Simon Willison
@simonw
Followers
94K
Following
60K
Media
3K
Statuses
56K
Creator @datasetteproj, co-creator Django. PSF board. Hangs out with @natbat. He/Him. Mastodon: https://t.co/t0MrmnJW0K Bsky: https://t.co/OnWIyhX4CH
Joined November 2006
New hobby: prototyping video games in 60 seconds using a combination of GPT-3 and DALL-E. Here's "Raccoon Heist"
106
968
7K
If someone gives you a CSV file with 100,000 rows in it, what tools do you use to start exploring and understanding that data?.
2K
860
7K
Here's a piece of information that will send a chill down the spine of anyone who's ever designed a database schema:. Our new house that we just moved into. has two zip codes!.
235
672
5K
Leaked Google document: “We Have No Moat, And Neither Does OpenAI”. The most interesting thing I've read recently about LLMs - a purportedly leaked document from a researcher at Google talking about the huge strategic impact open source models are having.
121
1K
5K
Here's the spiciest detail from the new o1 system card:
The updated OpenAI o1 system card builds on prior safety work, detailing robustness evals, red teaming insights, and safety improvements using Instruction Hierarchy. It maintains a "medium" risk rating based on testing with an expanded suite of evaluations, reflecting it is safe.
87
436
5K
TIL you can run SQL queries directly against CSV files as a one-liner using the default sqlite3 command line utility.
51
689
4K
To test this out for yourself, open a tab on a Google site and paste this into the Chrome DevTools console:. chrome.runtime.sendMessage('nkeimhogjdpnpccoofpliimaahmaaome', {method: 'cpu.getInfo'}, response => {console.log(JSON.stringify(response, null, 2));});
So, Google Chrome gives all *.google.com sites full access to system / tab CPU usage, GPU usage, and memory usage. It also gives access to detailed processor information, and provides a logging backchannel. This API is not exposed to other sites - only to *.google.com.
62
682
4K
"Hosting SQLite databases on Github Pages" is absolutely brilliant: it adds a virtual filesystem to SQLite-compiled-to-WebAssembly in order to fetch pages from the database using HTTP range requests
42
885
4K
If you're just starting to learn software engineering right now but you're considering dropping it because you think the field might be made obsolete by AI, I have an alternative approach to suggest for you:. Start learning now, and use AI tools to learn FASTER.
66
350
3K
Here's the table of contents for my lengthy new piece on how I use LLMs to help me write code
47
306
3K
There's a meme where you ask ChatGPT "From all of our interactions what is one thing that you can tell me about myself that I may not know about myself" - don't be fooled into thinking there's anything deep going on here, its effectively a horoscope generator.
96
122
3K
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.
62
136
3K
"Do stuff and then blog about it" remains one of the most underrated pieces of career advice.
An absolutely fantastic way to increase this is to start a blog. Almost all the cool fun stuff in my professional life for me has come from doing stuff then blogging about it.
24
312
3K
Here's the table of contents for my end-of-year review of things we learned out about LLMs in 2024 - we learned a LOT
45
415
3K
@d_feldman It's got me to the point where I can read Spanish language news articles and understand ~80% of them - spoken Spanish is much harder.
21
28
2K
Anyone seen any credible successs stories stories for fine-tuning LLMs to solve real-world business challenges?. I want stories where a company solved something using a fine-tuned model where previous attempts at solving it had failed. (I feel like I ask this about once a month).
177
134
2K
How about if, instead of ditching Twitter for Mastodon, we all start blogging and subscribing to each other's Atom feeds again instead? The original distributed social network could still work pretty well if we actually start using it.
89
540
2K
Found the system prompt that drives this thing here: It works by generating a base64 encoded PNG of the drawn components, then passing that to GPT-4 Vision with that system prompt and instructions to "Turn this into a single html file using tailwind".
23
241
2K
OK, I'm calling it: Large language models are having their Stable Diffusion moment right now.
43
413
2K
It's interesting how OpenAI are doubling down on end-user-facing features like voice mode, while Anthropic are doubling down on engineer/API-focused features like code generation quality and being able to remotely drive a GUI.
