Recent
Sept. 19, 2024
Moshi (via) Moshi is "a speech-text foundation model and full-duplex spoken dialogue framework". It's effectively a text-to-text model - like an LLM but you input audio directly to it and it replies with its own audio.
It's fun to play around with, but it's not particularly useful in comparison to other pure text models: I tried to talk to it about California Brown Pelicans and it gave me some very basic hallucinated thoughts about California Condors instead.
It's very easy to run locally, at least on a Mac (and likely on other systems too). I used uv and got the 8 bit quantized version running as a local web server using this one-liner:
uv run --with moshi_mlx python -m moshi_mlx.local_web -q 8
That downloads ~8.17G of model to a folder in ~/.cache/huggingface/hub/ - or you can use -q 4 and get a 4.81G version instead (albeit even lower quality).
The web’s clipboard, and how it stores data of different types.
Alex Harri's deep dive into the Web clipboard API, the more recent alternative to the old document.execCommand() mechanism for accessing the clipboard.
There's a lot to understand here! Some of these APIs have a history dating back to Internet Explorer 4 in 1997, and there have been plenty of changes over the years to account for improved understanding of the security risks of allowing untrusted code to interact with the system clipboard.
Today, the most reliable data formats for interacting with the clipboard are the "standard" formats of text/plain, text/html and image/png.
Figma does a particularly clever trick where they share custom Figma binary data structures by encoding them as base64 in data-metadata and data-buffer attributes on a <span> element, then write the result to the clipboard as HTML. This enables copy-and-paste between the Figma web and native apps via the system clipboard.
Sept. 18, 2024
The problem that you face is that it's relatively easy to take a model and make it look like it's aligned. You ask GPT-4, “how do I end all of humans?” And the model says, “I can't possibly help you with that”. But there are a million and one ways to take the exact same question - pick your favorite - and you can make the model still answer the question even though initially it would have refused. And the question this reminds me a lot of coming from adversarial machine learning. We have a very simple objective: Classify the image correctly according to the original label. And yet, despite the fact that it was essentially trivial to find all of the bugs in principle, the community had a very hard time coming up with actually effective defenses. We wrote like over 9,000 papers in ten years, and have made very very very limited progress on this one small problem. You all have a harder problem and maybe less time.
Things I’ve learned serving on the board of the Python Software Foundation
Two years ago I was elected to the board of directors for the Python Software Foundation—the PSF. I recently returned from the annual PSF board retreat (this one was in Lisbon, Portugal) and this feels like a good opportunity to write up some of the things I’ve learned along the way.
[... 2,703 words]Sept. 17, 2024
Oracle, it’s time to free JavaScript. (via) Oracle have held the trademark on JavaScript since their acquisition of Sun Microsystems in 2009. They’ve continued to renew that trademark over the years despite having no major products that use the mark.
Their December 2019 renewal included a screenshot of the Node.js homepage as a supporting specimen!
Now a group lead by a team that includes Ryan Dahl and Brendan Eich is coordinating a legal challenge to have the USPTO treat the trademark as abandoned and “recognize it as a generic name for the world’s most popular programming language, which has multiple implementations across the industry.”
In general, the claims about how long people are living mostly don’t stack up. I’ve tracked down 80% of the people aged over 110 in the world (the other 20% are from countries you can’t meaningfully analyse). Of those, almost none have a birth certificate. [...]
Regions where people most often reach 100-110 years old are the ones where there’s the most pressure to commit pension fraud, and they also have the worst records.
Something that I confirmed that other conference organisers are also experiencing is last-minute ticket sales. This is something that happened with UX London this year. For most of the year, ticket sales were trickling along. Then in the last few weeks before the event we sold more tickets than we had sold in the six months previously. […]
When I was in Ireland I had a chat with a friend of mine who works at the Everyman Theatre in Cork. They’re experiencing something similar. So maybe it’s not related to the tech industry specifically.
Serializing package requirements in marimo notebooks. The latest release of Marimo - a reactive alternative to Jupyter notebooks - has a very neat new feature enabled by its integration with uv:
One of marimo’s goals is to make notebooks reproducible, down to the packages used in them. To that end, it’s now possible to create marimo notebooks that have their package requirements serialized into them as a top-level comment.
This takes advantage of the PEP 723 inline metadata mechanism, where a code comment at the top of a Python file can list package dependencies (and their versions).
