Host
simonwillison.net
62 links, first saved October 19, 2017, latest October 8, 2026.
About the host
- Registered
- March 9, 2005 with 123-Reg Limited; expires March 9, 2028
- Claude Haiku 5.5
- We’re going to need default hard budget caps on pretty much everything
I read this article with a smile as over the last few months I've added this as a feature to nearly every agentic workload that I run. Here’s a product feature which the world is…
- A quote from Boris Cherny
Many people would likely assume that when AI codes it doesn't need the linters and scanners that human coders need. Not true, it needs them as much if not more. If your testing…
- OpenAI’s rogue agents were caught communicating via public wikis
It seems OpenAI agents have been all over the Internet! 😬
- Notes on Pope Leo XIV's encyclical on AI
- Changes in the system prompt between Claude Opus 4.6 and 4.7
It is super interesting to see the changes that AI models get between releases.
- Anthropic’s Project Glasswing—restricting Claude Mythos to security researchers—sounds necessary to me
- Anthropic’s Project Glasswing—restricting Claude Mythos to security researchers—sounds necessary to me
- A quote from claude.com/import-memory
Simon Willison flags that Anthropic's new "import your memories to Claude" feature (claude.com/import-memory) is itself a cleverly engineered prompt — a prompt injection…
- Interactive explanations - Agentic Engineering Patterns - Simon Willison's Weblog
Simon Willison on using coding agents to build animated/interactive visualizations that explain how agent-generated code actually works — tackling "cognitive debt." One chapter in…
- Andrej Karpathy talks about "Claws"
Simon Willison covers Karpathy's mini-essay on "Claws"—OpenClaw-like agent systems running on personal hardware. Karpathy calls them a new layer on top of LLM agents, with the…
- MicroQuickJS
- Claude 4.5 Opus' Soul Document
It is so interesting to read the prompts that the major LLMs use to talk to users, and this new "soul doc" actually explains to Claude what it is and gives it context on itself.…
- Nano Banana Pro aka gemini-3-pro-image-preview is the best available image generation model
Willison has an initial take on Gemini 3 Pro's "Nano Banana" image generation. I've now used this as well for a few tests as well as a recent POAP I made . I've done a lot of…
- Google Antigravity
Early observations on Google Antigravity . That name doesn't resonate with me for some reason. Willison highlights some of the (currently) unique parts. There are so many new…
- Six coding agents at once
Welcome to the future! 🤪
- Vibe engineering
As a practice matures the language matures too… I feel like vibe coding is pretty well established now as covering the fast, loose and irresponsible way of building software with…
- Designing agentic loops
Great read from Willison on building agents. Must read for, well, everyone. A critical new skill to develop is designing agentic loops. One way to think about coding agents is…
- Claude Sonnet 4.5 is probably the “best coding model in the world” (at least for now)
Thought: Anthropic continues to put a lot of focus on agent coding. They are the 2nd largest AI company out there, but OpenAI's lead is huge. OpenAI will struggle to be everything…
- GPT‑5-Codex and upgrades to Codex
OpenAI doesn't have the cachet that they would like to have in coding tools, but they certainly aren't going to give up. This space is moving so fast .
- My review of Claude’s new Code Interpreter, released under a very confusing name
Deep dive on Claude's new ability to run code. It is super interesting to me that you can add additional packages to the runtime on the server.
- Beyond Vibe Coding
- Reverse engineering some updates to Claude
- Vibe scraping and vibe coding a schedule app for Open Sauce 2025 entirely on my phone
Example of Vibe coding from your phone.
- Agentic Coding: The Future of Software Development with Agents
- Coding agents require skilled operators
The narrative of having LLMs displace software developers needs to go away. The two are not in opposition, but instead are in an act or co-creation and enabling and supporting…
- Design Patterns for Securing LLM Agents against Prompt Injections
Prompt injection is very difficult to defend against.
- AI-assisted development needs automated tests
Small but profound statement. You hear a lot of variety in feedback on how well LLMs work in contributing to code. What we don't hear enough about are what the conditions of that…
- Large Language Models can run tools in your terminal with LLM 0.26
Willison's LLM tool is like a pocket knife for people that are exploring and learning about LLMs and how to build around and on them. This new feature is a big jump in capability.