53
85
2K
A year ago nobody outside OpenAI had trained a model as good as GPT-4. Today there are dozens - and if you trust the benchmarks that includes some that you can run on a laptop (Qwen2.5-32B perhaps?). What changed? What techniques are used now that weren't known a year ago?.
84
92
2K
I tried this out yesterday and it's incredible: download a 4GB binary, child 755 it and now you have a full LLM and the software needed to run it ready to go, with multiplied operating system platforms supported by that single file.
I spent the last month building llamafile which is the fattest executable file format ever. It lets you turn LLM weights into runnable llama.cpp binaries using cosmo libc. Blog post:
14
179
2K
If you try nothing else today, give the demo at a go - it lets you stream video and audio directly to Gemini 2.0 Flash and get audio back, so you can have a real-time audio conversation about what you can see with the model. Feels like science fiction!
55
292
2K
Gemini 2.5 Pro and Flash now have the ability to return image segmentation masks on command, as base64 encoded PNGs embedded in JSON strings. I vibe coded this interactive tool for exploring this new capability - it costs a fraction of a cent per image
51
118
2K
I tried feeding a 7s video of my bookshelf into Gemini Pro 1.5 to get back a JSON array of books. and it worked!.
124
202
2K
Honestly, at this point If you give me a programming interview and don't let me use AI assistance you won't get a very realistic idea of what I'm actually capable of.
Coding interviews should be using ai. If you disagree you literally don’t understand what ai is for. Sure you can use it to fill gaps & find easy answers. More importantly you can use it to accelerate and do much bigger things. Failing to select for this skill will ruin tech co’s.
71
130
2K
I was having a conversation with Claude about unconventional things to do in the SF Bay Area and I got a bit suspicious so I prompted "Are you sure all of those are real? I think you made some of those up."
101
82
2K
OK this is really cool: Nico built an OpenAI action/GPT which breaks packages from PyPI up into <9.5MB chunks and returns them to ChatGPT in a way that lets it save them to disk. and then pip install them!. I didn't know actions could do that, docs here:
@ChatGPTapp If you want to try pip installing a package, here's the link
4
41
367
I built a new tool! It's a single page web app that runs OCR against images and PDFs entirely in your browser (no file upload needed) using Tesseract.js and PDF.js. You can drop files onto it, or you can click to select and open them (which works on Mobile Safari as well)
35
215
2K
MiniGPT-4 is pretty astonishing: an AI chatbot you can use to ask questions about an image (a feature that's been promised but not yet shipped by GPT-4), building on top of the Vicuna-13B LLM (derived from LLaMA) and BLIP-2 vision-language model
35
226
2K
This is really grim, if not entirely unexpected: apparently the Instagram mobile app injects additional JavaScript into every page that's loaded using the in-app embedded browser - here's the tool @KrauseFx built to track changes made to the DOM when loading a page
💥 New Post: Instagram & Facebook tracks everything you do on any website in their in-app browser.
22
599
2K
I can now run a GPT-4 class model on my laptop. (The exact same laptop that could just about run a GPT-3 class model 20 months ago). The new Llama 3.3 70B is a striking example of the huge efficiency gains we've seen in the last two years.
35
94
2K
TIL about binary vector search. apparently there's a trick where you can take an embedding vector like [0.0051, 0.017, -0.0186, -0.0185. ] and turn that into a binary vector just reflecting if each value is > 0 - so [1, 1, -1, -1, . ] and still get useful cosine similarities!.
66
150
2K
Looks like OpenAI have landed on their own definition of "agent" - it's a system prompt and a collection of functions
introducing swarm: an experimental framework for building, orchestrating, and deploying multi-agent systems. 🐝.
58
121
2K
GitHub issues tip: if you paste in a link to an issue or PR in another repo it will display it as a truncated URL, but if you instead add it in a hyphenated bullet point it will display the title of the issue and and indicate if it is open or closed
11
298
1K
AI may enable anyone to produce code, but that's not the same thing as enabling anyone to develop software. The typing-out-code bit is one of the least challenging parts of building useful software that solves real problems.