I tried this out by installing marimo using uv:
uv tool install --python=3.12 marimo
Then grabbing one of their example notebooks:
wget 'https://raw.githubusercontent.com/marimo-team/spotlights/main/001-anywidget/tldraw_colorpicker.py'
And running it in a fresh dependency sandbox like this:
marimo run --sandbox tldraw_colorpicker.py
Also neat is that when editing a notebook using marimo edit:
marimo edit --sandbox notebook.py
Just importing a missing package is enough for Marimo to prompt to add that to the dependencies - at which point it automatically adds that package to the comment at the top of the file:
Supercharging Developer Productivity with ChatGPT and Claude with Simon Willison (via) I'm the guest for the latest episode of the TWIML AI podcast - This Week in Machine Learning & AI, hosted by Sam Charrington.
We mainly talked about how I use LLM tooling for my own work - Claude, ChatGPT, Code Interpreter, Claude Artifacts, LLM and GitHub Copilot - plus a bit about my experiments with local models.
Do not fall into the trap of anthropomorphizing Larry Ellison. You need to think of Larry Ellison the way you think of a lawnmower. You don’t anthropomorphize your lawnmower, the lawnmower just mows the lawn - you stick your hand in there and it’ll chop it off, the end. You don’t think "oh, the lawnmower hates me" – lawnmower doesn’t give a shit about you, lawnmower can’t hate you. Don’t anthropomorphize the lawnmower. Don’t fall into that trap about Oracle.
Sept. 16, 2024
o1 prompting is alien to me. Its thinking, gloriously effective at times, is also dreamlike and unamenable to advice.
Just say what you want and pray. Any notes on “how” will be followed with the diligence of a brilliant intern on ketamine.
Sept. 15, 2024
UV — I am (somewhat) sold
(via)
Oliver Andrich's detailed notes on adopting uv. Oliver has some pretty specific requirements:
I need to have various Python versions installed locally to test my work and my personal projects. Ranging from Python 3.8 to 3.13. [...] I also require decent dependency management in my projects that goes beyond manually editing a
pyproject.tomlfile. Likewise, I am way too accustomed topoetry add .... And I run a number of Python-based tools --- djhtml, poetry, ipython, llm, mkdocs, pre-commit, tox, ...
He's braver than I am!
I started by removing all Python installations, pyenv, pipx and Homebrew from my machine. Rendering me unable to do my work.
Here's a neat trick: first install a specific Python version with uv like this:
uv python install 3.11
Then create an alias to run it like this:
alias python3.11 'uv run --python=3.11 python3'
And install standalone tools with optional extra dependencies like this (a replacement for pipx and pipx inject):
uv tool install --python=3.12 --with mkdocs-material mkdocs
Oliver also links to Anže Pečar's handy guide on using UV with Django.
How to succeed in MrBeast production (leaked PDF). Whether or not you enjoy MrBeast’s format of YouTube videos (here’s a 2022 Rolling Stone profile if you’re unfamiliar), this leaked onboarding document for new members of his production company is a compelling read.
It’s a snapshot of what it takes to run a massive scale viral YouTube operation in the 2020s, as well as a detailed description of a very specific company culture evolved to fulfill that mission.
It starts in the most on-brand MrBeast way possible:
I genuinely believe if you attently read and understand the knowledge here you will be much better set up for success. So, if you read this book and pass a quiz I’ll give you $1,000.
Everything is focused very specifically on YouTube as a format:
Your goal here is to make the best YOUTUBE videos possible. That’s the number one goal of this production company. It’s not to make the best produced videos. Not to make the funniest videos. Not to make the best looking videos. Not the highest quality videos.. It’s to make the best YOUTUBE videos possible.
The MrBeast definition of A, B and C-team players is one I haven’t heard before:
A-Players are obsessive, learn from mistakes, coachable, intelligent, don’t make excuses, believe in Youtube, see the value of this company, and are the best in the goddamn world at their job. B-Players are new people that need to be trained into A-Players, and C-Players are just average employees. […] They arn’t obsessive and learning. C-Players are poisonous and should be transitioned to a different company IMMEDIATELY. (It’s okay we give everyone severance, they’ll be fine).
The key characteristic outlined here, if you read between the hustle-culture lines, is learning. Employees who constantly learn are valued. Employees who don’t are not.