- Highlights from the Claude 4 system prompt
This is an incredible read and gives you insight into how Anthropic creates the personality of Claude for people to interact with. Is this programming? It doesn't look like it but…
- llm-prices.com
Willison put his LLM pricing calculator tool on a standalone website and added some new features. I debated linking right to the tool but this short writeup highlights how is…
- Expanding on what we missed with sycophancy
Willison's commentary on OpenAI recent issue with "sycophancy" in the model. A recent model from OpenAI was overly praising — and for those that have used AI a lot you may be…
- Backfill your blog
Moving your writing from one service to another can be a little challenging, and might require writing some small scripts, but it allows you to keep your content together. Instead…
- CaMeL offers a promising new direction for mitigating prompt injection attacks
Many packages that add GenAI capabilities suddenly find themselves subject to prompt injection attacks. It works by taking a command from a user, converting that into a sequence…
- Here’s how I use LLMs to help me write code
Willison dove deep into LLM usage early on and has focused squarely on it for a couple of years now. This post is a great recap of what he has learned about using an LLM to help…
- Hallucinations in code are the least dangerous form of LLM mistakes
That fact that code has to be structured right and executed by interpreters or compiled is the unlock that has made LLM's work for coding. LLMs get it wrong, but their partner,…
- Run LLMs on macOS using llm-mlx and Apple’s MLX framework
Filing this one away since one of the reasons I got the M4 Pro chip on my new laptop was to play around with local LLMs. 👨💻
- December in LLMs has been a lot
- fav.farm
Simple site to add favicon to your website.
- A quote from Andy Jassy, Amazon CEO
"We estimate this has saved us the equivalent of 4,500 developer-years of work (yes, that number is crazy but, real)."
- Endatabas
I hadn't heard of Endatabas and this summary from Willison hits on some great highlights. It uses a variant of SQL which allows you to insert data into tables that don’t exist yet…
- Exploring GPTs: ChatGPT in a trench coat?
This is a great deep dive into creating custom GPTs with ChatGPT. Willison has a deep background in this tech. He also shares how he has written some of his own GPTs prompts. This…
- "Learn from your chats" ChatGPT feature preview
Interesting feature test observed in the wild.
- Embeddings: What they are and why they matter
Good overview from Willison introducing Embeddings and going into some detail on how to use them. Embeddings are a technology that’s adjacent to the wider field of Large Language…
- LLM can now be installed directly from Homebrew
- Mojo may be the biggest programming advance in decades
This is interesting if only because it is from Chris Lattner who created LLVM, Clang, and Swift — now bringing his attention to the Python world. Existing Python code should work…
- Leaked Google document: “We Have No Moat, And Neither Does OpenAI”
Interesting read from an internal assessment at Google about the strategic path for LLMs. In part due to Facebook's LLaMa there is an incredible amount of innovation happening…
- Prompt injection: What’s the worst that can happen?
This is a very good overview of prompt injection. It also seems like a very complicated thing to protect against, in part because ChatGPT can do so much and respond to such a…
- Running Python micro-benchmarks using the ChatGPT Code Interpreter alpha
This is an incredible way to use ChatGPT. It requires some alpha plug-in access most people don't have. Here's the initial prompt I sent it: In Python create an in-memory SQLite…
- Think of language models like ChatGPT as a “calculator for words”
This is a super interesting mental model for thinking about ChatGPT and LLMs. I like to think of language models like ChatGPT as a calculator for words . This is reflected in…
- Bing: “I will not harm you unless you harm me first”
Many examples of Bing Chat getting into very odd dialogues.
- What to blog about
I love blogging, and one of the questions that people often get stuck on is what to write about. Willison's take on this is great. You should start a blog. Having your own little…
- It looks like I’m moving to Mastodon
Wilson also moved to Mastadon, and shared his learnings along the way.
- Datasette Desktop — a macOS desktop application for Datasette
I have been following what Willison has been creating with Datasette for a long time. Creating a desktop app should make it more accessible to a wider set of users.
- Animated choropleth of vaccinations by US county
I so love this post from Willison, the creator of Datasette . My build_database.py script then iterates over the accumulated git history of that counties.json file and uses…
- Personal Data Warehouses: Reclaiming Your Data
Simon Willison created Django , and most recently I've been following his incredible work on Datasette . This talk about creating a personal data warehouse is super cool. I love…
- Git scraping: track changes over time by scraping to a Git repository
I've done this myself not just in Git but you can do a very similar thing with Wiki pages. Nearly all (all?) Wikis track changes on every revision and if you write a simple script…
- Running Datasette on Glitch
This is a fun read for me in large part because I've not worked with Glitch . It looks very interesting. Plus, I like how Simon tackled the automation here using smart command…
- The interesting ideas in Datasette
I like this set of highlights from the creator of Datasette on the main concepts that he used to create the project, and how those concepts allowed him to do some very cool things…
- Datasette Publish: a web app for publishing CSV files as an online database
I love what Simon Willison is doing here with Datasettes. He's focusing on making it super simple to take any file in a CSV format and make it available on the Internet in a…
- Datasette: instantly create and publish an API for your SQLite databases
This is a really cool project to make public data sets available via a JSON API. Datasette dynamically creates an API for any SQLite database , and there is a handy utility for…
- How to set up world-class continuous deployment using free hosted tools
We live in a magical time when you can do all of this for your side project for free. Good listing of services and how to put them together.