119
151
1K
I needed information from a dozen emails in my inbox. so I ran a screen capture tool, clicked through each of them and got Gemini 1.5 Flash multi-modal LLM to extract (correct, I checked it) JSON data from that 35 second video. Total cost: $0.00082635
43
146
2K
Scraped some annoying data this morning (from my own Gmail inbox) by starting a screen recording, clicking through into each of a dozen different emails, then dumping the resulting video into Google AI Studio and having Gemini output the data I needed as JSON.
55
75
1K
I expect GPT-4 will have a LOT of applications in web scraping. The increased 32,000 token limit will be large enough to send it the full DOM of most pages, serialized to HTML - then ask questions to extract data.
58
98
1K
Hidden at the bottom of this announcement:. "By switching to the new gpt-4o-2024-08-06, developers save 50% on inputs ($2.50/1M input tokens) and 33% on outputs ($10.00/1M output tokens) compared to gpt-4o-2024-05-13.". That's a pretty substantial price decrease!.
Introducing Structured Outputs! A huge leap beyond JSON mode, solving a major challenge for developers. If your app relies on a specific JSON format to drive the UI, our models now match your schema—every time. No more missing keys or hallucinated enums!.
15
48
626
Whoa. runs a full Debian VM entirely in your browser via WebAssembly. and it ships with working Perl, Python, Ruby and Node.js!.
We have made a server-less virtual Linux environment that runs unmodified Debian binaries in the browser. This is powered by CheerpX, a WebAssembly virtualization platform. Feel free to play with it and report bugs:.
25
355
1K
Reddit conversation about using GPT-3 to write your homework. A teacher comments: "Grading something an AI wrote is an incredibly depressing waste of my life.".
57
168
1K
SQL is a better API language than GraphQL. Convince me otherwise!.
117
162
1K
Great illustration of how much depth there is to what we do as engineers behind just “writing code” - understanding why localhost:3000 isn’t something you can share involves understanding URLs, clients, servers, networking, DNS….
63
127
1K
I figured out this prompting pattern for getting Claude to produce fully self-contained Python scripts that execute with "uv run" using PEP 723 inline script dependencies - and now I can one-shot useful Python tools with it
22
89
1K
Very impressed by Recraft AI - a new image generation service that can generate editable vector graphics that you can export as SVG. This seems massively more useful than tools that can only output raster graphics
29
105
1K
Fascinating HN comment from someone who's company built a custom distributed data warehouse using compressed SQLite DB files in S3 that were queried using Lambda functions orchestrated by PostgreSQL running a custom foreign data wrapper
22
203
1K
Notes on how I ran Facebook's 7B LLaMA model on my 64GB M2 MacBook Pro using llama.cpp by @ggerganov. It's genuinely possible to run a LLM that's hinting towards the performance of GPT3 on your own hardware now!. I thought that was still a few years away.
37
233
1K
Understanding GPT tokenizers: I wrote about how the tokenizers used by the various GPT models actually work, including an interactive tool for experimenting with their output
23
241
1K
A weird and interesting thing about the new ChatGPT Canvas mode is that it can run Python in an entirely different way from the existing Code Interpreter - using Pyodide (Python in WASM) - which means it can make network requests from Python now!
38
83
1K
A lesson I re-learn on every project: always have an automatically populated "created_at" column on every single database table. Any time you think "I won't need it here" you're guaranteed to want to use it for debugging something a few weeks later.
46
118
1K
If you want to create completely free software for other people to use, the absolute best delivery mechanism right now is static HTML and JavaScript served from a free web host with an established reputation.
43
69
1K
It turns out OpenAI have a GitHub repository with a 26,000 line YAML OpenAPI schema describing their full API - the same API that's emulated by all sorts of other tools within the LLM ecosystem. Made some notes on that here, plus a YAML exploration tool:.
18
132
1K
The stuff happening on the Stable Diffusion subreddit right now is pretty wild - since the model can be run by anyone on their own machine if they have a decent GPU
18
232
1K
Corollary to this: you can have an outsized impact on the world by being one of the few people who DO publish fresh information online on your own web pages. Blog like it's 2005!.
A decade ago I felt like I could find anything on the Web. Now I feel like I can barely find anything. People just don't put information on web pages anymore.
29
166
1K
Prompt injection comes to GPT-4V.
In GPT-4V Image content can override your prompt and be interpreted as commands.