There’s a lot of stuff in there about YouTube virality, starting with the Click Thru Rate (CTR) for the all-important video thumbnails:
This is what dictates what we do for videos. “I Spent 50 Hours In My Front Yard” is lame and you wouldn’t click it. But you would hypothetically click “I Spent 50 Hours In Ketchup”. Both are relatively similar in time/effort but the ketchup one is easily 100x more viral. An image of someone sitting in ketchup in a bathtub is exponentially more interesting than someone sitting in their front yard.
The creative process for every video they produce starts with the title and thumbnail. These set the expectations for the viewer, and everything that follows needs to be defined with those in mind. If a viewer feels their expectations are not being matched, they’ll click away - driving down the crucial Average View Duration that informs how much the video is promoted by YouTube’s all-important mystical algorithms.
MrBeast videos have a strictly defined formula, outlined in detail on pages 6-10.
The first minute captures the viewer’s attention and demonstrates that their expectations from the thumbnail will be met. Losing 21 million viewers in the first minute after 60 million initial clicks is considered a reasonably good result! Minutes 1-3, 3-6 and 6-end all have their own clearly defined responsibilities as well.
Ideally, a video will feature something they call the “wow factor”:
An example of the “wow factor” would be our 100 days in the circle video. We offered someone $500,000 if they could live in a circle in a field for 100 days (video) and instead of starting with his house in the circle that he would live in, we bring it in on a crane 30 seconds into the video. Why? Because who the fuck else on Youtube can do that lol.
Chapter 2 (pages 10-24) is about creating content. This is crammed with insights into what it takes to produce surprising, spectacular and very expensive content for YouTube.
A lot of this is about coordination and intense management of your dependencies:
I want you to look them in the eyes and tell them they are the bottleneck and take it a step further and explain why they are the bottleneck so you both are on the same page. “Tyler, you are my bottleneck. I have 45 days to make this video happen and I can not begin to work on it until I know what the contents of the video is. I need you to confirm you understand this is important and we need to set a date on when the creative will be done.” […] Every single day you must check in on Tyler and make sure he is still on track to hit the target date.
It also introduces the concept of “critical components”:
Critical components are the things that are essential to your video. If I want to put 100 people on an island and give it away to one of them, then securing an island is a critical component. It doesn’t matter how well planned the challenges on the island are, how good the weather is, etc. Without that island there is no video.
[…]
Critical Components can come from literally anywhere and once something you’re working on is labeled as such, you treat it like your baby. WITHOUT WHAT YOU’RE WORKING ON WE DO NOT HAVE A VIDEO! Protect it at all costs, check in on it 10x a day, obsess over it, make a backup, if it requires shipping pay someone to pick it up and drive it, don’t trust standard shipping, and speak up the second anything goes wrong. The literal second. Never coin flip a Critical Component (that means you’re coinfliping the video aka a million plus dollars)
There’s a bunch of stuff about communication, with a strong bias towards “higher forms of communication”: in-person beats a phone call beats a text message beats an email.
Unsurprisingly for this organization, video is a highly valued tool for documenting work:
Which is more important, that one person has a good mental grip of something or that their entire team of 10 people have a good mental grip on something? Obviously the team. And the easiest way to bring your team up to the same page is to freaken video everything and store it where they can constantly reference it. A lot of problems can be solved if we just video sets and ask for videos when ordering things.
I enjoyed this note:
Since we are on the topic of communication, written communication also does not constitute communication unless they confirm they read it.
And this bit about the value of consultants:
Consultants are literally cheat codes. Need to make the world's largest slice of cake? Start off by calling the person who made the previous world’s largest slice of cake lol. He’s already done countless tests and can save you weeks worth of work. […] In every single freakin task assigned to you, always always always ask yourself first if you can find a consultant to help you.
Here’s a darker note from the section “Random things you should know”:
Do not leave consteatants waiting in the sun (ideally waiting in general) for more than 3 hours. Squid game it cost us $500,000 and boys vs girls it got a lot of people out. Ask James to know more
And to finish, this note on budgeting:
I want money spent to be shown on camera ideally. If you’re spending over $10,000 on something and it won’t be shown on camera, seriously think about it.
I’m always interested in finding management advice from unexpected sources. For example, I love The Eleven Laws of Showrunning as a case study in managing and successfully delegating for a large, creative project.
I don’t think this MrBeast document has as many lessons directly relevant to my own work, but as an honest peek under the hood of a weirdly shaped and absurdly ambitious enterprise it’s legitimately fascinating.
