chatgpt-case-study/