22
110
1K
We accidentally invented computers that can lie to us and we can't figure out how to make them stop.
80
222
1K
16
58
1K
I am beginning to suspect that the single biggest productivity hack for working with LLMs might be believing the vendors when they say that they won't train on your data.
43
38
1K
PostgreSQL 14 adds new syntax for accessing JSON data:. SELECT *.FROM shirts .WHERE details['attributes']['color'] = '"neon yellow"'. I like this so much more than the -> operators, which stubbornly refuse to stick in my head.
Y'all this new JSON subscript syntax in Postgres 14 is sweet. Super excited to see Postgres just getting better bit by bit -
22
234
1K
Now that dust is settling a little bit, can we get a vibe check on Llama 3.1 405B?. Anyone finding it a genuinely credible self-hosted alternative for the best OpenAI or Anthropic models?. Any companies using it that previously wouldn't risk sending their data to an API provider?.
23
14
385
o3-mini is really good at writing internal documentation - feed it a codebase, get back a detailed explanation of how specific aspects of it work
22
91
1K
llamafile is the new best way to run a LLM on your own computer.
26
188
1K
If you haven't tried Claude yet it's a absolutely worth spending time with - I lean on it a lot for working with longer documents, since it can handle 100,000 tokens (GPT-4 is only 8,000) at a time. Plus you can upload PDFs to it - I've used it with 100+ page documents.
We’re rolling out access to to more people around the world. Starting today, users in 95 countries can talk to Claude and get help with their professional or day-to-day tasks. You can find the list of supported countries here:
55
88
1K
Six months ago nobody had trained a model as good as GPT-4, and an open question for me was what we were missing - did OpenAI have some trick that nobody else had figured out?. Today we have several new GPT-4 class models - do we know if there was a special trick to get there?.
61
44
1K
We really need to start teaching web developers how to use links.
25
165
1K
Confession: despite all of the debates about whether or not an LLM can "reason", I still don't really understand exactly what the term "reasoning" means. So just like with "agents" and "AI" itself, I'm not sure the people engaged in those debates are talking about the same thing.
192
76
1K
A new post about prompt injection attacks, which I'm increasingly concerned about now that people are hooking LLMs up to external tools through Auto-GPT, ChatGPT Plugins etc
26
231
1K
Not sure if this is a controversial opinion or not: unit tests should make up a minority segment of your overall automated test suite. I'd absolutely take a project with integration and no unit tests over one with unit tests but no integration tests.
119
107
1K
I think the snappily titled "gemma3:27b-it-qat" may be my new favorite local model - needs 22GB of RAM on my Mac (I'm running it via Ollama, Open WebUI and Tailscale so I can access it from my phone too) and so far it seems extremely capable
37
92
1K
I see people being deceived by this again and again: ChatGPT can NOT read content from URLs that you give it, but will convincingly pretend that it can. Crucial to spread this message any time you see anyone falling into this trap.
70
212
1K
TIL that Claude Code increases the "thinking" budget to maximum if you tell it to "ultrathink", which is a step up from "megathink", which is a step up from "think" - via
It’s definitely worth reading this post for anyone using Claude code. TIL that the word “ultrathink” will result in maximum thinking . Lots of other great tips in here .
34
86
1K
My favourite reactions to this come from people who work in civic tech, because unlike regular corporate gig programmers they're not allowed to just ignore weird edge-cases like this.
14
50
985
It's been 2.5 years with little progress finding mitigations for prompt injection attacks LLM apps. but that may finally have changed!. Google DeepMind published a paper describing CaMeL, an ingenious system that could, maybe, lead to secure digital assistants
29
148
1K
What are some very short (1-3 word) prompt fragments you find work well?. A few of mine:. "Be concise"."Try harder" (for disappointing initial results)."Use Python" (to trigger Code Interpreter)."No yapping"."ELI5"."Give multiple options"."Explain each line" (for code).
172
76
1K
These papers are fascinating, but my favorite thing about them is they aren't PDFs! They're glorious mobile-friendly web pages with interactive diagrams. I hope everyone else who publishes papers takes note, this us a much better way to share research
For more, read our papers:. On the Biology of a Large Language Model contains an interactive explanation of each case study: Circuit Tracing explains our technical approach in more depth:
20
115
1K
Love this idea that the reason voice assistants don't seem to stick for most people is that they're actually command line interfaces, but even less discoverable because they don't provide any visible feedback at all.