Speed matters (via) Jamie Brandon in 2021, talking about the importance of optimizing for the speed at which you can work as a developer:
Being 10x faster also changes the kinds of projects that are worth doing.
Last year I spent something like 100 hours writing a text editor. […] If I was 10x slower it would have been 20-50 weeks. Suddenly that doesn't seem like such a good deal any more - what a waste of a year!
It’s not just about speed of writing code:
When I think about speed I think about the whole process - researching, planning, designing, arguing, coding, testing, debugging, documenting etc.
Often when I try to convince someone to get faster at one of those steps, they'll argue that the others are more important so it's not worthwhile trying to be faster. Eg choosing the right idea is more important than coding the wrong idea really quickly.
But that's totally conditional on the speed of everything else! If you could code 10x as fast then you could try out 10 different ideas in the time it would previously have taken to try out 1 idea. Or you could just try out 1 idea, but have 90% of your previous coding time available as extra idea time.
Jamie’s model here helps explain the effect I described in AI-enhanced development makes me more ambitious with my projects. Prompting an LLM to write portions of my code for me gives me that 5-10x boost in the time I spend typing code into a computer, which has a big effect on my ambitions despite being only about 10% of the activities I perform relevant to building software.
I also increasingly lean on LLMs as assistants in the research phase - exploring library options, building experimental prototypes - and for activities like writing tests and even a little bit of documentation.
[… OpenAI’s o1] could work its way to a correct (and well-written) solution if provided a lot of hints and prodding, but did not generate the key conceptual ideas on its own, and did make some non-trivial mistakes. The experience seemed roughly on par with trying to advise a mediocre, but not completely incompetent, graduate student. However, this was an improvement over previous models, whose capability was closer to an actually incompetent graduate student.
Sept. 14, 2024
It's a bit sad and confusing that LLMs ("Large Language Models") have little to do with language; It's just historical. They are highly general purpose technology for statistical modeling of token streams. A better name would be Autoregressive Transformers or something.
They don't care if the tokens happen to represent little text chunks. It could just as well be little image patches, audio chunks, action choices, molecules, or whatever. If you can reduce your problem to that of modeling token streams (for any arbitrary vocabulary of some set of discrete tokens), you can "throw an LLM at it".
Notes on running Go in the browser with WebAssembly (via) Neat, concise tutorial by Eli Bendersky on compiling Go applications that can then be loaded into a browser using WebAssembly and integrated with JavaScript. Go functions can be exported to JavaScript like this:
js.Global().Set("calcHarmonic", jsCalcHarmonic)
And Go code can even access the DOM using a pattern like this:
doc := js.Global().Get("document")
inputElement := doc.Call("getElementById", "timeInput")
input := inputElement.Get("value")
Bundling the WASM Go runtime involves a 2.5MB file load, but there’s also a TinyGo alternative which reduces that size to a fourth.
Sept. 13, 2024
Believe it or not, the name Strawberry does not come from the “How many r’s are in strawberry” meme. We just chose a random word. As far as we know it was a complete coincidence.
There is superstition about creativity, and for that matter, about thinking in every sense, and it's part of the history of the field of artificial intelligence that every time somebody figured out how to make a computer do something - play good checkers, solve simple but relatively informal problems - there was a chorus of critics to say, but that's not thinking.
Sept. 12, 2024
o1-mini is the most surprising research result I've seen in the past year
Obviously I cannot spill the secret, but a small model getting >60% on AIME math competition is so good that it's hard to believe
LLM 0.16.
New release of LLM adding support for the o1-preview and o1-mini OpenAI models that were released today.
Notes on OpenAI’s new o1 chain-of-thought models
OpenAI released two major new preview models today: o1-preview and o1-mini (that mini one is not a preview)—previously rumored as having the codename “strawberry”. There’s a lot to understand about these models—they’re not as simple as the next step up from GPT-4o, instead introducing some major trade-offs in terms of cost and performance in exchange for improved “reasoning” capabilities.
Sept. 11, 2024
Pixtral 12B. Mistral finally have a multi-modal (image + text) vision LLM!
I linked to their tweet, but there’s not much to see there - in now classic Mistral style they released the new model with an otherwise unlabeled link to a torrent download. A more useful link is mistral-community/pixtral-12b-240910 on Hugging Face, a 25GB “Unofficial Mistral Community” copy of the weights.
Pixtral was announced at Mistral’s AI Summit event in San Francisco today. It has 128,000 token context, is Apache 2.0 licensed and handles 1024x1024 pixel images. They claim it’s particularly good for OCR and information extraction. It’s not available on their La Platforme hosted API yet, but that’s coming soon.
A few more details can be found in the release notes for mistral-common 1.4.0. That’s their open source library of code for working with the models - it doesn’t actually run inference, but it includes the all-important tokenizer, which now includes three new special tokens: [IMG], [IMG_BREAK] and [IMG_END].
Sept. 10, 2024
Notes from my appearance on the Software Misadventures Podcast
I was a guest on Ronak Nathani and Guang Yang’s Software Misadventures Podcast, which interviews seasoned software engineers about their careers so far and their misadventures along the way. Here’s the episode: LLMs are like your weird, over-confident intern | Simon Willison (Datasette).
[... 1,740 words]Telling the AI to "make it better" after getting a result is just a folk method of getting an LLM to do Chain of Thought, which is why it works so well.
Sept. 9, 2024
Why GitHub Actually Won (via) GitHub co-founder Scott Chacon shares some thoughts on how GitHub won the open source code hosting market. Shortened to two words: timing, and taste.
There are some interesting numbers in here. I hadn't realized that when GitHub launched in 2008 the term "open source" had only been coined ten years earlier, in 1998. This paper by Dirk Riehle estimates there were 18,000 open source projects in 2008 - Scott points out that today there are over 280 million public repositories on GitHub alone.
Scott's conclusion:
We were there when a new paradigm was being born and we approached the problem of helping people embrace that new paradigm with a developer experience centric approach that nobody else had the capacity for or interest in.
files-to-prompt 0.3.
New version of my files-to-prompt CLI tool for turning a bunch of files into a prompt suitable for piping to an LLM, described here previously.
It now has a -c/--cxml flag for outputting the files in Claude XML-ish notation (XML-ish because it's not actually valid XML) using the format Anthropic describe as recommended for long context:
files-to-prompt llm-*/README.md --cxml | llm -m claude-3.5-sonnet \
--system 'return an HTML page about these plugins with usage examples' \
> /tmp/fancy.html
The format itself looks something like this:
<documents>
<document index="1">
<source>llm-anyscale-endpoints/README.md</source>
<document_content>
# llm-anyscale-endpoints
...
</document_content>
</document>
</documents>Sept. 8, 2024
uv under discussion on Mastodon. Jacob Kaplan-Moss kicked off this fascinating conversation about uv on Mastodon recently. It's worth reading the whole thing, which includes input from a whole range of influential Python community members such as Jeff Triplett, Glyph Lefkowitz, Russell Keith-Magee, Seth Michael Larson, Hynek Schlawack, James Bennett and others. (Mastodon is a pretty great place for keeping up with the Python community these days.)
The key theme of the conversation is that, while uv represents a huge set of potential improvements to the Python ecosystem, it comes with additional risks due its attachment to a VC-backed company - and its reliance on Rust rather than Python.
Here are a few comments that stood out to me.
As enthusiastic as I am about the direction uv is going, I haven't adopted them anywhere - because I want very much to understand Astral’s intended business model before I hook my wagon to their tools. It's definitely not clear to me how they're going to stay liquid once the VC money runs out. They could get me onboard in a hot second if they published a "This is what we're planning to charge for" blog post.
As much as I hate VC, [...] FOSS projects flame out all the time too. If Frost loses interest, there’s no PDM anymore. Same for Ofek and Hatch(ling).
I fully expect Astral to flame out and us having to fork/take over—it’s the circle of FOSS. To me uv looks like a genius sting to trick VCs into paying to fix packaging. We’ll be better off either way.
Even in the best case, Rust is more expensive and difficult to maintain, not to mention "non-native" to the average customer here. [...] And the difficulty with VC money here is that it can burn out all the other projects in the ecosystem simultaneously, creating a risk of monoculture, where previously, I think we can say that "monoculture" was the least of Python's packaging concerns.
I don’t think y’all quite grok what uv makes so special due to your seniority. The speed is really cool, but the reason Rust is elemental is that it’s one compiled blob that can be used to bootstrap and maintain a Python development. A blob that will never break because someone upgraded Homebrew, ran pip install or any other creative way people found to fuck up their installations. Python has shown to be a terrible tech to maintain Python.