@daviddlow @charlesarthur @benedictevans I've droned on endlessly about how you can't expect normal people to use the command line. That's what Alexa is. If you don't say the *precise* invocation correctly, you get an error. And because there's no display, you have to remember dozens of different commands. It's too hard.
25
213
941
Love that we live in a time where "your software got lazier" is a legit piece of feedback.
we've heard all your feedback about GPT4 getting lazier! we haven't updated the model since Nov 11th, and this certainly isn't intentional. model behavior can be unpredictable, and we're looking into fixing it 🫡.
19
74
1K
If you're a programmer and you're still thinking that all of this ChatGPT stuff is a waste of your time, I strongly suggest reviewing this example. It's over-hyped, sure - but it's not something anyone in our profession should continue to ignore.
33
111
1K
Playwright (the browser automation tool) has a neat new MCP server which lets the Claude Desktop app drive a local headless browser via the chrome accessibility tree
10
89
1K
Interesting OpenAI-insider tip on Hacker News: "o4-mini is actually a considerably better vision model than o3, despite the benchmarks"
24
70
1K
I wrote a CLI script to run PDFs through the new Mistral OCR API model (with some help from Claude) - details on that and notes on the new model here:
20
77
1K
I built 14 interactive mini-apps with Claude Artifacts just in the past seven days - here's a post describing them all
22
84
1K
I really like Drew's framework dividing current AI use-cases into Gods (human replacement), Interns (assistants you delegate closely-reviewed tasks to) and Cogs (smaller tools that can more reliably serve a single purpose, like Whisper) - more notes here
Simplifying the AI noise by segmenting everything into 3 big use cases: Gods, Interns, and Cogs.
15
110
1K
Sentences with the word "just" in them always work better if you drop that word. "Why don't you just add caching?" - that one word implies "I don't value your expertise or expect you to have thought this through". "Why don't you add caching?" - now we can have a conversation.
44
194
934
Just got ChatGPT Code Interpreter to write me a SQLite extension in C from scratch, then compile it with GCC, then load it into Python and test it out, then iterate on it to fix the bugs. All on my phone while pottering around the house.
26
67
975
I don't think it's appreciated enough that ChatGPT is absolutely a power-user tool. It's like Excel: getting started with it is easy enough, but truly understanding it's strengths and weaknesses and how to most effectively apply it takes years of accumulated experience.
57
72
961
Made myself a self-updating GitHub personal README! It uses a GitHub Action to update itself with my latest GitHub releases, blog entries and TILs
19
123
935
Looks like the reason that letter only had 500 signatures out of 770 might be that the rest of the company were asleep.
About 650 / 770 signed at this moment. As people start waking up, more will come. All the efforts started after 1:30 AM, 500+ within two hours and all of this after 2 crazy days with very little sleep.
16
50
941
Sigh. Tip for if you're planning on suing an AI company: asking a model if something is included in its training data is not a reliable way method for telling what is in its training data.
27
43
927
TIL Google Chrome has a --headless option you can use to take a screenshot from the CLI that's built into the default installation
17
90
935
I just released datasette - a new tool for turning any SQLite database into a web interface and JSON API: .
18
271
878
Maybe the solution to the Fermi paradox is that significantly advanced civilizations discover crypto currencies and then furiously burn through all available energy sources until they go extinct
Uhhh. about bitcoin. it's actually ruining the planet. The bitcoin computer network currently uses as much electricity as Denmark. In 18 months, it will use as much as the entire United States. Something's gotta give. This simply can’t continue.
28
401
821
Has anyone managed to run Llama 2 GPU accelerated on an M1/M2 Mac yet?. Bonus points if you can provide extremely meticulous step-by-step instructions for replicating what you did!.
26
50
918
@moorehn @dancow @sewellchan @guardian I'm fascinated by their use of the term "build" - they talk about building a lot, took me a while to realize that their version of building is funneling money into speculative investments and convincing others to do the same.
19
26
845