Just dropping in here to say that corporate capture of the Python ecosystem is the #1 keeps-me-up-at-night subject in my community work, so I watch Astral with interest, even if I'm not yet too worried.
I'm reminded of this note from Armin Ronacher, who created Rye and later donated it to uv maintainers Astral:
However having seen the code and what uv is doing, even in the worst possible future this is a very forkable and maintainable thing. I believe that even in case Astral shuts down or were to do something incredibly dodgy licensing wise, the community would be better off than before uv existed.
I'm currently inclined to agree with Armin and Hynek: while the risk of corporate capture for a crucial aspect of the Python packaging and onboarding ecosystem is a legitimate concern, the amount of progress that has been made here in a relatively short time combined with the open license and quality of the underlying code keeps me optimistic that uv will be a net positive for Python overall.
Update: uv creator Charlie Marsh joined the conversation:
I don't want to charge people money to use our tools, and I don't want to create an incentive structure whereby our open source offerings are competing with any commercial offerings (which is what you see with a lost of hosted-open-source-SaaS business models).
What I want to do is build software that vertically integrates with our open source tools, and sell that software to companies that are already using Ruff, uv, etc. Alternatives to things that companies already pay for today.
An example of what this might look like (we may not do this, but it's helpful to have a concrete example of the strategy) would be something like an enterprise-focused private package registry. A lot of big companies use uv. We spend time talking to them. They all spend money on private package registries, and have issues with them. We could build a private registry that integrates well with uv, and sell it to those companies. [...]
But the core of what I want to do is this: build great tools, hopefully people like them, hopefully they grow, hopefully companies adopt them; then sell software to those companies that represents the natural next thing they need when building with Python. Hopefully we can build something better than the alternatives by playing well with our OSS, and hopefully we are the natural choice if they're already using our OSS.
Teresa T is name of the whale in Pillar Point Harbor near Half Moon Bay
There is a young humpback whale in the harbor at Pillar Point, just north of Half Moon Bay, California right now. Their name is Teresa T and they were first spotted on Thursday afternoon.
[... 254 words]Sept. 7, 2024
json-flatten, now with format documentation.
json-flatten is a fun little Python library I put together a few years ago for converting JSON data into a flat key-value format, suitable for inclusion in an HTML form or query string. It lets you take a structure like this one:
{"foo": {"bar": [1, True, None]}
And convert it into key-value pairs like this:
foo.bar.[0]$int=1
foo.bar.[1]$bool=True
foo.bar.[2]$none=None
The flatten(dictionary) function function converts to that format, and unflatten(dictionary) converts back again.
I was considering the library for a project today and realized that the 0.3 README was a little thin - it showed how to use the library but didn't provide full details of the format it used.
On a hunch, I decided to see if files-to-prompt plus LLM plus Claude 3.5 Sonnet could write that documentation for me. I ran this command:
files-to-prompt *.py | llm -m claude-3.5-sonnet --system 'write detailed documentation in markdown describing the format used to represent JSON and nested JSON as key/value pairs, include a table as well'
That *.py picked up both json_flatten.py and test_json_flatten.py - I figured the test file had enough examples in that it should act as a good source of information for the documentation.
This worked really well! You can see the first draft it produced here.
It included before and after examples in the documentation. I didn't fully trust these to be accurate, so I gave it this follow-up prompt:
llm -c "Rewrite that document to use the Python cog library to generate the examples"
I'm a big fan of Cog for maintaining examples in READMEs that are generated by code. Cog has been around for a couple of decades now so it was a safe bet that Claude would know about it.
This almost worked - it produced valid Cog syntax like the following:
[[[cog
example = {
"fruits": ["apple", "banana", "cherry"]
}
cog.out("```json\n")
cog.out(str(example))
cog.out("\n```\n")
cog.out("Flattened:\n```\n")
for key, value in flatten(example).items():
cog.out(f"{key}: {value}\n")
cog.out("```\n")
]]]
[[[end]]]
But that wasn't entirely right, because it forgot to include the Markdown comments that would hide the Cog syntax, which should have looked like this:
<!-- [[[cog -->
...
<!-- ]]] -->
...
<!-- [[[end]]] -->
I could have prompted it to correct itself, but at this point I decided to take over and edit the rest of the documentation by hand.
The end result was documentation that I'm really happy with, and that I probably wouldn't have bothered to write if Claude hadn't got me started.